SMART GRIDS Clouds , C ommuni c a t i o n s , O p e n Sourc e, and Aut o m a t i o n Devices, Circuits, and Systems Series Editor Krzysztof Iniewski CMOS Emerging Technologies Research Inc., Vancouver, British Columbia, Canada PUBLISHED TITLES: Atomic Nanoscale Technology in the Nuclear Industry Taeho Woo Biological and Medical Sensor Technologies Krzysztof Iniewski Building Sensor Networks: From Design to Applications Ioanis Nikolaidis and Krzysztof Iniewski Circuits at the Nanoscale: Communications, Imaging, and Sensing Krzysztof Iniewski Electrical Solitons: Theory, Design, and Applications David Ricketts and Donhee Ham Electronics for Radiation Detection Krzysztof Iniewski Embedded and Networking Systems: Design, Software, and Implementation Gul N. Khan and Krzysztof Iniewski Energy Harvesting with Functional Materials and Microsystems Madhu Bhaskaran, Sharath Sriram, and Krzysztof Iniewski Graphene, Carbon Nanotubes, and Nanostuctures: Techniques and Applications James E. Morris and Krzysztof Iniewski High-Speed Photonics Interconnects Lukas Chrostowski and Krzysztof Iniewski Integrated Microsystems: Electronics, Photonics, and Biotechnology Krzysztof Iniewski Integrated Power Devices and TCAD Simulation Yue Fu, Zhanming Li, Wai Tung Ng, and Johnny K.O. Sin Internet Networks: Wired, Wireless, and Optical Technologies Krzysztof Iniewski Low Power Emerging Wireless Technologies Reza Mahmoudi and Krzysztof Iniewski Medical Imaging: Technology and Applications Troy Farncombe and Krzysztof Iniewski PUBLISHED TITLES: MEMS: Fundamental Technology and Applications Vikas Choudhary and Krzysztof Iniewski Microluidics and Nanotechnology: Biosensing to the Single Molecule Limit Eric Lagally and Krzysztof Iniewski MIMO Power Line Communications: Narrow and Broadband Standards, EMC, and Advanced Processing Lars Torsten Berger, Andreas Schwager, Pascal Pagani, and Daniel Schneider Nano-Semiconductors: Devices and Technology Krzysztof Iniewski Nanoelectronic Device Applications Handbook James E. Morris and Krzysztof Iniewski Nanoplasmonics: Advanced Device Applications James W. M. Chon and Krzysztof Iniewski Nanoscale Semiconductor Memories: Technology and Applications Santosh K. Kurinec and Krzysztof Iniewski Novel Advances in Microsystems Technologies and Their Applications Laurent A. Francis and Krzysztof Iniewski Optical, Acoustic, Magnetic, and Mechanical Sensor Technologies Krzysztof Iniewski Radiation Effects in Semiconductors Krzysztof Iniewski Semiconductor Radiation Detection Systems Krzysztof Iniewski Smart Grids: Clouds, Communications, Open Source, and Automation David Bakken and Krzysztof Iniewski Smart Sensors for Industrial Applications Krzysztof Iniewski Technologies for Smart Sensors and Sensor Fusion Kevin Yallup and Krzysztof Iniewski Telecommunication Networks Eugenio Iannone Testing for Small-Delay Defects in Nanoscale CMOS Integrated Circuits Sandeep K. Goel and Krishnendu Chakrabarty Wireless Technologies: Circuits, Systems, and Devices Krzysztof Iniewski FORTHCOMING TITLES: 3D Circuit and System Design: Multicore Architecture, Thermal Management, and Reliability Rohit Sharma and Krzysztof Iniewski Circuits and Systems for Security and Privacy Farhana Sheikh and Leonel Sousa CMOS: Front-End Electronics for Radiation Sensors Angelo Rivetti Gallium Nitride (GaN): Physics, Devices, and Technology Farid Medjdoub and Krzysztof Iniewski High Frequency Communication and Sensing: Traveling-Wave Techniques Ahmet Tekin and Ahmed Emira High-Speed Devices and Circuits with THz Applications Jung Han Choi and Krzysztof Iniewski Labs-on-Chip: Physics, Design and Technology Eugenio Iannone Laser-Based Optical Detection of Explosives Paul M. Pellegrino, Ellen L. Holthoff, and Mikella E. Farrell Metallic Spintronic Devices Xiaobin Wang Mobile Point-of-Care Monitors and Diagnostic Device Design Walter Karlen and Krzysztof Iniewski Nanoelectronics: Devices, Circuits, and Systems Nikos Konofaos Nanomaterials: A Guide to Fabrication and Applications Gordon Harling and Krzysztof Iniewski Nanopatterning and Nanoscale Devices for Biological Applications Krzysztof Iniewski and Seila Selimovic Optical Fiber Sensors and Applications Ginu Rajan and Krzysztof Iniewski Organic Solar Cells: Materials, Devices, Interfaces, and Modeling Qiquan Qiao and Krzysztof Iniewski Power Management Integrated Circuits and Technologies Mona M. Hella and Patrick Mercier Radio Frequency Integrated Circuit Design Sebastian Magierowski Semiconductor Device Technology: Silicon and Materials Tomasz Brozek and Krzysztof Iniewski FORTHCOMING TITLES: Soft Errors: From Particles to Circuits Jean-Luc Autran and Daniela Munteanu VLSI: Circuits for Emerging Applications Tomasz Wojcicki and Krzysztof Iniewski Wireless Transceiver Circuits: System Perspectives and Design Aspects Woogeun Rhee and Krzysztof Iniewski SMART GRIDS Clouds , C ommuni c a t i o n s , O p e n Sourc e, and Aut o m a t i o n EDITED BY David Bakken Wa s h in g to n Sta te University School of Ele c tr ic a l E n g in e e r in g and C omputer S cience MANAGING EDITOR Krzysztof Iniewski CMOS E m e r g in g Te c h n o lo g ies R esearch Inc. Va n c o u v e r, Br itis h Co lu mbia, C anada Boca Raton London New York CRC Press is an imprint of the Taylor & Francis Group, an informa business MATLAB® is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks does not warrant the accuracy of the text or exercises in this book. This book’s use or discussion of MATLAB® software or related products does not constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular use of the MATLAB® software. CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2014 by Taylor & Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Version Date: 20140305 International Standard Book Number-13: 978-1-4822-0612-8 (eBook - PDF) This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright. com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com Contents Preface.................................................................................................................... xiii Editors ...................................................................................................................... xv Contributors ...........................................................................................................xvii Chapter 1 Mission-Critical Cloud Computing for Critical Infrastructures .......... 1 Thoshitha Gamage, David Anderson, David Bakken, Kenneth Birman, Anjan Bose, Carl Hauser, Ketan Maheshwari, and Robbert van Renesse Chapter 2 Power Application Possibilities with Mission-Critical Cloud Computing .......................................................................................... 17 David Bakken, Pranavamoorthy Balasubramanian, Thoshitha Gamage, Santiago Grijalva, Kory W. Hedman, Yilu Liu, Vaithianathan Venkatasubramanian, and Hao Zho Chapter 3 Emerging Wide-Area Power Applications with Mission-Critical Data Delivery Requirements .............................................................. 33 Greg Zweigle Chapter 4 GridStat: High Availability, Low Latency, and Adaptive Sensor Data Delivery for Smart Generation and Transmission ..................... 55 David E. Bakken, Harald Gjermundrød, and Ioanna Dionysiou Chapter 5 A Distributed Framework for Smart Grid Modeling, Monitoring, and Control ................................................................... 115 Alfredo Vaccaro and Eugenio Zimeo Chapter 6 Role of PLC Technology in Smart Grid Communication Networks....................................................................................... 133 Angeliki M. Sarai, Artemis C. Voulkidis, Spiros Livieratos, and Panayotis G. Cottis Chapter 7 Power Grid Network Analysis for Smart Grid Applications............ 151 Zhifang Wang, Anna Scaglione, and Robert J. Thomas ix x Chapter 8 Contents Open Source Software, an Enabling Technology for Smart Grid Evolution .................................................................................. 179 Russell Robertson, Fred Elmendorf, and Shawn Williams Chapter 9 Contribution of Microgrids to the Development of the Smart Grid .............................................................................. 191 Tine L. Vandoorn and Lieven Vandevelde Chapter 10 Microgrids ........................................................................................ 213 Mietek Glinkowski, Adam Guglielmo, Alexandre Oudalov, Gary Rackliffe, Bill Rose, Ernst Scholtz, Lokesh Verma, and Fang Yang Chapter 11 Integrating Consumer Advance Demand Data in Smart Grid Energy Supply Chain ....................................................................... 251 Tongdan Jin, Chongqing Kang, and Heping Chen Chapter 12 Photovoltaic Energy Generation and Control for an Autonomous Shunt Active Power Filter................................................................. 275 Ayman Blorfan, Damien Flieller, Patrice Wira, Guy Sturtzer, and Jean Mercklé Chapter 13 Self-Tuning and Self-Diagnosing Simulation ................................... 311 Jin Ma Chapter 14 A Consensus-Based Fully Distributed Load Management Algorithm for Smart Grid ................................................................ 333 Yinliang Xu, Wei Zhang, and Wenxin Liu Chapter 15 Expert Systems Application for the Reconiguration of Electric Distribution Systems ........................................................................ 359 Horacio Tovar-Hernández and Guillermo Gutierrez-Alcaraz Chapter 16 Load Data Cleansing and Bus Load Coincidence Factors ............... 375 Wenyuan Li, Ke Wang, and Wijarn Wangdee Chapter 17 Smart Metering and Infrastructure .................................................. 399 Wenpeng Luan and Wenyuan Li xi Contents Chapter 18 Vision of Future Control Centers in Smart Grids ............................ 421 Fangxing Li, Pei Zhang, Sarina Adhikari, Yanli Wei, and Qinran Hu Index ...................................................................................................................... 435 Preface While electric interconnections have had different kinds and levels of intelligence in them for many decades, in the last 6 years the notion of the “smart grid” has come seemingly out of nowhere to be on the minds of not just power engineers but policy makers, regulators, rate commissions, and the general public. Inherent in the notion of the smart grid is the ability to communicate much more sensor data and have far more computations at many more locations using these data. The purpose of this book is to give power engineers, information technology workers in the electric sector, and others a snapshot of the state of the art and practice today as well as a peek into the future regarding the smart grid. There is a special focus on new kinds of communications and computations enabled or necessitated by the smart grid. This book is divided into four parts. Part I deals with cloud computing, whose use is being seriously considered for planning and operational use in a number of utilities and independent system operators/regional transmission organizations as of March 2014. Cloud computing has the potential to deploy massive amounts of computational resources to help grid operations, especially under contingency situations. Chapter 1 describes the mission-critical features that cloud computing infrastructures must support in order to be appropriate for operational use in power grids. It also describes the Advanced Research Projects Agency-Energy GridCloud project to develop such technologies. Chapter 2 describes a handful of “killer apps” for cloud computing in power grid operations. It has been written by leading power researchers. Part II deals with wide-area communications for power grids. Chapter 3 describes a wide range of power application programs that have extreme communications requirements over wide distances. Such applications are becoming more widely deployed as grids come under more pressure with every passing year. Chapter 4 describes GridStat, a middleware communications framework designed from the ground up to meet these challenging requirements. The chapter includes a detailed analysis of how different technologies used in today’s grids such as multiprotocol label switching, Internet protocol multicast, IEC 61850, and others are inadequate for the applications described in Chapter 3 and the requirements derived from them in Chapter 4. Chapter 5 presents an advanced framework based on the service-oriented architecture approach for integrated modeling, monitoring, and control. Chapter 6 analyzes the role of power line communication, which is also called broadband over power lines, in the smart grid. Power line communication/broadband over power lines technologies can provide additional redundant paths for data delivery in a grid, and ones that have failure characteristics other than traditional network communications infrastructures. Finally, Chapter 7 describes a novel approach for estimating the statistical properties of power grids. This is an important irst step toward having more broadly reusable power algorithms with greater conidence, as computer scientists and mathematicians have done for centuries. Part III deals with open source, something common in other industries that is starting to draw great interest from utilities and has great potential to help stimulate innovation in power grids (which suffer from a far higher degree of “vendor lock-in” than xiii xiv Preface most other industries). Chapter 8 explains what open source software is and its history. It then overviews a number of freely available open source power application programs. Part IV deals with the broad category of automation. Chapter 9 explains how microgrids it into the smart grid landscape and how they can contribute to its operations. Chapter 10 describes in detail the design and operation of microgrids. Chapter 11 introduces a virtual energy provisioning concept by which utilities can collect and aggregate advanced demand information in order to better manage smart grid supply chains. Chapter 12 describes a new technique for better managing photovoltaic energy while limiting harmonic pollution. Chapter 13 provides an approach for two-way interactions between simulations and an operational wide-area measurement system that is both self-tuning and self-diagnosing. Chapter 14 describes an approach for load management in smart grids that is stable, distributed and employs multiagent techniques. Chapter 15 details the use of an expert system application to enable electric distribution systems to be reconigured in new and advantageous way. Chapter 16 describes an approach for both cleansing the load curve data and calculating bus load coincidence factors in order to better exploit smart meter data. Chapter 17 overviews an advanced metering infrastructure system and its components, discusses its beneits, and summarizes a variety of applications by which smart metering and infrastructure supports both planning and operations. Finally, Chapter 18 offers a vision of how smart grid control centers may look in the future. David E. Bakken Pullman, Washington Krzysztof (Kris) Iniewski Vancouver, British Columbia MATLAB® is a registered trademark of The MathWorks, Inc. For product information, please contact: The MathWorks, Inc. 3 Apple Hill Drive Natick, MA 01760-2098 USA Tel: 508 647 7000 Fax: 508-647-7001 E-mail: [email protected] Web: www.mathworks.com MATLAB® and Simulink® are trademarks of the MathWorks, Inc. and are used with permission. The MathWorks does not warrant the accuracy of the text or exercises in this book. This book’s use or discussion of MATLAB® and Simulink® software or related products does not constitute endorsement or sponsorship by the MathWorks of a particular pedagogical approach or particular use of the MATLAB® and Simulink® software. Editors David Bakken is a professor of computer science in the School of Electrical Engineering and Computer Science at Washington State University and chief scientist at GridStat, Inc. His research interests include wide-area distributed computing systems, middleware implementation, and dependable computing. Since 1999, he has been working closely with researchers in his department’s very strong electric power group on helping rethink the way data delivery is done in power grids over the wide area, and is considered the world’s leading expert on this. His GridStat data-delivery software has inluenced the shape of the emerging NASPInet. He is a frequent visitor and lecturer at utilities, electrical engineering departments, and power meetings worldwide. Prior to Washington State University, Dr. Bakken was a research scientist at BBN (Cambridge, MA), which built the irst Internet in 1969. There he was coinventor of the Quality Objects middleware framework, in which the Defense Advanced Research Projects Agency invested more than 50 BBN person-years, which was integrated with approximately 10 other research projects in various demonstrations, and which lew in Boeing experimental aircraft. Dr. Bakken has worked for Boeing and consulted for Amazon.com, Harris Corp., Realtime Innovations, Intel, TriGeo Network Security, and others. He holds a MS (1990) and a PhD (1994) in computer science from the University of Arizona, and Bachelor of Science degrees in computer science and mathematics from Washington State University (1985). He is the author of over 100 publications and coinventor of three patents. Krzysztof (Kris) Iniewski manages R&D at Redlen Technologies Inc., a startup company in Vancouver, Canada. Redlen’s revolutionary production process for advanced semiconductor materials enables a new generation of more accurate, alldigital, radiation-based imaging solutions. Kris is also president of CMOS Emerging Technologies Research Inc. (www.cmosetr.com), an organization of high-tech events covering communications, microsystems, optoelectronics, and sensors. In his career, Dr. Iniewski has held numerous faculty and management positions at the University of Toronto, the University of Alberta, Simon Fraser University, and PMC-Sierra Inc. He has published over 100 research papers in international journals and conferences. He holds 18 international patents granted in the United States, Canada, France, Germany, and Japan. He is a frequent invited speaker and has consulted for multiple organizations internationally. He has written and edited several books for CRC Press, Cambridge University Press, IEEE Press, Wiley, McGraw-Hill, Artech House, and Springer. His personal goal is to contribute to healthy living and sustainability through innovative engineering solutions. In his leisure time, Kris can be found hiking, sailing, skiing, or biking in beautiful British Columbia. He can be reached at [email protected]. xv Contributors Sarina Adhikari Department of Electrical Engineering and Computer Science University of Tennessee Knoxville, Tennessee Anjan Bose School of Electrical Engineering and Computer Science Washington State University Pullman, Washington David Anderson School of Electrical Engineering and Computer Science Washington State University Pullman, Washington Heping Chen Ingram School of Engineering Texas State University San Marcos, Texas David Bakken School of Electrical Engineering and Computer Science Washington State University Pullman, Washington Pranavamoorthy Balasubramanian School of Electrical, Computer, and Energy Engineering Arizona State University Tempe, Arizona Kenneth Birman Department of Computer Science Cornell University Ithaca, New York Ayman Blorfan Modelling, Intelligence, Process and Systems Laboratory Université de Haute Alsace Mulhouse, France and National Institute of Applied Science Strasbourg, France Panayotis G. Cottis School of Electrical and Computer Engineering National Technical University of Athens Athens, Greece Ioanna Dionysiou Department of Computer Science University of Nicosia Nicosia, Cyprus Fred Elmendorf Grid Protection Alliance Chattanooga, Tennessee Damien Flieller National Institute of Applied Science Research Group of Electrical and Electronics in Nancy Strasbourg, France Thoshitha Gamage School of Electrical Engineering and Computer Science Washington State University Pullman, Washington xvii xviii Contributors Harald Gjermundrød Department of Computer Science University of Nicosia Nicosia, Cyprus Tongdan Jin Ingram School of Engineering Texas State University San Marcos, Texas Mietek Glinkowski ABB Inc. Raleigh, North Carolina Chongqing Kang Department of Electrical Engineering Tsinghua University Beijing, China Santiago Grijalva School of Electrical and Computer Engineering Georgia Institute of Technology Atlanta, Georgia Fangxing Li Department of Electrical Engineering and Computer Science University of Tennessee Knoxville, Tennessee Adam Guglielmo ABB Inc. Raleigh, North Carolina Wenyuan Li School of Electrical Engineering Chongqing University Chongqing, China Guillermo Gutierrez-Alcaraz Department of Electrical Engineering Instituto Tecnológico de Morelia Morelia, Mexico Carl Hauser School of Electrical Engineering and Computer Science Washington State University Pullman, Washington Kory W. Hedman School of Electrical, Computer, and Energy Engineering Arizona State University Tempe, Arizona Qinran Hu Department of Electrical Engineering and Computer Science University of Tennessee Knoxville, Tennessee and BC Hydro Vancouver, Canada Yilu Liu Department of Electrical Engineering and Computer Science University of Tennessee Knoxville, Tennessee Wenxin Liu Klipsch School of Electrical and Computer Engineering New Mexico State University Las Cruces, New Mexico Spiros Livieratos School of Pedagogical and Technological Education Athens, Greece Wenpeng Luan State Grid Smart Grid Research Institute Beijing, China xix Contributors Jin Ma School of Electrical and Information Engineering The University of Sydney New South Wales, Australia Anna Scaglione Department of Electrical and Computer Engineering University of California Davis, California Ketan Maheshwari Argonne National Laboratory Lemont, Illinois Ernst Scholtz ABB Inc. Raleigh, North Carolina Jean Mercklé Modelling, Intelligence, Process and Systems Laboratory Université de Haute Alsace Mulhouse, France Guy Sturtzer National Institute of Applied Science Research Group of Electrical and Electronics in Nancy Strasbourg, France Alexandre Oudalov ABB Inc. Raleigh, North Carolina Robert J. Thomas Department of Electrical and Computer Engineering Cornell University Ithaca, New York Gary Rackliffe ABB Inc. Raleigh, North Carolina Robbert van Renesse Department of Computer Science Cornell University Ithaca, New York Russell Robertson Grid Protection Alliance Chattanooga, Tennessee Bill Rose ABB Inc. Raleigh, North Carolina Angeliki M. Sarai School of Electrical and Computer Engineering National Technical University of Athens Athens, Greece Horacio Tovar-Hernández Department of Electrical Engineering Instituto Tecnológico de Morelia Morelia, Mexico Alfredo Vaccaro Department of Engineering University of Sannio Benevento, Italy Lieven Vandevelde Department of Electrical Energy Ghent University Ghent, Belgium Tine L. Vandoorn Department of Electrical Energy Ghent University Ghent, Belgium xx Contributors Vaithianathan Venkatasubramanian School of Electrical Engineering and Computer Science Washington State University Pullman, Washington Patrice Wira Modelling, Intelligence, Process and Systems Laboratory Université de Haute Alsace Mulhouse, France Lokesh Verma ABB Inc. Raleigh, North Carolina Yinliang Xu Klipsch School of Electrical and Computer Engineering New Mexico State University Las Cruces, New Mexico Artemis C. Voulkidis School of Electrical and Computer Engineering National Technical University of Athens Athens, Greece Ke Wang School of Computing Science Simon Fraser University Vancouver, Canada Zhifang Wang Department of Electrical and Computer Engineering Virginia Commonwealth University Richmond, Virginia Wijarn Wangdee The Sirindhorn International Thai–German Graduate School of Engineering King Mongkut’s University of Technology North Bangkok Bangkok, Thailand Yanli Wei Department of Electrical Engineering and Computer Science University of Tennessee Knoxville, Tennessee Shawn Williams Grid Protection Alliance Chattanooga, Tennessee Fang Yang ABB Inc. Raleigh, North Carolina Pei Zhang Grid Operations and Planning Electric Power Research Institute Palo Alto, California Wei Zhang Klipsch School of Electrical and Computer Engineering New Mexico State University Las Cruces, New Mexico Hao Zhu Department of Electrical and Computer Engineering University of Illinois Champaign, Illinois Eugenio Zimeo Department of Engineering University of Sannio Benevento, Italy Greg Zweigle Schweitzer Engineering Laboratories, Inc. Pullman, Washington 1 Mission-Critical Cloud Computing for Critical Infrastructures Thoshitha Gamage, David Anderson, David Bakken, Kenneth Birman, Anjan Bose, Carl Hauser, Ketan Maheshwari, and Robbert van Renesse CONTENTS 1.1 Introduction ...................................................................................................... 1 1.1.1 Cloud Computing.................................................................................. 2 1.1.2 Advanced Power Grid ...........................................................................4 1.2 Cloud Computing’s Role in the Advanced Power Grid .................................... 5 1.2.1 Berkeley Grand Challenges and the Power Grid ..................................7 1.3 Model for Cloud-Based Power Grid Applications ............................................ 8 1.4 GridCloud: A Capability Demonstration Case Study ......................................9 1.4.1 GridStat.................................................................................................9 1.4.2 Isis2...................................................................................................... 10 1.4.3 TCP-R ................................................................................................. 11 1.4.4 GridSim .............................................................................................. 12 1.4.5 GridCloud Architecture ...................................................................... 12 1.5 Conclusions ..................................................................................................... 13 References ................................................................................................................ 15 1.1 INTRODUCTION The term cloud is becoming prevalent in nearly every facet of day-to-day life, bringing up an imperative research question: how can the cloud improve future critical infrastructures? Certainly, cloud computing has already made a huge impact on the computing landscape and has permanently incorporated itself into almost all sectors of industry. The same, however, cannot be said of critical infrastructures. Most notably, the power industry has been very cautious regarding cloud-based computing capabilities. This is not a total surprise: the power industry is notoriously conservative about changing its operating model, and its rate commissions are generally focused on short-term goals. With thousands of moving parts, owned and operated by just as many stakeholders, even modest changes are dificult. Furthermore, continuing to 1 2 Smart Grids operate while incorporating large paradigm shifts is neither a straightforward nor a risk-free process. In addition to industry conservatism, progress is slowed by the lack of comprehensive cloud-based solutions meeting current and future power grid application requirements. Nevertheless, there are numerous opportunities on many fronts—from bulk power generation, through wide-area transmission, to residential distribution, including at the microgrid level—where cloud technologies can bolster power grid operations and improve the grid’s eficiency, security, and reliability. The impact of cloud computing is best exempliied by the recent boom in e-commerce and online shopping. The cloud has empowered modern customers with outstanding bargaining power in making their purchasing choices by providing up-to-date pricing information on products from a wide array of sources whose computing infrastructure is cost-effective and scalable on demand. For example, not long ago air travelers relied on local travel agents to get the best prices on their reservations. Cloud computing has revolutionized this market, allowing vendors to easily provide customers with web-based reservation services. In fact, a recent study shows that online travel e-commerce skyrocketed from a mere $30 billion in 2002 to a staggering $103 billion, breaking the $100 billion mark for the irst time in the United States in 2012 [1]. A similar phenomenon applies to retail shopping. Nowadays, online retail shops offer a variety of products, ranging from consumer electronics, clothing, books, jewelry, and video games to event tickets, digital media, and lots more at competitive prices. Mainstream online shops such as Amazon, eBay, Etsy, and so on provide customers with an unprecedented global marketplace to both buy and sell items. Almost all major US retail giants, such as Walmart, Macy’s, BestBuy, Target, and so on, have adopted a hybrid sales model, providing online shops to complement the traditional in-store shopping experience. A more recent trend is lash sale sites (Fab, Woot, Deals2Buy, Totsy, MyHabit, etc.), which offer limited-time deals and offers. All in all, retail e-commerce in the United States increased by as much as 15% in 2012, totaling $289 billion. To put this into perspective, the total was $72 billion 10 years earlier. Such rapid growth relied heavily on cloud-based technology to provide the massive computing resources behind online shopping. 1.1.1 CLOUD COMPUTING What truly characterizes cloud computing is its business model. The cloud provides on-demand access to virtually limitless hardware and software resources meeting the users’ requirements. Furthermore, users only pay for resources they use, based on the time of use and capacity. The National Institute of Standards and Technology (NIST) deines ive essential cloud characteristics: on-demand self-service, broad network access, resource pooling, rapid elasticity, and measured service [2]. The computational model of the cloud features two key characteristics— abstraction and virtualization. The cloud provides its end users with well-deined application programming interfaces (APIs) that support requests to a wide range of hardware and software resources. Cloud computing supports various conigurations (central processing unit [CPU], memory, platform, input/output [I/O], networking, storage, servers) and capacities (scale) while abstracting resource management Mission-Critical Cloud Computing for Critical Infrastructures 3 (setup, startup, maintenance, etc.), underlying infrastructure technology, physical space, and human labor requirements. The end users see only APIs when they access services on the cloud. For example, users of Dropbox, the popular cloud-based online storage, only need to know that their stored items are accessible through the API; they do not need any knowledge of the underlying infrastructure supporting the service. Furthermore, end users are relieved of owning large computing resources that are often underused. Instead, resources are housed in large data centers as a shared resource pool serving multiple users, thus optimizing their use and amortizing the cost of maintenance. At the same time, end users are unaware of where their resources physically reside, effectively virtualizing the computing resources. Cloud computing provides three service models: software as a service (SaaS), platform as a service (PaaS), and infrastructure as a service (IaaS). Each of these service models provides unique APIs. Services can be purchased separately, but are typically purchased as a solution stack. The SaaS model offers end-point business applications which are customizable and conigurable based on speciic needs. One good example is the Google Apps framework, which offers a large suite of end-user applications (email, online storage, streaming channels, domain names, messaging, web hosting, etc.) that individuals, businesses, universities, and other organizations can purchase individually or in combination. Software offered in this manner has a shorter development life cycle, resulting in frequent updates and up-to-date versions. The life-cycle maintenance is explicitly handled by the service provider, who offers the software on a pay-per-use basis. Since the software is hosted in the cloud, there is no explicit installation or maintenance process for the end users in their native environment. Some of the prominent SaaS providers include Salesforce, Google, Microsoft, Intuit, Oracle, and so on (Figure 1.1). The PaaS model offers a development environment, middleware capabilities, and a deployment stack for application developers to build tailor-made applications or host prepurchased SaaS. Amazon Web Services (AWS), Google App Engine, and Microsoft Azure are a few examples of PaaS. In contrast to SaaS, PaaS does not abstract development life-cycle support, given that most end users in this model are application developers. Nevertheless, the abstraction aspect of cloud computing is Software as a service (application) Platform as a service (operating system) Infrastructure as a service (hardware) FIGURE 1.1 Cloud service models as a stack. 4 Smart Grids still present in PaaS, where developers rely on underlying abstracted features such as infrastructure, operating system, backup and version control features, development and testing tools, runtime environment, worklow management, code security, and collaborative facilities. The IaaS model offers the fundamental hardware, networking, and storage capabilities needed to host PaaS or custom user platforms. Services offered in IaaS include hardware-level provisioning, public and private network connectivity, (redundant) load balancing, replication, data center space, and irewalls. IaaS relieves end users of operational and capital expenses. While the other two models also provide these features, here they are much more prominent, since IaaS is the closest model to actual hardware. Moreover, since the actual hardware is virtualized in climate-controlled data centers, IaaS can shield end users from eventual hardware failures, greatly increasing availability and eliminating repair and maintenance costs. A popular IaaS provider, Amazon Elastic Compute Cloud (EC2), offers 9 hardware instance families in 18 types [3]. Some of the other IaaS providers include GoGrid, Elastic Hosts, AppNexus, and Mosso [4]. 1.1.2 ADVANCED POWER GRID Online shopping is just one of many instances where cloud computing is making its mark on society. The power grid, in fact, is currently at an interesting crossroads in this technological space. One fundamental capability that engineers are striving to improve is the grid’s situational awareness—its real-time knowledge of grid state—through highly time-synchronized phasor measurement units (PMUs), accurate digital fault recorders (DFRs), advanced metering infrastructure (AMI), smart meters, and signiicantly better communication. The industry is also facing a massive inlux of ubiquitous household devices that exchange information related to energy consumption. In light of these new technologies, the traditional power grid is being transformed into what is popularly known as the “smart grid” or the “advanced power grid.” The evolution of the power grid brings its own share of challenges. The newly introduced data have the potential to dramatically increase accuracy, but only if processed quickly and correctly. True situational awareness and real-time control decisions go hand in hand. The feasibility of achieving these two objectives, however, heavily depends on three key features: 1. The ability to capture the power grid state accurately and synchronously 2. The ability to deliver grid state data reliably and in a timely manner over a (potentially) wide area 3. The ability to rapidly process large quantities of state data and redirect the resulting information to appropriate power application(s), and, to a lesser extent, the ability to rapidly acquire computing resources for on-demand data processing Emerging power applications are the direct beneiciaries of rapid data capture, delivery, processing, and retrieval. One such example is the transition from Mission-Critical Cloud Computing for Critical Infrastructures 5 conventional state estimation to direct state calculation. Beginning in the early 1960s, the power grid has been employing supervisory control and data acquisition (SCADA) technology for many of its real-time requisites, such as balancing load against supply, demand response, and contingency detection and analysis. SCADA uses a slow, cyclic polling architecture in which decisions are based on unsynchronized measurements that may be several seconds old. Consequently, the estimated state lags the actual state most. Thus, state estimation gives very limited insight and visibility into the grid’s actual operational status. In contrast, tightly time-synchronized PMU data streams deliver data under strict quality of service (QoS) guarantees—low latency and high availability—allowing control centers to perform direct state calculations and measurements. The capabilities that come with the availability of status data make creating a real-time picture of the grid’s operational state much more realistic [5]. There are also many myths surrounding the operations of a power grid in conjunction with big data and its eficient use. The following is a nonexhaustive list of some of these myths. 1. Timeliness: Real-time data is a relative term. Often the application requirements dictate the timeliness needs. Modern software and hardware technologies provide many workarounds on the timeliness of data availability on wide area networks with average bandwidths. One of them is selective packet dropping. This technique guarantees a minimum QoS while delivering information to recipients in a timely manner. Smart power grids will greatly beneit from these techniques. 2. Security and Safety: Security and safety are concerns often cited by decision makers when considering new technologies. While absolute security is impossible, most concerns arising from data security issues have been technically addressed. One large factor that affects security is human errors and oversights. Often, insuficient emphasis is given to this side of security. More and more emphasis is given to the communication channels. Securing an already secure channel only results in performance losses and overheads. 3. Cost: The cost of maintaining information infrastructures has become a major portion of budgets for large industries, and is a substantial challenge in running a sustainable, data-centered architecture. Thanks to data centers and cloud computing infrastructures, these challenges are being successfully addressed. Clouds facilitate outsourcing of large-scale computational infrastructures while achieving provably reliable QoS. 1.2 CLOUD COMPUTING’S ROLE IN THE ADVANCED POWER GRID Cloud computing can play a vital role in improving the advanced power grid’s situational awareness and the ability to derive better control decisions. As mentioned earlier, emerging power applications will leverage large amounts of data in making control decisions affecting the stability and reliability of the grid. 6 Smart Grids Analyzing and processing such large amounts of data require data parallelism and massive computational capabilities well beyond general-purpose computing. Beyond data analysis, the future grid can greatly beneit from much more extensive simulation and analysis to remediate stressful situations. These are spontaneous special purpose applications (e.g., system integrity protection schemes [SIPS], also known as remedial action schemes [RASs] or special protection schemes [SPSs]) [6], each with different needs—real time, scaling, and computational—that are triggered by grid disturbances such as frequency oscillations, voltage luctuations, line overloads, and blackouts. Moreover, the number of power grid applications and their computational needs can only be expected to increase as the grid evolves. Managing this variety of applications and needs presents a challenge. Keeping these applications running idle on dedicated hardware until the speciic condition is triggered is both ineficient and expensive. An elegant solution is presented here which utilizes cloud computing and its rapid elasticity. Power grid applications can utilize the cloud to rapidly deploy an application-speciic infrastructure using IaaS and PaaS to achieve new levels of availability and scalability. Availability and scalability are properties that are much harder to meet in a piecemeal fashion, but are inherent features of the cloud and easily adoptable. Cloud-based solutions also beneit entities at different levels of the control hierarchy, giving them the ability to perform an independent, replicated analysis on the same sensor data. The ability to elastically manage resources in the presence of a grid disturbance is extremely attractive in comparison with in-house solutions, which could be overprovisioned or underprovisioned at the time of need. Another area where cloud computing performs well is in supporting the varying needs of the growing ecosystem of power applications. Both PaaS and SaaS will be useful for developing and supporting power applications. PaaS for the power grid will need to encompass industry best practices, engineering standards, compliance requirements, and data privacy and security requirements as properties of the platform itself. The CAP theorem [7] argues that simultaneously achieving three key properties—consistency, availability, and partition tolerance— is impossible in distributed systems. As a result, and especially since their apps are not mission critical, present-day commercial clouds often sacriice consistency in favor of availability. Cloud environments that are used for power applications must be able to guarantee high-assurance properties, including consistency, fault tolerance, and real-time responsiveness, in order to support the anticipated needs of power applications. While PaaS enables power researchers and developers to expand the power application ecosystem, SaaS can abstract essential properties and requirements to provide end-user application solutions. Grid incident-speciic applications can be offered as SaaS, readily deployable by power utilities. The success of power grid SaaS depends heavily on the completeness and the richness of power grid PaaS. The overarching challenge lies in ensuring that power applications delivered across SaaS/PaaS models inherently carry the necessary high-assurance properties. The subtle intricacies of high-assurance properties, which are often outside the power engineering realm, will necessitate a different approach to cloud computing as well as a stronger mesh between power engineering and computer science. Mission-Critical Cloud Computing for Critical Infrastructures 7 1.2.1 BERKELEY GRAND CHALLENGES AND THE POWER GRID The Berkeley view of the cloud [8,9] outlines 10 grand challenges and opportunities for cloud computing. The following list reviews some of these challenges and their implications for cloud-based power grid applications: 1. Ensuring High Service Availability: Consistency is arguably one of the most critical requirements for cloud-based power applications [10], but availability is a close second. Many of the early adopters of cloud technology support availability as a liveness property, while smart-grid applications depend on availability as a safety property. Availability also relates to the issue of whether cloud-based power grid applications should follow stateful or stateless models. The ability to design stateful applications often depends on the availability of state information. Achieving high availability requires avoiding single point of failure scenarios and potential bottlenecks. The general consensus is that the cloud promotes and provides high availability. However, using cloud services from a single service provider allows a single point of failure [9]. Interoperability between different cloud vendors for the sake of availability for power grid applications merely because of the many proprietary and market advantages is a far-fetched ambition. Perhaps one solution would be for the power grid community to manage and operate its own cloud, either overlaying existing commercial clouds or as a private infrastructure with built-in replication at all levels. Such an initiative, however, would be dictated by the many economic drivers. 2. Eliminating Data Transfer Bottlenecks: Large amounts of high-frequency data must cross the cloud boundary to reach power applications running within the cloud. Application responsiveness is directly tied to the time lines with which data reach their destination. The outermost layer of the cloud can have a dual role as a sensor data aggregator for sources outside the cloud and as a multiplexer toward the applications within the cloud. Thus, a suficiently large number of replicated cloud end points for sensor data must be provided in order to prevent a potential data transfer bottleneck. 3. Assuring Data Conidentiality and Availability: For a community historically notorious for a conservative modus operandi, sharing sensor data is a frightening proposition. Power grid entities operate under industry regulations and standards that can prevent data sharing in many circumstances. Additionally, companies are reluctant to share market-sensitive data that could give away economic advantage. Power application data traversing the cloud must be secured, meeting all compliance requirements, so that they cannot be used by unintended parties. Thus, the cloud will need to provide adequate data sanitization and iltering capabilities to protect application data. 4. Performance Predictability under Scaling: The enormous amount of data that some power applications require, combined with the impacts of virtualized I/O channels and interrupts, leads to unpredictable performance 8 Smart Grids during elastic scaling. Different IaaS vendors exhibit different I/O performance, resource acquisition, and release time characteristics. The computational performance on current cloud computing infrastructures also shows signs of strain under scaling [11]. High-end batch processing applications will require improved resource sharing and scheduling capabilities for virtual machines to ensure strict QoS demands. 1.3 MODEL FOR CLOUD-BASED POWER GRID APPLICATIONS Many of the cloud adoption challenges outlined in Section 1.2 are essentially about supporting highly scalable, highly assured behaviors and stringent communication guarantees. These are properties that are rarely found in today’s commercial cloud infrastructures, which are optimized to support mobile applications and web-based e-commerce applications. The notions of real-time responsiveness, guaranteed consistency, data security, and fault tolerance are signiicantly more forgiving in these applications than in infrastructure control, supplying little incentive for current commercial clouds to embrace the type of changes necessary to support critical infrastructure systems. Figure 1.2 visually represents an abstract architectural model for cloud-based power applications. The architecture includes three basic components: 1. A real-time data collection and transport infrastructure 2. A soft state, elastic outer cloud tier that supports data collection, data preprocessing, temporary archiving, data sanitization and iltering, and multiplexing to services residing in interior tiers 3. Interior cloud tiers hosting services and applications, and supporting data processing, analysis, batch processing, persistent storage, and visualization functions The data collection and transportation infrastructure sits between the physical sensors and the outermost tier of the cloud, and is the communication backbone of Outermost cloud tier (data collectors) Interior cloud tier (applications) Data multiplexing Computation Sensors Data collection and transportation Data aggregation Hard archiving Batch processing Visualization Data filtering FIGURE 1.2 Analysis Soft archiving Data sanitization An abstract architectural model for a cloud-based power grid application. Mission-Critical Cloud Computing for Critical Infrastructures 9 the overall architecture. This component is responsible for delivering data that are produced outside the cloud to the irst-tier cloud collectors with strong QoS guarantees such as guaranteed delivery, ultrahigh availability, ultralow latency, and guaranteed latency. The soft state, outermost cloud tier provides the interface to data lowing to the applications hosted in the interior tiers. The primary objective of this tier is to provide high availability, to exhibit rapid elasticity, and to forward correct data to the appropriate applications. To aid in this process, this tier will also host auxiliary applications that provide data sanitization, iltering of bad data, buffering (or soft achieving), data preprocessing, and forwarding capabilities. Availability and fault tolerance are heightened by replicated shards—nodes that collect data from a group of sensors—and by mapping sensor data sources appropriately to the shards. The interior cloud tiers host the actual applications that consume data from the shards and perform analysis, computation, and batch processing tasks. Additionally, the results of these deeper computations may be delivered at high rates to visualization applications residing inside and outside the cloud. 1.4 GRIDCLOUD: A CAPABILITY DEMONSTRATION CASE STUDY An Advanced Research Projects Agency-Energy (ARPA-E)-funded, high-proile research collaboration between Cornell University and Washington State University is spearheading efforts to develop, prototype, and demonstrate a powerful and comprehensive software platform realizing the cloud computing needs of the future power grid. Appropriately named GridCloud [12], this research project aims to bring together best-of-breed, already existing high-assurance distributed system technologies as a basis to innovate new cloud architectural models for the monitoring, management, and control of power systems. The technologies integrated in this effort include GridStat [13,14], Isis2 [15,16], TCP-R [17], and GridSim [18]. A brief description of each of these technologies is presented here. 1.4.1 GRIDSTAT GridStat implements a data delivery overlay network framework designed from the bottom up to meet the challenging requirements of the electric power grid (see Chapter 4). Power grids today are increasingly populated with high-rate, time-synchronized sensors that include PMUs and DFRs, whose functionalities are actually blurring. High-rate, time-synchronized data are expected to form the basis of many monitoring and control applications with a wide variety of delivery requirements and conigurations across such dimensions as geographic scope, latency, volume, and required availability [19]. These needs cannot be met by Internet protocol (IP) multicast, which forces all subscribers of a given sensor variable to get all updates at the highest rate that any subscriber requires. They also cannot be met by multiprotocol label switching (MPLS), which is not designed to provide per-message guarantees (only overall statistical guarantees) and also only has three bits (eight categories) with which to categorize the millions of different sensor lows that will likely be deployed in 5–10 years. GridStat delivers rate-based updates of sensor variables with a wide range of QoS+ guarantees (latency, rate, availability) that include support for ultralow latency 10 Smart Grids and ultrahigh availability, which are implemented by sending updates over redundant disjoint paths, each of which meets the end-to-end latency requirements for the given subscription. Additionally, GridStat enables different subscribers to a given sensor variable to require different QoS+ guarantees, which can greatly reduce bandwidth requirements and improve scalability. GridStat’s data delivery plane is a lat graph of forwarding engines (FEs), each of which stores the state for every subscription whose updates it forwards. FEs forward sensor updates on each outgoing link at the highest rate that any downstream subscriber requires. They drop updates that are not needed downstream, based on the expressed rate requirements of subscribers. GridStat’s management plane is implemented as a hierarchy of QoS brokers that can be mapped onto the natural hierarchy of the power grid. Each node in the hierarchy is designed to contain policies for resource permissions, security permissions, aggregation, and adaptations to anomalies. With these policies, the management plane calculates the paths required for the data delivery (with the given number of disjoint paths) and updates the forwarding tables in the FEs. Applications interact with GridStat using publisher and subscriber software libraries through which the applications’ requirements for QoS are conveyed to the management plane. GridStat incorporates mechanisms for securing communication between the management plane entities and those of the data plane. Security mechanisms for end-to-end message security between publishers and subscribers are modular and conigurable, allowing different data streams and applications to fulill different security and real-time requirements [20]. GridStat in the power grid provides the opportunity to respond to different power system operating conditions with different communication conigurations. GridStat provides a mechanism by which communication patterns can be rapidly changed among multiple preconigured modes in response to anticipated power system contingencies. 1.4.2 ISIS2 Isis2 is a high-assurance replication and coordination technology that makes it easy to capture information at one location and share it in a consistent, fault-tolerant, secure manner with applications running at other locations—perhaps great numbers of them. This system revisits a powerful and widely accepted technology for replicating objects or computations, but with a new focus on running at cloud scale, where the system might be deployed onto thousands of nodes and supporting new styles of machine-learning algorithms. Isis2 enables massive parallelism, strong consistency, and automated fault tolerance, and requires little sophistication on the part of its users. With Isis2, all computational nodes and applications sharing the same data see it [the data?] evolve in the same manner and at nearly the same time, with delays often measured in hundreds of microseconds. The system also supports replicated computation and coordination: with Isis2 one could marshal 10,000 machines to jointly perform a computation, search a massive database, or simulate the consequences of control actions, all in a manner that is fast, secure against attack or intrusion, and correct even if some crashes occur. The form of assurance offered by Isis2 is associated with a formal model that merges two important prior models—virtual synchrony [21] and Paxos [22]. Isis2 Mission-Critical Cloud Computing for Critical Infrastructures 11 embeds these ideas into modern object-oriented programming languages. Isis2 is used to create two new components for GridCloud: a version of the technology specialized for use in wide-area power systems networks, and support for high-assurance smart-grid applications that are hosted in cloud computing data centers. The GridCloud researchers believe that Isis2 can be used to support services that run on standard cloud infrastructures and yet (unlike today’s cloud solutions) are able to guarantee continuous availability, automatically adapting under attack so that intruders cannot disrupt the grid even if a few nodes are compromised. They are also analyzing and demonstrating the best options for building cloud services that respond to requests in a time-critical manner. 1.4.3 TCP-R GridCloud will tie together a very large number of components, including sensors, actuators, forwarding elements and aggregators, cloud-based services, and so on, using Internet standards. For best performance, it is important that related components communicate using persistent, stateful connections. Stateful connections reduce retransmissions and wasteful connections, and provide better low control. The standard for stateful connections in the Internet is transmission control protocol (TCP). TCP provides network connections that provide reliable irst in, irst out (FIFO) communication as well as fair low provisioning using adaptive congestion windows. Consider a cloud service that sends commands to a large number of actuators. The cloud service consists of a cluster of a few hundred servers. To keep actuators simple, and also to allow lexibility in evolving the cloud service, the cloud service should appear to the actuators as a single end point with a single TCP/IP address. While an actuator will receive commands from a particular server machine in the cluster, it appears to the actuators (and their software) as if the cloud service is a single, highly reliable, and fast machine. It is desirable to maintain this illusion even when connections migrate between server machines for load balancing, for hardware or software upgrades, or when rebooting cloud servers. TCP connections, unfortunately, are between socket end points, and, using current operating systems abstractions, socket end points cannot migrate or survive process failures. Also, the cloud service would have to maintain a TCP socket for every actuator. This does not scale well, as each TCP socket involves storing a lot of state information. Replacing TCP with a radically different protocol would not be feasible today. Operating systems and even networking hardware implement TCP connections very eficiently. TCP is the dominant communication protocol on the Internet, and Internet routers have evolved to support TCP eficiently, easily scaling to many millions of simultaneous TCP connections. TCP-R proposes to support standard TCP connections, but to extend them with a technology that addresses the shortcomings mentioned above. The essential idea is to extend the cloud service with a ilter that intercepts and preprocesses TCP packets. The ilter is scalable and maintains little state per TCP connection (on the order of 32 bytes). It has only soft state (i.e., it does not have to store its state persistently across crashes, greatly simplifying fault tolerance). The ilter allows servers to migrate TCP connections, and TCP connections 12 Smart Grids to survive server failure and recovery. Originally developed to maintain TCP connections between border gateway protocol (BGP) (Internet routing) servers across failures and subsequent recovery [23], TCP-R is extended into a scalable technology for a cluster serving client end points and also to “park” connections that are not currently live. 1.4.4 GRIDSIM GridSim is a real-time, end-to-end power grid simulation package that is unique in its integration of a real-time simulator, data delivery infrastructure, and multiple applications all running in real time. The goal of this project is to simulate power grid operation, control, and communications on a grid-wide scale (e.g., the Western Interconnection), as well as to provide utilities with a way to explore new equipment deployments and predict reactions to contingencies. The ability to simulate operation under different equipment deployment conigurations includes large-scale conigurations of PMUs. With the objective of simulating real-world equipment usage, and usage in conjunction with readily available industry equipment, the GridSim simulation package uses the industry standard C37.118 data format for all streaming measurement data. The irst element in the GridSim platform is a transient power stability simulator, specially modiied to output streaming data in real time. The output data are encoded into C37.118 and sent to a huge number of substation processes. At each of these processes, the data are aggregated, as would be done in a real power utility substation. The data are also sent to any of the substation-level power applications that are running. Both the raw substation data as well as any power application outputs are then published to GridStat. GridStat allows the substation data to be distributed as they would be in the real world. Published data can be sent via redundant paths, 1◊many communication (publish-subscribe, whose degenerate version is network-level multicast), and so on. The lexibility provided by the GridStat data delivery middleware allows subscription applications to be easily integrated into the system with minimal reconiguration. Published data are available to any subscribers of GridStat, including the two applications included in the GridSim simulation, the hierarchical state estimator (HSE), and the oscillation and damping monitor (Figure 1.3). 1.4.5 GRIDCLOUD ARCHITECTURE GridCloud was designed with the expectation that the developers of the advanced power grid will require easy access to large computing resources. Tasks may require large-scale computation, or may involve such large amounts of data that simply hosting and accessing the data will pose a substantial scalability challenge. This leads us to believe that cloud computing will play an important role in the future grid, supplementing the roles played by existing data center architectures. The compelling economics of cloud computing, the ease of creating “apps” that might control household power consumption (not a subject that has been mentioned yet), and the remarkable scalability of the cloud all support this conclusion. 13 Mission-Critical Cloud Computing for Critical Infrastructures Powertech TSAT simulator Measurement generator Static data generator Simulated power system C37.118 generator GridStat FE FE Substationlevel simulation Substation Substation N O Substation Su O Substation SE Su OM Substation SE Sub OM Substation SE Subs OM Substation SE Substation gateway SE FE FE FE FE Substation 1 Controllevel applications OpenPDC Oscillation monitor State estimator FIGURE 1.3 The GridSim architecture. (From Anderson, D., Zhao, C., Hauser, C., Venkatasubramanian, V., Bakken, D., and Bose, A., IEEE Power and Energy Magazine, 10, 49–57, 2012.) Figure 1.4 shows the architecture of GridCloud. The representative application used in this case is a HSE [18,24,25]. Data sources represent PMUs, which stream data to data collectors across a wide area using GridStat. The HSE comprises several substation-level state estimators that aggregate, ilter, and process PMU data before forwarding them to a control center-level state estimator. The input and the irst-level computation are inherently sharded at substation granularity. Furthermore, computations are inherently parallel between substations. Thus, the HSE has a natural mapping in GridCloud with substation state estimators residing in the outermost tier of the cloud while the control center state estimator is moved to the interior tier. The substation state estimators are replicated to increase fault tolerance and availability. The consistency of the replicas is managed through Isis2. TCP-R is used to provide fail-over capabilities for connections across the cloud. 1.5 CONCLUSIONS This chapter presents a roadmap of how cloud computing can be used to support the computational needs of the advanced power grid. Today’s commercial cloud computing infrastructure lacks the essential properties required by power grid applications. These deiciencies are explained and a cloud-based power grid application architecture is presented which overcomes these dificulties using well-known distributed system constructs. Furthermore, the GridSim project, which instantiates this model, is presented as a case study example. 14 EC2 cloud COL11 COL21 S1 COLN1 P1 S-SE11 S-SE21 Isis2 S-SEN1 TCP-R FE Control center SE COL FE GridStat (UDP) P2 1 2 COL22 FE COLN2 FE S-SE12 S-SE22 S-SE1 S-SE2 S-SEN2 FE FE Results S2 Computation FE Data origin Localhost Local client and visualizer S-SEN FE PM SK COL1M COL2M COLNM GridCloud S-SE2M S-SENM Output Smart Grids FIGURE 1.4 GridCloud architecture. S-SE1M Mission-Critical Cloud Computing for Critical Infrastructures 15 REFERENCES 1. G. Fulgoni, A. Lipsman and I. Essling, State of the U.S. Online Retail Economy in Q4 2012, February 2013. Available at: http://goo.gl/BCnbn. 2. P. Mell and T. Grance, The NIST Deinition of Cloud Computing: Recommendations of the National Institute of Standards and Technology. The National Institute of Standards and Technology, Gaithersburg, MD, 2011. 3. A. W. Services, Amazon elastic compute cloud: User guide API version 2013-02-01. http://awsdocs.s3.amazonaws.com/EC2/latest/ec2-ug.pdf (accessed 3 June 2013). 4. R. Prodan and S. Ostermann, A survey and taxonomy of infrastructure as a service and web hosting cloud providers, in Proceedings of the 10th IEEE/ACM International Conference on Grid Computing, October 13–15, 2009, Banff, AB. 5. D. Novosel, V. Madani, B. Bhargava, K. Vu and J. Cole, Dawn of the grid synchronization, IEEE Power and Energy Magazine, 6(1), 49–60, 2008. 6. S. Horowitz, D. Novosel, V. Madani and M. Adamiak, System-wide protection, IEEE Power and Energy Magazine, 6(September), 34–42, 2008. 7. S. Gilbert and N. Lynch, Brewer’s conjecture and the feasibility of consistent, available, partition-tolerant web services, ACM SIGACT News, 33(2), 51–59, 2002. 8. M. Armbrust, A. Fox, R. Grifith, A. D. Joseph, R. Katz, A. Konwinski, G. Lee, et al., A view of cloud computing, Communications of the ACM, 53(4), 50–58, 2010. 9. M. Armbrust, A. Fox, R. Grifith, A. D. Joseph, R. Katz, A. Konwinski, G. Lee, et al., Above the clouds: A Berkeley view of cloud computing, Department of Electrical Engineering and Computer Sciences, University of California Berkeley, Report UCB/ EECS 28, 2009. 10. K. Birman, L. Ganesh and R. V. Renesse, Running smart grid control software on cloud computing architectures, in Proceedings of the Computational Needs for the Next Generation Electric Grid Workshop, Cornell University, April, 2011, New York. 11. A. Iosup, S. Ostermann, M. Yigitbasi, R. Prodan, T. Fahringer and D. H. J. Epema, Performance analysis of cloud computing services for many-tasks scientiic computing, IEEE Transactions on Parallel and Distributed Systems, 22(6), 931–945, 2011. 12. K. Birman, A. Bose, D. Bakken and C. Hauser, GridControl: A software platform to support the smart grid, A Cornell University and Washington State University Research Collaboration, 2011. Available at: http://www.cs.cornell.edu/Projects/gridcontrol/index. html#gridcloud (accessed 3 June 2013). 13. H. Gjermundrod, D. Bakken, C. Hauser and A. Bose, GridStat: A lexible QoS-managed data dissemination framework for the power grid, IEEE Transactions on Power Delivery, 24(1), 136–143, 2009. 14. C. Hauser, D. Bakken and A. Bose, A failure to communicate: Next generation communication requirements, technologies, and architecture for the electric power grid, IEEE Power and Energy Magazine, 3(2), 47–55, 2005. 15. K. P. Birman, D. A. Freedman, Q. Huang and P. Dowell, Overcoming CAP with consistent soft-state replication, IEEE Computer, 45(2), 50–58, 2012. 16. K. Birman, Isis2 cloud computing library, 2013. Available at: http://isis2.codeplex.com/. 17. A. Agapi, K. Birman, R. M. Broberg, C. Cotton, T. Kielmann, M. Millnert, R. Payne, R. Surton and R. van Renesse, Routers for the cloud: Can the Internet achieve 5-nines availability?, IEEE Internet Computing, 15, 72–77, 2011. 18. D. Anderson, C. Zhao, C. Hauser, V. Venkatasubramanian, D. Bakken and A. Bose, A virtual smart grid: Real-time simulation for smart grid control and communications design, IEEE Power and Energy Magazine, 10(1), 49–57, 2012. 19. D. E. Bakken, A. Bose, C. H. Hauser, D. E. Whitehead and G. C. Zweigle, Smart generation and transmission with coherent, real-time data, Proceedings of the IEEE, 99(6), 928–951, 2011. 16 Smart Grids 20. E. Solum, C. Hauser, R. Chakravarthy and D. Bakken, Modular over-the-wire conigurable security for long-lived critical infrastructure monitoring systems, in Proceedings of the Third ACM International Conference on Distributed Event-Based Systems (DEBS), Nashville, TN, 2009. 21. K. Birman, Reliable Distributed Systems: Technologies, Web Services, and Applications, Springer, New York, 2005. 22. L. Lamport, “Paxos made simple, ACM SIGACT News, 32(4), 18–25, 2001. 23. R. Surton, K. Birman, R. Broberg, T. Marian and R. van Renesse, Man-in-the-Middle TCP recovery, Cornell University. Available at: http://goo.gl/9S4Jz (accessed 3 June 2013). 24. W. Jiang, V. Vittal and G. Heydt, A distributed state estimator utilizing synchronized phasor measurements, IEEE Transactions on Power Systems, 22(2), 563–571, 2007. 25. M. Zhao and A. Abur, Multi area state estimation using synchronized phasor measurements, IEEE Transactions on Power Systems, 20(2), 611–617, 2005. 2 Power Application Possibilities with Mission-Critical Cloud Computing David Bakken, Pranavamoorthy Balasubramanian, Thoshitha Gamage, Santiago Grijalva, Kory W. Hedman, Yilu Liu, Vaithianathan Venkatasubramanian, and Hao Zho CONTENTS 2.1 Overview......................................................................................................... 18 2.2 Robust Adaptive Topology Control ................................................................ 18 References ................................................................................................................ 19 2.3 Adaptive Real-Time Transient Stability Controls ...........................................20 Reference ................................................................................................................. 21 2.4 Prosumer-Based Power Grid........................................................................... 21 2.4.1 Introduction ........................................................................................ 21 2.4.2 Prosumer-Based Control Architecture ............................................... 22 2.4.3 Computational Challenges .................................................................. 23 2.4.4 Cloud Computing in the Future Electric Grid .................................... 23 2.4.5 Economic Dispatch of Stochastic Energy Resources ......................... 23 2.4.6 Cloud-Based Apps ..............................................................................24 2.4.7 Scenario Analysis and Transmission Planning ..................................24 2.4.8 Model Integration ...............................................................................24 2.4.9 Exploiting and Abstracting Self-Similarity ........................................25 References ................................................................................................................25 2.5 Wide-Area Frequency Monitoring .................................................................26 2.5.1 Introduction ........................................................................................26 2.5.2 FNET Architecture .............................................................................26 2.5.3 FNET Applications ............................................................................. 27 2.5.4 Cloud Computing and FNET.............................................................. 27 2.5.5 Rapidly Elastic Data Concentrators .................................................... 27 2.5.6 Computational Requirement Flexibility .............................................28 2.5.7 Cloud-Based Applications ..................................................................28 References ................................................................................................................ 29 17 18 Smart Grids 2.6 Oscillation Mitigation Strategies .................................................................... 29 References ................................................................................................................ 30 2.7 Automatic Network Partitioning for Steady-State Analysis ........................... 30 References ................................................................................................................ 31 2.1 OVERVIEW As we have seen in Chapter 1, not only is cloud computing coming to the grid, but mission-critical implementations such as GridCloud can provide mission-critical properties. This chapter explores new applications enabled by such technology. While this chapter is only scratching the surface of what is likely to be routine in a decade, we hope that it provides a tantalizing glimpse of what is possible. What, then, is mission-critical cloud computing? To recap, in a nutshell, it • Keeps the same fast throughput as generic commercial cloud platforms • Does not deliberately trade off this throughput to allow “inconsistencies,” for example, when a replica does a state update on a copy of the state but this update is “forgotten” • Is much more predictable (and faster) in terms of ramp-up time, central processing unit (CPU) performance per node, and number of nodes Therefore, the question for power application developers is how they can use: • Hundreds of processors in steady state. • Thousands or tens of thousands of processors when a contingency is reached or is being approached. Note: often there are many minutes of advanced warning of this, sometimes an hour or more. • Data from all participants in a grid that is enabled quickly when a crisis is approached (though, for market reasons, not necessarily during steady state). With this in mind, we now present groundbreaking applications that can exploit such mission-critical cloud platforms. 2.2 ROBUST ADAPTIVE TOPOLOGY CONTROL Balasubramanian and Hedman The electric power transmission system is one of the most complex systems available today. Traditionally, bulk power transmission systems (lines and transformers) are treated as static assets, even though these resources are controllable. However, it is known that transmission topology control has been used in the past and is still being used for corrective-based applications; for example, PJM uses corrective topology control as a special protection scheme (SPS) [1]. These switching actions are primarily taken on an ad hoc basis, determined by the system operators based on past historical data rather than in an automated way based on decision support tools. Past research has demonstrated the ability of topology control to help improve voltage proiles, increase transfer capacity, improve system reliability, and provide cost beneits to the system [2–8]. Even Power Application Possibilities with Mission-Critical Cloud Computing 19 though transmission topology control can provide these beneits, harnessing such lexibility from the transmission network in existing operational procedures is limited due to the computational challenges of optimizing the transmission topology. More recently, sensitivity-based methods have been proposed as a mechanism to reduce the computational complexity [9–12]. The robust adaptive topology control method develops a sensitivity-based heuristic, which reduces the computational time of the topology control problem. An expression is derived indicating the impact of changing the state of a transmission line on the objective. This expression is used to generate a line-ranking system with the potential candidate lines for switching based on a direct current (dc) optimal power low, which builds on the work of [12]. This approach selects a single feasible switching action per iteration, which provides an improvement to the system. The advantage of this method is that it solves linear programs iteratively to come up with a beneicial line-switching solution, which is computationally simple as compared with other methods employing mixed integer programming. All the possible switching solutions are lined up in the ranked list, with the switching action most likely to be beneicial placed at the top of the list. As the list is formed based on a sensitivity study, the switching action is not guaranteed to improve the system. Hence, the switching actions need to be checked for alternating current (ac) feasibility and whether they truly provide an improvement in the objective before they are implemented. This is done by selecting the irst action from the ranked list and simulating the switching to ind the improvement in the system. If the switching is not beneicial, the next action in the ranked list is checked for improvement. This process is continued until a beneicial switching action is found. While such a procedure is a heuristic, prior work has shown substantial economic savings [9] as well as strong performance in comparison with global optimization techniques [12]. The processing time taken to come up with a beneicial switching action could be signiicantly reduced if this process were parallelized so that all the proposed switching solutions could be checked at once. This opens up enormous opportunities for the application of cloud computing to transmission-switching applications, which would drastically reduce the computational time and improve the solution quality, as the best solution from the ranking list could be identiied very quickly. With prior research demonstrating cost savings of close to 4% for a $500 billion industry [3], there is a great opportunity for advanced decision support tools to ill this technological need, in terms of both algorithm sophistication and advanced computing capabilities, such as cloud computing. REFERENCES 1. PJM, Manual 3: Transmission Operations, Revision: 40, 2012. Available at: http://www. pjm.com/~/media/documents/manuals/m03.ashx. 2. W. Shao and V. Vittal, Corrective switching algorithm for relieving overloads and voltage violations, IEEE Transactions on Power Systems, 20(4), 1877–1885, 2005. 3. K. W. Hedman, M. C. Ferris, R. P. O’Neill, E. B. Fisher, and S. S. Oren, Co-optimization of generation unit commitment and transmission switching with N-1 reliability, IEEE Transactions on Power Systems, 25(2), 1052–1063, 2010. 4. K. W. Hedman, R. P. O’Neill, E. B. Fisher, and S. S. Oren, Optimal transmission switching with contingency analysis, IEEE Transactions on Power Systems, 24(3), 1577–1586, 2009. 20 Smart Grids 5. A. Korad and K. W. Hedman, Robust corrective topology control for system reliability, IEEE Transactions on Power Systems, 28(4), 4042–4051, 2013. 6. K. W. Hedman, R. P. O’Neill, E. B. Fisher, and S. S. Oren, Smart lexible just-in-time transmission and lowgate bidding, IEEE Transactions on Power Systems, 26(1), 93–102, 2011. 7. E. B. Fisher, R. P. O’Neill, and M. C. Ferris, Optimal transmission switching, IEEE Transactions on Power Systems, 23(3), 1346–1355, 2008. 8. K. W. Hedman, S. S. Oren, and R. P. O’Neill, A review of transmission switching and network topology optimization, in Proceedings of IEEE Power and Energy Society General Meeting, July 2011, Detroit, MI. 9. P. A. Ruiz, J. M. Foster, A. Rudkevich, and M. C. Caramanis, On fast transmission topology control heuristics, in Proceedings of IEEE Power and Energy Society General Meeting, July 2011, Detroit, MI. 10. J. M. Foster, P. A. Ruiz, A. Rudkevich, and M. C. Caramanis, Economic and corrective applications of tractable transmission topology control, in Proceedings of 49th Annual Allerton Conference on Communication, Control, and Computing, pp. 1302–1309, September 2011, Monticello, IL. 11. P. A. Ruiz, J. M. Foster, A. Rudkevich, and M. C. Caramanis, Tractable transmission topology control using sensitivity analysis, IEEE Transactions on Power Systems, 27(3), 1550–1559, 2012. 12. J. D. Fuller, R. Ramasra, and A. Cha, Fast heuristics for transmission line switching, IEEE Transactions on Power Systems, 27(3), 1377–1386, 2012. 2.3 ADAPTIVE REAL-TIME TRANSIENT STABILITY CONTROLS Venkatasubramanian The power system is expected to undergo major changes in the next decade, resulting from rapid growth in system loads (such as electric cars) and from increased dependence on renewable intermittent generation. To face up to these challenges, power utilities are making major upgrades to wide-area monitoring and control technologies, with impetus from major federal investments in the past few years. Power system operation is designed to withstand small- and large-scale disturbances. However, when the system is subjected to several large disturbances in a short span of time, it may become vulnerable to blackouts. Some recent events, such as the 2012 San Diego blackout and the 2003 Northeastern blackout, point to the need for adaptive real-time transient stability control designs that are speciically designed on an adaptive premise of making control decisions during the evolution of the event. In the present-day power system, wide-area transient stability controls such as remedial action schemes (RAS) or SPS are hard-coded control algorithms that are triggered by a central controller in response to the occurrence of speciic contingencies based on preset switching logic. When the system is subject to any “unknown” set of contingencies that is not part of the RAS controller logic, the system operation typically switches to a “safe mode” whereby interarea power transfers are limited to low conservative settings. The tie-line transfers remain at safe low values until the reliability coordinator completes a new set of transient stability simulation studies, which results in signiicant economic losses due to operation at nonoptimal power transfer levels. Cloud computing emerges as an ideal platform for handling transient stability mitigation issues, both for the present-day power system and for future control designs. In present-day operation, whenever the system operation is found to be in Power Application Possibilities with Mission-Critical Cloud Computing 21 one of the “unknown” operating conditions, the reliability coordinator can dial in a vast amount of cloud-based processing power to carry out the massive number of new transient stability simulations needed for determining the safe transfer limits. In the future, we need to rethink the design of transient stability controls such as RAS or SPS schemes. The massive computational capability offered by cloud computing opens up truly novel futuristic control schemes for mitigating transient stability events, as proposed in [1]. In the present-day power system, simulation studies are performed off-line for a “guesstimated” list of potential contingencies, and RAS schemes are implemented for a subset of problematic N−2 or higher-order contingencies whenever needed. Such RAS schemes, then, only work for a limited number of potential scenarios. Moreover, the respective control actions in these RAS schemes are also designed to be conservative, being based on off-line studies. Zweigle and Venkatasubramanian [1] propose to select and implement transient stability controls based on simulations of the system in real time during the evolution of the events themselves. Wide-area monitoring from an abundance of phasor measurement units (PMU) in the future will pave the way for real-time monitoring of the state and system topology of the full-size power system. Combining this real-time state information with real-time simulations will allow us to evaluate which control actions are optimally suited to the system at the present time, and the decisions are fully adaptive to whatever the system conditions are. Since the controller continues to monitor the system in a closed-loop fashion, the proposed control schemes are also robust with respect to simulation errors and communication or actuation failures. The formulation is not restrictive to any subset of contingencies, and can handle low-probability events consisting of multiple outages, such as those that have served as precursors to large blackouts in the past. In this proposed formulation, denoted as adaptive real-time transient stability controls, massive processing power is needed to carry out “what if” simulations of many potential control candidates in parallel before deciding on whether any control action is needed and which speciic action(s) will be implemented. The system monitoring and simulations of “what if” scenarios will continue throughout the event until the system has been stabilized. Once the controller recognizes that the system has returned to its normal state, the controller returns to dormant system monitoring mode, and cloud resources can be released. Details of the control algorithms can be seen in [1]. REFERENCE 1. G. Zweigle and V. Venkatasubramanian, Wide-area optimal control of electric power systems with application to transient stability for higher order contingencies, IEEE Transactions on Power Systems, 28(3), 2313–2320, 2013. 2.4 PROSUMER-BASED POWER GRID Gamage and Grijalva 2.4.1 IntroductIon The electric power grid, in a bid to improve its sustainability, is aggressively exploring ways to integrate distributed renewable energy generation and storage devices 22 Smart Grids at many levels. The most obvious integration is at the level of generation, where renewable generation sources such as large wind turbine and solar panel farms will supplement and eventually (it is hoped) supplant traditional nonrenewable power generation sources. Another natural integration is at the distribution level, where relatively smaller-scale renewable energy generation by utilities and other power distribution entities offers cheaper and greener energy options to customers. While not on the same bulk scale as the generation or distribution level, an emerging trend in recent years is for end consumers who are typically below the distribution level (e.g., households, microgrids, and energy buildings) to generate their own power using renewable sources and become self-sustainable and energy-independent of the grid. A fascinating aspect of this changing energy landscape is the drastic changes in the roles of the players involved. For example, end consumers, in addition to their typical energy consumption role, are economically motivated to sell excess energy and provide energy storage services to the grid. Modern utilities also go beyond their traditional energy distribution role in buying energy from end consumers when available. Similar role augmentations can be observed at all levels of the modern electric power grid [1]. As a consequence, what traditionally has been a one-way energy transfer—from bulk generation, through transmission and distribution, to end-user consumption—is transforming into a two-way energy exchange. 2.4.2 Prosumer-Based control archItecture The key entities in this evolving electric power grid are prosumers, economically motivated power system participants that can consume, produce, store, and transport energy [1,2,3]. Examples of prosumers include a building equipped with an electric vehicle parking lot that provides storage services to a utility, a hybrid (generates its own power but is not completely off-grid) renewable energy microgrid that sells excess power at peak generation to a neighboring microgrid or utility, and two households with renewable resources exchanging power on demand. In fact, almost all current power grid players can be conceptualized as prosumers [2]. A key characteristic of prosumers is that they assume different roles depending on the situation and underlying conditions; a prosumer who is a consumer at one moment can become a producer at the next. This is analogous to a peer-to-peer model versus a typical client–server model in computer science. Prosumers interact with one another through the services they offer. For example, a utility prosumer that aggregates heterogeneous home-user prosumers provides consumption and storage services to a distributed independent service operator (ISO) prosumer. Moreover, at any given instant, different prosumers may operate under different satisfaction and objective functions (comfort, cost, eficiency, security, sustainability, reliability, etc.) that will directly inluence their control decisions. Thus, the control architecture that revolves around prosumers is radically different from the current centralized hierarchical control architecture. Instead of waiting for control decisions to trickle down the hierarchy, a prosumer-based control architecture promotes autonomy by taking proactive and distributed control steps that relect the local, internal, and external state of the network of prosumers. Power Application Possibilities with Mission-Critical Cloud Computing 2.4.3 23 comPutatIonal challenges An important, but often overlooked, trait of the future electric power grid is its multidimensional, multiscale nature. The existing electricity industry is confronted by a massive invasion of energy-aware ubiquitous household devices that aim to empower consumers to make intelligent and mindful energy choices. The industry itself is invested in improving the grid’s visibility and real-time sensing capabilities by deploying PMUs, smart meters, and other intelligent and highly accurate electronic devices. An inadvertent effect of this inlux of new devices is that they produce data with temporal, spatial, and scenario signiicance that are useful in planning, simulation, operation, and control of the future power grid [4]. If the next evolution of the power grid is a prosumer-oriented two-way energy exchange, the true beneit of such massive quantities of multidimensional data profoundly depends on the ability to extract and render usable, useful, and mission-critical information at the right time and deliver it to the right destinations. Inarguably, the computational complexity of this daunting task is beyond the capability of today’s general-purpose computers. Additional major driving forces, such as federal and industry initiatives to incorporate renewable energy resources, various energy eficiency programs, and unconventional and novel system behaviors, coupled with consumer perceptions about the type of services they expect from the future power grid, add granularity on multiple scales to this computational challenge. 2.4.4 cloud comPutIng In the Future electrIc grId There are numerous computational challenges and limitations associated with existing computational models and analytical tools [4]. A primary objective of today’s power engineers and researchers alike is to design, develop, and innovate new models and analytical tools that can support and fully exploit massive multidimensional and multiscale data. Below are a few such examples that have cloud computing implications in the domain of potential solutions. 2.4.5 economIc dIsPatch oF stochastIc energy resources The inherent unpredictability and the variable nature of stochastic energy resources such as wind energy are making it dificult for utilities with large penetration of renewables to fully exercise real-time economic dispatch with a iner granularity [5]. The research community has already identiied the need to replace current economic dispatch software with short-term stochastic scheduling software that has a granularity of seconds, minutes, or a few hours [6,7]. Such software has highly elastic computational needs that scale up and down based on the conditions, the level of penetration, and the availability of information, thus offering an invitation to exploit the cloud’s rapid elasticity feature [8] to provision software, platform, and computational infrastructure with different conigurations on demand, rather than in-house computational stacks that offer little lexibility, scalability, or availability. 24 2.4.6 Smart Grids cloud-Based aPPs Most internal events experienced by prosumers are not visible to external entities. However, there are circumstances in which knowledge of such events becomes signiicant beyond the immediate control domain. For example, conventional load forecasting by ISOs within a 2% error bound will need to be tightened even further with the deployment of distributed energy resources and utility-scale storage devices [9]. The enablers needed for such capability are widely available and acceptable new modeling and simulation methods—single- versus three-phase modeling, nonstandardization of operational and planning models, node/breaker versus bus/branch modeling, data privacy, operational liability—which can be offered as softwareas-a-service (SaaS) solutions. 2.4.7 scenarIo analysIs and transmIssIon PlannIng Fundamentally, the power grid has resorted to using an N−1 contingency analysis of bulk transmission system security. The reality, however, is that, to achieve a higher level of system reliability, the power grid must be capable of evaluating a large number of plausible contingencies beyond N−1 conditions. This is certainly a computationally very intensive process. A heuristically feasible option would be to systematically prioritize searches, reduce scenario spaces, and make dynamically updated multidimensional information available to a wider audience of power grid entities. This is still beyond the capability of general-purpose computers, and would require a suficiently large high-performance computing facility. It is highly unrealistic to assume that every power grid entity will be excited about the proposition of individually housing such a highly expensive facility, not to mention the highly skilled workforce required to maintain it. A more likely solution is to utilize cloud computing infrastructures on the basis of need and usage, while greatly reducing the capital investment needed and abstracting the underlying maintenance cost and workforce requirements. Cloud computing can commission massive parallel computations to explore the huge permutation space on demand, and can also share the indings much more quickly with all interested entities. Even the data aggregation, mining, and analysis tasks can be ofloaded onto cloud-based computational platforms that offer better scalability features under changing conditions. 2.4.8 model IntegratIon Power systems at various spatial scales are modeled differently. Information exchange and seamless modeling integration require signiicant processing capabilities. Often, the data from more detailed models have to be aggregated and transferred “by hand” to other applications. Cloud-based modeling solutions, on the other hand, can greatly improve this process by using in-the-cloud dynamic and highly automated aggregation and delivery capabilities. These capabilities can be encapsulated as a library of models with a consistent description of phenomena, and various scales can be generalized and offered as platform-as-a-service (PaaS) options with well-deined application programming interfaces (APIs) to system developers who Power Application Possibilities with Mission-Critical Cloud Computing 25 can custom-build models that suit the individual spatial scaling requirements of different power system entities. Moreover, entities such as utilities or microgrids that operate individual control systems with homogeneous objective functions (e.g., cost minimization) can be aggregated for better control stability. Thus, the natural hierarchy of spatially separated power grid entities can be uniied under common protocols while still abiding by any regulatory or compliance standards. 2.4.9 exPloItIng and aBstractIng selF-sImIlarIty Self-similarity is the notion of entities at various levels of a system hierarchy exhibiting similar characteristics [10,11]. This is observable in the power grid, where entities at a lower level of the system are governed by physics similar to that of higher-level entities, but at a lower scale. What this means for a prosumer-based power grid is that multiscale prosumers at different levels can beneit from models that are based on the same reference model, which enables expansion and collapse of lower-level information as well as seamless integration and aggregation at higher levels. Since similar physics takes place at various hierarchical levels of the power system, data at different aggregation levels can be theoretically collapsed or expanded dynamically depending on the required granularity. For example, a combined securityconstrained optimal power low application may start with transmission-level congestion management, but certain buses (distribution systems) may be dynamically expanded, depending on the solution state, to address the internal constraints or conditions of the distribution system. This again raises an important computational issue—the ability to dynamically expand highly available computational capabilities on demand—which is yet another cloud computing possibility. REFERENCES 1. S. Grijalva, M. Costley, and N. Ainsworth, Prosumer-based control architecture for the future electricity grid, in Proceedings of 2011 IEEE International Conference on Control Applications (CCA), 28–30 September 2011, Denver, CO. 2. S. Grijalva and M. Tariq, Prosumer-based smart grid architecture enables a lat, sustainable electricity industry, in 2011 IEEE PES Innovative Smart Grid Technologies (ISGT), 17–19 January 2011, Anaheim, CA. 3. M. Tariq, S. Grijalva, and M. Wolf, Towards a distributed, service-oriented control infrastructure for smart grid, in Proceedings of the 2011 IEEE/ACM International Conference on Cyber-Physical Systems (ICCPS), 11–14 April 2011, Chicago, IL. 4. S. Grijalva, Research needs in multi-dimensional, multi-scale modeling and algorithms for next generation electricity grids, in DOE Conference on Computing Challenges for the Next-Generation Power Grid, 2011, Ithaca, NY. 5. M. H. Albadi and E. F. El-Saadany, Overview of wind power intermittency impacts on power systems, Electric Power Systems Research, 80(6), 627–632, 2010. 6. P. A. Ruiz, C. R. Philbrick, and P. W. Sauer, Wind power day-ahead uncertainty management through stochastic unit commitment policies, in IEEE/PES Power Systems Conference and Exposition (PSCE‘09), 15–18 March 2009, Seattle, WA. 7. W. B. Powell, A. George, H. S. Simão, A. W. Lamont, and J. Stewart, SMART: A stochastic multiscale model for the analysis of energy resources, technology, and policy, INFORMS Journal on Computing, 24(4), 665–682, 2012. 26 Smart Grids 8. P. Mell and T. Grance, The NIST Deinition of Cloud Computing: Recommendations of the National Institute of Standards and Technology, The National Institute of Standards and Technology, 2011, Gaithersburg, MD. 9. S. Fan, L. Chen, and W.-J. Lee, Short-term load forecasting using comprehensive combination based on multimeteorological information, IEEE Transactions on Industry Applications, 45(4), 1460–1466, 2009. 10. J. Maver, Self-similarity and points of interest, IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(7), 1211–1226, 2010. 11. C. Vandekerckhove, Macroscopic simulation of multiscale systems within the equationfree framework. PhD Thesis, Katholieke Universiteit Leuven, 2008, Leuven, Belgium. 2.5 WIDE-AREA FREQUENCY MONITORING Gamage and Liu 2.5.1 IntroductIon The power system frequency monitoring network (FNET), proposed in 2001 and established in 2004, is a pioneering wide-area monitoring system in the United States. Composed of a network of over 200 frequency disturbance recorder (FDR) devices—a member of the PMU family—that are spread throughout the country, the FNET serves the entire North American power grid through advanced situational awareness techniques, such as real-time event alerts, accurate event location estimation, animated event visualization, and postevent analysis [1–3]. An FDR is a single-phase PMU that is capable of measuring voltage phase angle, amplitude, and frequency from a single-phase voltage source, which can provide useful information for power system event recognition and status estimation. The frequency measurement algorithms in FDRs are highly accurate, with virtually zero error in the 52–70 Hz frequency range, and real hardware accuracy of ±0.0005 Hz, which outperforms even some of the commercial PMUs. The power grid operates in a very narrow frequency range, such that frequency excursions outside this range give clues to the problems. Frequency is a universal parameter across the entire interconnected power system that provides insight into generation electromechanical transients, generation-demand dynamics, and system operations such as load shedding, breaker reclosing, and capacitor bank switching. This characteristic allows frequency monitoring to be as informative at the distribution level as it is at the transmission level. 2.5.2 Fnet archItecture The overall FNET architecture, illustrated in Figure 2.1, comprises a widely deployed network of FDRs that report phasor measurements to either a local processing unit or a remote data center through Ethernet. The outermost layer of the data center is a data concentrator for the incoming ield FDR data. The data concentrator performs certain data preprocessing tasks, such as creating global positioning system (GPS) time-aligned records from the received data before forwarding them to FNET applications and storage agents in the interior levels of the data center. Power Application Possibilities with Mission-Critical Cloud Computing Client FDR 1 Satellite FDR 2 27 Firewall Data streaming Real-time applications and router Ethernet WAN Real-time data sharing Firewall Data concentrator and router FDR n Firewall and router Data storage Non-real-time applications Client Sensors Medium and clients Data center FIGURE 2.1 The FNET architecture is composed of widely deployed FDRs that transmit phasor measurements to either a local client or a remote data center for processing and longterm storage. Data centers host a multilayer agent hierarchy comprising a data concentrator in the top layer, and a data storage agent and real-time and non-real-time application agents in subsequent layers. 2.5.3 Fnet aPPlIcatIons There are two types of FNET application classes—real-time applications and nonreal-time applications [4]. Real-time applications have stringent timing requirements; responses need to be produced within seconds or subseconds after receiving the data. Examples of real-time FNET applications include frequency monitoring, event trigger, interarea oscillation trigger, line-trip detection and identiication, and event visualization. Non-real-time applications, on the other hand, have much more relaxed timing requirements. Event location, interarea oscillation modal analysis, data historians, and associated visualizations are some examples of non-real-time applications. 2.5.4 cloud comPutIng and Fnet Fundamental to the FNET is the ability to process and analyze measurements from FDRs in an eficient and effective manner. There is an inherent communications requirement even before data arrive at a data center for processing. Timely delivery of data with an adequately high availability is critical for an accurate depiction of the system state. 2.5.5 raPIdly elastIc data concentrators The data concentrator layer is pivotal to the underlying FNET applications, since it acts as the main interface to the data center for the incoming phasor measurements. High availability and high consistency are two key data concentrator properties 28 Smart Grids that ultimately dictate the accuracy of the system state representation for the FNET applications. One way to ensure high availability is by making the data concentrators fault-tolerant through replication. When data concentrators are suficiently replicated, it also becomes important to make sure that the replicas are consistent with each other. Inconsistent replicas adversely affect the underlying applications by forwarding out-of-sync or stale measurement records. On the measurement side, the granularity of the sensor measurements can rapidly scale up and down based on speciic grid conditions. Thus, the data concentrators must be capable of absorbing large quantities of measurement data arriving at high rates when the grid is stressed as well as when it is operating under normal conditions. This requires an underlying computational infrastructure that can rapidly scale on demand, which is an inherent feature of the cloud [5]. 2.5.6 comPutatIonal requIrement FlexIBIlIty The multilayer computational data center hierarchy has different computational needs at different layers. The outermost data concentrator units are mostly preprocessing units (aligning data as GPS–time-synced records) that require more primary memory than secondary memory. Preprocessed records are streamed immediately to real-time applications and are only buffered when the number of records reaches the memory cache limit. Large memory caches result in faster processing times. Data collection and record padding are not interrupted by the saving procedure, because the memory cache is designed as a stack; only the oldest portion of the data is saved. On the contrary, FNET applications have higher CPU and processing requirements than those of data concentrators. Furthermore, different applications can have different coniguration requirements (number of CPUs, memory, storage, network, etc.) based on the class of the application (real-time or non-real-time) and the type of spatial scope of the data that are being handled. Clearly, given the large capital investment, skilled workforce requirements, and maintenance cost, supporting such diverse computational requirements using in-house solutions is not cost effective to any power grid entity. A much more feasible solution would be to take advantage of cloud computing facilities, which can provide any type of platform or infrastructure requirement as a service, greatly improving the overall value proposition. 2.5.7 cloud-Based aPPlIcatIons The most intriguing aspect of using cloud computing in the FNET is supporting FNET applications as SaaS [5]. The FNET provides a rich pool of applications for power grid entities at all levels, both spatially and temporally. Prebuilt applications can be hosted as SaaS for interested grid entities to use at any scale, based on their speciic and individual organizational requirements. Furthermore, the cloud can provide a PaaS model for application developers to develop new applications under a uniied reference model. The data delivery to these applications is already handled by the data concentration level and can be offered as a well-deined API. This not only greatly Power Application Possibilities with Mission-Critical Cloud Computing 29 simpliies the instrumentation aspect of application development, but also helps to standardize the application development life cycle and data processing pipeline. REFERENCES 1. B. Qiu, L. Chen, V. Centeno, X. Dong, and Y. Liu, Internet based frequency monitoring network (FNET), in Proceedings of the IEEE Power Engineering Society Winter Meeting, vol. 3, pp. 1166–1171, 28 January–1 February 2001, Columbus, OH. 2. Y. Liu, A US-wide power systems frequency monitoring network, in Proceedings of the IEEE PES Power Systems Conference and Exposition (PSCE‘06), 29 October–1 November 2006, Atlanta, GA. 3. Power Information Technology Lab, FNET Web Display, University of Tennessee 2013. Available at: http://fnetpublic.utk.edu/sample_events.html. 4. Y. Zhang, P. Markham, T. Xia, L. Chen, Y. Ye, Z. Wu, Z. Yuan, et al., Wide-area frequency monitoring network (FNET) architecture and applications, IEEE Transactions on Smart Grid, 1(2), 159–167, 2010. 5. P. Mell and T. Grance, The NIST Deinition of Cloud Computing: Recommendations of the National Institute of Standards and Technology, The National Institute of Standards and Technology, 2011, Gaithersburg, MD. 2.6 OSCILLATION MITIGATION STRATEGIES Venkatasubramanian Real-time oscillation monitoring tools such as the oscillation monitoring system (OMS), developed at Washington State University based on synchrophasors, are being implemented in many control centers all over the world. Excellent progress has been made recently on centralized [1,2] and distributed algorithms [3] for widearea oscillation detection and analysis. Typically, the monitoring tools issue alarms to the operators whenever the damping ratio of any of the local or interarea modes goes below a preset threshold, say 2%. Unbounded growth of negatively damped oscillations can lead to tripping of critical system components and may even lead to blackouts such as the August 10, 1996, western American blackout. The presence of poorly damped or sustained oscillations over a period of time can cause damage to rotor shafts and also impact on consumer power quality, with potentially signiicant economic consequences. Therefore, it is important to mitigate poorly damped oscillations as quickly as possible. However, the choice of what control actions should be used to mitigate different types of oscillatory modes is a nontrivial task. As real-time oscillation monitoring tools are adopted by the industry, there is an urgent need to develop mitigation strategies whereby operators can respond to oscillation alarms in the form of speciic operator actions whenever such oscillation alarms occur. Cloud computation is an ideal platform for developing such mitigation control strategies. The problem can be handled in two ways: 1. Baseline studies: Massive analysis of historical synchrophasor data can be carried out using oscillation monitoring engines such as OMS, and the results can be recorded together with corresponding supervisory control and data acquisition (energy management system) data. Correlation studies 30 Smart Grids can be conducted to identify patterns in how different modes are sensitive to different tie lows and generation patterns. Cloud computing enables analysis of different groups of synchrophasor data in parallel to target different interarea modes, and, possibly, all the observable local modes of relevant generators. 2. Online simulations: Whenever the oscillation engine issues an alarm, cloud resources can be called on to carry out a parallel simulation of a large number of feasible candidate operator actions to decide which actions are the most effective in mitigating the speciic mode-related alarm. For instance, a critical tie-line interface may be identiied, and the necessary reduction or increase of tie-line low can be determined based on the simulation studies to improve the damping of a problematic interarea oscillatory mode. REFERENCES 1. G. Liu, J. Quintero, and V. Venkatasubramanian, Oscillation monitoring system based on wide area synchrophasors in power systems, in Proceedings of the 2007 IREP Symposium Bulk Power System Dynamics and Control, VII, 19–24 August 2007, Charleston, SC. 2. V. Venkatasubramanian and G. Liu, Systems and methods for electromechanical oscillation monitoring, US Patent 8,000,914, 2011. 3. J. Ning, X. Pan, and V. Venkatasubramanian, Oscillation modal analysis from ambient synchrophasor measurements using distributed frequency domain optimization, IEEE Transactions on Power Systems, 28(2), 1960–1968, 2013. 2.7 AUTOMATIC NETWORK PARTITIONING FOR STEADY-STATE ANALYSIS Zhu Cloud computing is emerging as a new paradigm for contemporary computational tasks with explosive demand for computing resources [1]. It offers thousands of computers instant and affordable access to powerful remote computing platforms. Cloud computing resources can match up to the highest instantaneous demand peaks, while providing lexibility on a pay-by-use basis when the demand peaks are over. Thanks to its attractive computational abilities, cloud computing is advocated as a great opportunity for diverse power system computational tasks, ranging from operations and control to economics and planning [2]. One key factor to enable cloud computing for the power grid would be an eficient network partitioning scheme. Typically, the interconnected bulk electric system is divided into several zones, based on speciic applications. Perhaps the most obvious example is the division of North America into Regional Transmission Organization territories for market administration and transmission management purposes [3]. Partitioning the grid into zones with minimal intercoupling serves as a basis for eficiently paralleling computational tasks, since it leads to more succinct information exchange among the computers. However, with the growing size of the power grid, an Power Application Possibilities with Mission-Critical Cloud Computing 31 exciting future direction would be to develop an automatic tool to partition any electric network depending on the level of granularity as well as the application domain. Beneiting from eficient parallelism, cloud computing can be immediately applied to power system steady-state analysis, which is routinely performed in all control centers. In this realm, cloud computing with conidentiality is suggested in [4] for the optimal power low problem—a crucial power system operational task. Nonetheless, it would be more interesting to perform the security analysis when the grid is approaching or reaching stressed conditions, leveraging the increase in number of processors from the existing order of dozens to thousands. In fact, the potential of parallel high-performance computing has been preliminarily investigated in [5] for massive contingency analysis on the Western Electricity Coordinating Council power grid. Nonetheless, it is not possible to apply cloud offerings directly to the power grid without some further considerations. For example, the GridCloud architecture project (see Chapter 1), developed by Cornell and Washington State University researchers, aims to keep the throughput offered by the commercial cloud offerings while eliminating a deliberate trade-off of consistencies. This feature may not seem necessary for applications such as web browsing. However, it is of great importance for power system computational tasks that involve storing replicas of, for example, some state variables on different computers. Another issue with the cloud offerings is that they can take many minutes to ramp up. This motivates the power grid control centers to seek close cooperation with a cloud provider. As another option, separate cloud installations for power grids may be more viable in this regard. For example, RTE in France has its own data center, and multiple ISOs in the eastern US grid are reportedly giving serious thought to developing their own GridCloud pilots in 2014. REFERENCES 1. K. Birman, Guide to Reliable Distributed Systems: Building High-Assurance Applications and Cloud-Hosted Services, 2012, Springer, London. 2. G. Dán, R. B. Bobba, G. Gross, and R. H. Campbell, Cloud computing for the power grid: From service composition to assured clouds, in Proceedings of the 5th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud), June 2013, San Jose, CA. 3. North American Electric Reliability Corporation (NERC), 2013. Available at: http:// www.nerc.com/. 4. A. R. Borden, D. K. Molzahn, P. Ramanathan, and B. C. Lesieutre, Conidentialitypreserving optimal power low for cloud computing, in Proceedings of the 50th Annual Allerton Conference on Communication, Control, and Computing (Allerton), 1–5 October 2012, Monticello, IL. 5. Z. Huang, Y. Chen, and J. Nieplocha, Massive contingency analysis with high performance computing, in Proceedings of the IEEE Power and Energy Society General Meeting, 26–30 July 2009, Calgary, AB. 3 Emerging Wide-Area Power Applications with Mission-Critical Data Delivery Requirements Greg Zweigle CONTENTS 3.1 3.2 3.3 3.4 3.5 Introduction .................................................................................................... 33 System Analysis .............................................................................................. 35 Situational Awareness ..................................................................................... 37 Stability Assessment ....................................................................................... 38 Wide-Area Control ......................................................................................... 39 3.5.1 Automatic............................................................................................ 39 3.5.2 Operator Initiated ...............................................................................40 3.6 Transient Instability Mitigation ...................................................................... 41 3.7 System Protection ...........................................................................................46 3.8 Summary Tables ............................................................................................. 48 References ................................................................................................................ 50 3.1 INTRODUCTION Our electric power system plays a critical role in present-day society. The many beneits of electricity include its versatility, reliability, safety, and low cost. In terms of versatility, when new sources of energy such as wind or solar are considered, the best method to transmit their power to consumers is by converting it to electricity. Other transmission options, for example, generating a hydrogen gas intermediate, are signiicantly more complicated and expensive to implement. In terms of reliability and economics, the interconnected electric power system has demonstrated exceptional robustness over many decades of operation and with very low cost. Worldwide, electric power is responsible for delivering approximately one-third of overall requirements, and the percentage of total power consumption in the form of electricity is growing due to trends in the electriication of automobile transportation and increases in the computing and communication infrastructure [1]. Because of its importance to society, there is a continuous need to improve the reliability and economics of electric power. Key to this is maintaining correct operation 33 34 Smart Grids during challenging events, including equipment failure, electrical faults, and any weather-induced variability affecting renewable generation sources. Hierarchies of protection and control, starting at a local distributed level and extending to a wide-area centralized level, are applied to keep the system operating in acceptable ranges during times of stress. At the local level, devices for equipment protection are deployed to quickly disconnect devices that develop a fault. This isolates failures and prevents disturbance spreading. When the fault is temporary, protection devices reclose the line back into service. This keeps the system strong against future disturbances. Equipment protection primarily utilizes local signals to make decisions. Recently, there is a move toward building systems that communicate between neighboring devices in a distributed manner to aid selectivity, security, and dependability. At the top, most centralized level, human operators monitor and control the system centrally. They receive signals that are communicated over a wide geographical area. Between the local and top levels of observation, a variety of distributed control systems are applied for responding to situations that require wide-area information but must operate with intervals that are too short for a human to process and make a good decision [2]. Modern electric power systems are designed to connect generators and consumers of power over large distances. This creates a unique engineering challenge for reliable operation, because of the communication demands that result when state measurements need to be shared with mission-critical applications for proper control and operation. The electric power system, or a speciied subset, can span hundreds or even thousands of miles. Measuring, transmitting, and verifying measurements collected over these long distances require a sophisticated communication infrastructure. The distances involved are particularly challenging for control of electric power systems. Most other control systems, such as cell phone receivers or automobile engine regulators, are packaged within relatively small spaces. Communication within these systems is fast. Long-distance communication networks can result in long latencies, dificulty in aligning measurements, and unreliability in reception quality. Control systems are particularly sensitive to these impacts. Historically, engineers have overcome the challenge of geographic distance for power system measurements by only measuring quantities that change slowly. Both the voltage magnitude and injected power in a transmission system vary slightly over an interval of several seconds. This allows for larger communication delays and simpliies aligning the measurements in time. Limiting measurements to those that change slowly is acceptable for many applications, but, in general, it is a problem because the state of the power system consists not only of the magnitude but also the phase angle. Phase angles change rapidly. Furthermore, depending on the type of control required, the required measurements can also include topology, synchronous machine rotor angle, and even the internal conditions of governors, voltage regulators, and other local control devices. In the past, and in many systems even today, the fast-changing phase angle is estimated at rates as slow as every minute, and is treated on a relative basis. In addition to providing angle values, estimators calculate missing data from physical locations where direct measurement is impractical or for cases when communication system issues lead to data interruption. These estimators also improve the accuracy of measurements. State estimation was previously based on a nonlinear method, driven by the need to derive angles from power measurements [3]. Because of its nonlinearity, Wide-Area Power Applications with Mission-Critical Data Delivery 35 the traditional state estimation is computationally expensive and can sometimes fail to converge. This leads to intervals when critical data are unavailable, not because of communication failures but because of algorithmic limitations. Convergence problems have become a larger issue recently because renewables result in a violation of the assumption that the power system is changing slowly. Good estimates are needed even during times of fast dynamics. Over the past several decades, a new measurement approach has evolved. A shared, high-accuracy time source enables measuring devices to precisely timestamp signals before transmitting them over a communication network. One common method of receiving this accurate time is from the global positioning system (GPS). When measurements are the network voltage and current related states, they are called synchrophasors [4]. More recently, this measurement approach has also been applied to rotor angles and other connected machine and device internal states [5]. The time-stamped measurement technology has unlocked new ways to control electric power, increase its reliability, and decrease its cost. Time-synchronized systems now directly measure voltage and current phase angles. This enables the application of a linear calculation for illing in missing values [6]. The linear calculation does not have convergence issues. All values, whether directly measured or calculated, are available during times of both slow and fast system dynamics. Also, with these measurements simple applications become possible, such as a textbased meter command at the measuring devices to help an engineer determine whether line phases have been swapped [7]. Direct measurements also provide the means for independent topology checking at the substation level [8] and enable the state estimation to be distributed [9], which results in a more reliable system. This chapter overviews new or signiicantly improved power system applications that are now possible due to wide-area time-synchronized measurements. The emphasis is on those applications with mission-critical delivery requirements. The ability to communicate information over large distances and meet quality of service constraints is facilitated by the advanced communication systems and distributed computing architectures covered in other chapters of this book. Applications are divided into ive categories: system analysis, situational awareness, stability assessment, wide-area control, and system protection. Each area is surveyed to explain emerging applications and research directions. For wide-area control, one of the most dynamically advancing ields due to these new measurement techniques, a speciic application is explained in full detail. Common data delivery requirements for each category are summarized, and special focus is given to interaction with the communication subsystem. In general, the categories are listed in order of increasing performance requirements. These requirements then become the basis for communication infrastructure design and analysis, as described in Chapter 4. 3.2 SYSTEM ANALYSIS Wide-area time-synchronized measurements provide a new tool for power system analysis. Figure 3.1 shows an architecture suitable for collecting this information, acting as a total system data recorder. Buses are labeled 1, 2, …, N, each representing a separate substation. At each substation a phasor data concentrator (PDC) 36 Smart Grids 1 N 2 PDC PDC PDC Communication network Data recorder FIGURE 3.1 Data collection architecture. buffers the measurements [10]. The purpose of the PDC is to provide an intermediate storage location at the substation, which is accessed in the case of interrupted communication to the central data recording device. The PDC also provides an additional security layer between the substation and the external communication network. Layered security is an important strategy for achieving overall system resilience against cyberattacks. In some implementations the PDC time-aligns the measurements and then sends all measurements for a given time instant together as a single communicated entity. While phasor data concentrator is a common term in North America, others might refer to this intermediate communication and computation device by another name, and additional capabilities and features are possible with more complicated data delivery architectures. However, the term phasor data concentrator is retained here because of its common usage in the industry. One application of the collected data is fault recording. The electric power system is very reliable and major events are fortunately rare. However, when an event occurs, getting to the root cause is important. Also, although minor events are common and have little impact on power system customers, gaining information from minor disturbances helps improve the system and avoids major disturbances. Timesynchronized measurements enable event reports to be combined into a single view of what happened and how the disturbance evolved over a wide area. Measurements are received from a variety of devices, such as digital fault recorders, protective relays, and automation controllers, and are then communicated through the PDC to a centralized location, as shown in Figure 3.1. These measurements might be sampled at different rates, but with high-accuracy time stamps it is simple to combine them into a single time base. Another application of these archived data is model veriication [11,12]. Good models are the prerequisite for effective power system planning and it is important to Wide-Area Power Applications with Mission-Critical Data Delivery 37 validate models against measured data to verify their accuracy. Time-synchronized measurements provide a way to combine data from various parts of the electric power system and apply this information to test models [13]. Often, these data are from relatively small events, such as the frequency response of the system after a generator trips. Also, when load tap-changing transformers make a step change, the time-stamped measurements help identify load characteristics. Because these new time-synchronized systems directly measure voltage and current angles, they provide a baseline state when the system is in the normal operating range. This enables power low models to be veriied. With respect to the communication system and data delivery requirements, collecting events puts almost no restriction on latency. Receiving the complete set of events with as much as a minute of delay is acceptable. This means that, if the overall communication is interrupted, it is possible to retrieve data buffered at the PDC at a later time, after the communication system has been restored. However, a high sample rate is required and the volume of data is large. Present systems sample as fast as 60 times per second, but in the near future sample rates on the order of 240 samples per second will become common. Higher sample rates are important for capturing and analyzing details of the power system’s response to disturbances. 3.3 SITUATIONAL AWARENESS Operators continuously monitor the electric power system to help maintain reliability. Existing situational awareness is built on the supervisory control and data acquisition (SCADA) architecture, which is an asynchronously scanned system with update rates on the order of seconds. Slowly changing voltage magnitude and power low measurements are adequately represented with this architecture. However, when the power system changes rapidly, as is becoming more common with the increase in wind and solar renewable energy sources, then these measurements become suspect during transients. Over the scan interval there are no guarantees that the measurements are acquired at the same time, and this causes problems. For example, consider the case of power measurements on parallel lines and the need to know the total power low. This is necessary for assessing the stability of the Guatemalan power system [14]. During fast dynamics, the power on each line can transition from nominal levels to levels requiring action in seconds. With a multisecond SCADA update rate, it is not possible to correctly calculate the total power at a 1 s interval. A solution through the application of time-synchronized measurements is described later in this chapter. Also, SCADA displays cannot effectively show phase angles and fast-changing machine state values because of the asynchronous scan. New time-synchronized systems update as frequently as 60 times per second. All measurements are taken at a common time instant. The angles are directly measured and displayed. Operators obtain a more detailed understanding of the system state [15–19]. Additionally, these details improve the ability of alarming and event notiication applications to provide information based on sophisticated assessments [20]. The important beneits of the new information include increasing the conidence of operators. They know that they are seeing the exact power system conditions in 38 Smart Grids 1 N 2 NI NI NI Communication network Operations FIGURE 3.2 A representative situational awareness architecture. real time. For example, if a line is taken out of service, the steady state as well as the transient result on the voltage and current level is immediately visible. If the system exhibits any undesirable characteristics, such as sudden oscillations, the operator sees them immediately and can take action. As another beneit, standing line angles across disconnected lines are directly measured. The lack of these measurements has posed problems historically. This was a contributor to the Southern California outage in September 2011 [21]. A line was open and the adjusted power lows resulted in a large angle developing. The operators could not always see this angle and take the actions to get this line back into service quickly. With time-stamped information the phase angle across each line is directly measured and displayed. Figure 3.2 shows an architecture suitable for situational awareness and also operational control. In contrast to Figure 3.1, communication paths are bidirectional because of the need for operators to issue control actions. These might include opening a breaker or redispatching generation power. The PDC is replaced with a network interface. There is no need for a signiicant amount of intermediate data buffering because data are needed as fast as they are generated. Going back at a later time to retrieve data after a network outage provides little beneit. Communication requirements for operator displays are driven by the need for near real-time streaming data. Some amount of latency is tolerated because humans have a limit on their ability to respond. Latency on the order of seconds is acceptable. However, the update rate should remain on the order of 30–60 samples per second because this can show higher frequency oscillations and other detailed information of interest. 3.4 STABILITY ASSESSMENT Assessing the electrical security of a power system is always an important need. These assessments relate to several classes of possible instability: voltage, frequency, Wide-Area Power Applications with Mission-Critical Data Delivery 39 transient [22], and oscillatory. With high-rate, streaming, time-synchronized measurements, the accuracy and prediction power of security assessments are improved [23–25]. For example, SCADA systems, with asynchronous update rates, do not allow ixed-rate signal processing applications. With the new time-synchronized systems and direct state measurement, the assessment algorithms can use well-known iltering and frequency-domain processing algorithms. For voltage stability assessment, as an example, a simple method is the calculation of stability indices based on reactive power measurements [26]. The voltage stability index is based on the difference between the maximum value that the system can support and the present value. QVmargin ( ti ) = Qcollapse ( ti ) − Qmeasured ( ti ) (3.1) The reactive power Qmeasured is collected at each bus at time instant ti. Then, the inductive loading at each bus is incremented by a real-time simulation until the system becomes unstable. This is the estimated value Qcollapse. The difference is the reactive power margin at that time. When QVmargin crosses a deined threshold, the system is assessed as becoming unable to handle the reactive power demands. It is important that all measurements are taken at the same time to improve accuracy and to ensure that measurement artifacts are not causing artiicial convergence issues. More sophisticated methods utilizing time-synchronized measurements have also been developed [27,28]. Once a system is assessed as becoming unstable, the information is provided to the operators, who can then adjust reactive power or generation dispatch to alleviate the situation. 3.5 WIDE-AREA CONTROL A wide range of diverse applications fall under the category of wide-area control. They are either distributed and automated or centralized and operator initiated. All of these control applications are based on a directly measured time-synchronized system state. This section overviews several of these applications and then examines in detail an emerging model-based approach suitable to mitigate transient instability. 3.5.1 automatIc An example of a distributed automatic application is collecting voltage measurements to increase the control coordination optimality of a set of static var compensators (SVC). The SVC is a reactive power device that acts in a fast manner to improve the voltage proile of the power system. These devices monitor voltages locally and inject reactive power to keep that local voltage stable and within preset ranges. Using the streaming nature of time-synchronized measurements, there are systems in service that also receive more distant measurements and make control adjustments that not only correct for local variations but also hold voltages at remote locations within proper boundaries [29]. Another wide-area control application is interarea oscillation damping [30–32]. These oscillations are driven by rotating machine controls interacting with each 40 Smart Grids other through the electrical coupling of the power system. Again, using the streaming nature of time-synchronized measurements results in a system in which remote oscillations are measured and then corrected at the point of initiation. Islanding control for distributed generation is a challenging problem that has beneited from direct state measurements. Present standards require distributed generators to stop producing power when their region is disconnected from the main power system [33]. A key challenge is detecting this disconnected state. Currently, many systems make this assessment through measurement of breaker status. When prespeciied breakers indicate that they have opened, a controller determines that the distributed generation site has islanded and sends disconnection commands. The problem with this approach is its complexity. Power systems can consist of many possible connection arrangements, so the controller must track all of them in determining the islanded state. Also, when connections are reconigured it is required that the controller be reconigured to relect this new arrangement. Another approach is for each distributed generator to continuously push its synthesized frequency against the bulk power system received frequency. The connected bulk power system is much stronger than the distributed generator, so when the local system is not islanded the attempted frequency perturbation is unsuccessful. When the system islands, the frequency rapidly changes and this indicates the islanded condition. The problem with this approach is coordinating the various frequency forces. In contrast to these existing methods, the comparison of measurements across a point of interconnection has both simpliied the problem and resulted in more reliable controllers [34,35]. Time-stamped measurements allow the direct comparison of the phase angle at various points of the power system. When the region containing distributed generation goes into an islanded condition, the phase angle differences, along with their rate of change, provide a strong indication of islanding. The controller then sends appropriate commands for disconnection of the distributed generators. In the future, this same scheme will allow control over the real and reactive power output of distributed generators in response to instantaneous power system conditions. So, as utilities install the more advanced time-stamped schemes to meet present interconnection requirements, these same systems become ready for more advanced applications. It is required that communication of data for automatic wide-area control is fast and reliable. These applications tend to be closed loop, and latency directly affects the performance and stability of the loop. Latency must be low. Speciic values are quantiied at the end of this chapter. 3.5.2 oPerator InItIated The precise time nature of the modern power system enables accurate timing not only of measurements, but also of control actions. Consider, for example, the scenario of an operator removing a line from service. First, commands are sent to open the breaker on each side of the line. The result is that the voltage decreases because of a higher effective impedance. Local tap-changing transformers and SVCs automatically adjust to correct the voltage. Meanwhile, the operator can switch manually operated reactive power devices, such as shunt capacitors, into service. There is a signiicant time delay between automatic local controls and operator-initiated Wide-Area Power Applications with Mission-Critical Data Delivery 41 controls. As a result, the system voltage and power lows experience transients. Once an operator has put shunt capacitors into service, the local control devices react to return more closely to their original state. The various signal level changes add variability to the power system, which is already experiencing larger signal swings due to renewable generation. As control devices respond to varying voltage levels, this adds stress and reduces their life span. Operators can now generate recipe lists of commands, each with an associated time stamp [36], and execute all required changes simultaneously. These lists are distributed to the various control actuation points in the system. When a certain operation is required, for example, removing a line from service, the collection of distributed recipes all execute in lockstep. This prevents transients and reduces system stress. Other operator control applications include black-starting generators and monitoring line angles prior to reinserting a line into service. In these applications it is primarily the measured angles that provide the needed capability. Operator-assisted controls tend to have less strict communication requirements than the feedback controls. For example, with recipes, it is possible to send the commands seconds, minutes, or even hours ahead of the control execution time. This allows for feedback testing to ensure that the requested commands are validated, thus increasing the cybersecurity of the system. 3.6 TRANSIENT INSTABILITY MITIGATION The electric system interconnects large synchronous machines that provide the bulk of generated power. During normal operation, the rotors of these machines operate in coordination and maintain slowly varying angle differences. However, during large contingencies it is possible for power imbalances to disturb the rotor angle relationships beyond the capability of restoring forces. In these cases a set of machines can swing out of step with each other. This condition is called transient instability. A typical initiating condition is a fault that signiicantly unloads a section of the system, which causes the affected machines to accelerate past the point of stability. Modern, fast power system protection has provided signiicant resilience against this problem [37]. However, there are still cases in which transient instability can develop. These are often high-order contingencies in which multiple lines are lost, often in conjunction with the loss of primary protection and the resulting delay of backup protection causing longer fault duration; they cannot adapt to the present conditions. System integrity protection schemes, also called remedial action schemes, can provide a certain level of resiliency against transient instability. However, they are limited to the speciic conditions studied and must be conigured for these cases in advance of the disturbance. With time-stamped measurements it is now possible to build feedback control schemes that measure a wide-area collection of power system states and respond to a much larger range of contingency cases than is currently possible. A speciic example that demonstrates these ideas in detail is presented next. Figure 3.3 shows a block diagram of the controller, focusing on three buses out of the larger power system. Each device controller (DC) takes measurements of the network state and, when connected to a generator bus, the machine state. The net work state consists of time-stamped voltage V j k and current I j k measurements 42 Smart Grids Controller Communication network DC Machine controller Machine controller DC Other sections of power system DC FIGURE 3.3 Architecture for mitigation of transient instability. at each bus in the power system. Buses range over j = 1, …, N. The machine state consists of the rotor angle θi[k], rotor ield excitation Efd,i[k], and prime mover power Pm,i[k]. Machines range over i = 1, …, M. The index k represents the time stamp. For transient stability control a suitable measurement sample rate is fnom, meaning that the time stamps are incremented at an interval of T = 1/fnom. This is the rate at which data are transmitted over the communication network to the controller. Although currently not all installed systems are able to acquire such detailed machine measurements, this technology is becoming more viable. Time-stamped rotor measurements are included in the appendix of the latest Institute of Electrical and Electronics Engineers (IEEE) C37.118 standard [38] and are implemented today in some advanced controllers [39]. Furthermore, time-stamped measurements of the additional synchronous machine state values are under investigation [40]. The controller monitors the power system dynamics and determines the need for control actions. For transient instability correction, the control actions selected here consist of tripping generation, shedding load, or inserting series compensation on select lines. Other control actions are possible. Time-stamped received signals are processed and controls are selected through the methodology of model-driven assessment and performance metric measures [41,42]. The electric power system is nonlinear by nature. This makes controller designs particularly challenging. Although linearization or model-free approaches are possible, the strong nonlinearities that characterize rotor angle–related instability make a model-based controller attractive. Figure 3.4 shows the basics of operation. At time tF the power system experiences a large contingency that, for example, involves faults on multiple lines. Local protection devices clear these faults at time tT. Subsequently, it is the task of the controller to determine whether the state trajectories are evolving toward stability and whether control actions are required. If control actions are required, then it is the task of the controller to select an appropriate control and send it to the appropriate DC. These tasks are performed by predicting the state trajectories according to a model of the power system. The initial state is from the measured time-stamped state values. Then, for each control option, the state trajectories are predicted. A cost measure is applied to the resulting trajectories. The irst control action for the sequence of control actions that results in a minimal cost is selected for application. Because the option of not applying any controls is included in the Wide-Area Power Applications with Mission-Critical Data Delivery Evolution 43 Modeling Predicted trajectory 1 Predicted trajectory 2 Reference trajectory State Predicted trajectory for no controls tF tT Control 1 Control K FIGURE 3.4 Basic operation of a model-based controller with prediction. considered list, the algorithm inherently assesses stability. When the system is not evolving toward instability the controller takes no action. The cost measures consist of both the state cost, Cstate, and the control cost, Ccontrol. Each is calculated individually based on the received state measurements and a prediction of the future state. The total cost is the product of these two individual costs. J = CstateCcontrol (3.2) The state cost Cstate is the summation of the individual cost for three classes of states. These are the rotor angle Cθ, frequency Cω , and bus voltage magnitude CV. (3.3) Cstate = Cθ + Cω + CV Individual state costs in turn are calculated as follows. Equation 3.4 provides the cost measure for the rotor angle, following a mean-square formulation. The center of inertia angle θɶ i [ k ] = θi [ k ] − ΣiM=1 Hi θi [ k ] ΣiM= 0 Hi, where Hi is the inertia for machine i, at time kT is compared against the average of the center of inertia angle θɶ i over window W. This difference is normalized by the maximum expected angle, π, and then squared. The duration of the prediction window is WT seconds. The resulting cost measure for the rotor angle of machine i is deined as Cθ′ ( i ). A large value of Cθ′ ( i ) indicates that the rotor angle state for machine i has deviated signiicantly from a possible stable path. 1 Cθ′ ( i ) = 2 W θɶ i k − θɶ i π k =1 W ∑ 2 (3.4) The total cost for all of the rotor angles in the complete power system is the summation of the individual machine rotor angle costs. M Cθ = ∑C′ (i ) θ i =1 (3.5) 44 Smart Grids For frequency, the rate of change of the rotor angle ωi[k] is compared with the nominal frequency ωnom and normalized by a frequency limit, ωlimit. This quantity is then summed over the prediction window. 1 Cω′ ( i ) = 2 W ωi k − ωnom ωlimit k =1 W ∑ 2 (3.6) The total cost of the frequency for all machines is the summation of the individual machine frequency costs. M Cω = ∑C′ (i ) (3.7) ω i =1 The inal state-related cost measure is based on the voltage magnitude, computed for each of the N network buses. The voltage is compared with the predisturbance equilibrium voltage for that bus. The difference is divided by the reference voltage, squared, and then summed over the prediction window. 1 CV′ ( i ) = 2 W Vi k − Veq,i Vmax k =1 W ∑ 2 (3.8) The total cost of the voltage magnitude is the summation of individual bus voltage costs, normalized by the ratio between the number of bus voltages and the number of machines in the system. M CV = N 2 N ∑C′ (i ) V (3.9) i =1 While the state costs are calculated with a mean-square approach, the control costs, Ccontrol, are taken from tabulated values. The total control cost is then a summation of the individual control actions, taken sequentially over the prediction window. In Equation 3.10, the number of controls considered over the prediction window is W′. W′ Ccontrol = ∑C j (3.10) h =1 Table 3.1 provides an example cost allocation for control actions. The table approach allows the selection of each value based on the individual needs of the application. In Table 3.1, two sets of costs are provided. When the cost (Equation 3.3) for the case without application of control is below a speciied threshold, then all tested controls use the values in the second column of Table 3.1. When the cost (Equation 3.3) is above this threshold, then the system is estimated as becoming unstable and the tested controls in the third column are applied. In either case, the inal selected control set is based on the sequence that minimizes (Equation 3.2), independently of whether the cost table selection is based on a stable or an unstable threshold check. 45 Wide-Area Power Applications with Mission-Critical Data Delivery TABLE 3.1 Control Action Costs Structural Change Control Action No control action Series capacitance Load shedding Generator tripping Base Case is Modeled Stable 1 2 4 3 Base Case is Modeled Unstable 4 1 3 2 The capability of this controller is demonstrated with the IEEE 39-bus system. Figure 3.5 shows the system response to an example of three lines simultaneously faulted. The upper left signals are the phase angles of each generator referenced to the center of inertia angle. The lower left signals are the frequency values, referenced to the nominal frequency. The right of the igure shows the voltage magnitudes for all 39 buses. Assume that the fault is cleared by backup protection with a long clearing time. The probability of losing three lines is very low. However, it is these low-probability events that can lead to large cascading outages. An advantage of applying timestamped measurements with wide-area control in a feedback architecture is the ability to respond effectively to both common and rare system events. Figure 3.6 shows the system response after correction by the controller. The controller receives time-synchronized measurements, estimates the response against a model for a set of control options, and then selects the control that minimizes that total cost. The result is that the controller selects tripping a generator 100 ms after the 4 6 × 10 1.5 2 0 –2 1 0 5 10 15 20 15,000 0.5 10,000 mHz Voltage (p.u.) Degrees 4 5,000 0 –5,000 0 5 10 Time (s) 15 20 0 0 5 10 Time (s) FIGURE 3.5 Unstable response of the 39-bus system to a three-line outage. 15 20 46 Smart Grids 1.5 100 Degrees 50 0 –50 1 0 5 10 15 20 800 600 Voltage (p.u.) –100 0.5 mHz 400 200 0 –200 0 5 10 Time (s) 15 20 0 0 5 10 Time (s) 15 20 FIGURE 3.6 Stable response after control application. fault is cleared, then inserting series capacitance 200 ms after the fault is cleared, and inally shedding some load 300 ms after the fault is cleared. The advantage of this method is that controls are selected based on minimizing a combined cost of state trajectory deviation and control action costs. This helps to ind a solution that minimizes the negative effect of the controls while simultaneously stabilizing the system. Automated control for resolving transient instability requires a very reliable communication network. As seen in Figure 3.3, measurements are collected and controls are sent over a wide area. Latency on the order of less than 10 ms is required for this application. This allows time for measurement iltering and control selection such that the irst control action can take place only 100 ms after the initiating event. 3.7 SYSTEM PROTECTION The most demanding application in terms of communication requirements is system protection. Time-synchronized measurements can now be found in line-current differential and distributed bus differential protection schemes [43]. For wide-area protection, new approaches to transient stability [44,45] and voltage control [46] have become available. A detailed example is the correction of oscillations using time-stamped measurements [14]. In Central America, a 230 kV transmission network connects from Mexico through Guatemala and south to Panama. The Central American system capacity is approximately 7 GW, which is signiicantly smaller than the power system in Mexico. Historically, oscillations in the Central American system have required disconnection from the Mexican system during certain intervals of the day. This negatively impacted the reliability and economics of the Guatemalan system. The wholesale operator of Guatemala, the Administrador del Mercado Mayorista Wide-Area Power Applications with Mission-Critical Data Delivery 47 (AMM), decided to implement a system that could detect these oscillations and proactively disconnect lines. This eliminates the need to isolate based on time of day and, instead, acts only when necessary. Thus, a measurement-based control system is in place instead of predeined table lookup actions. What has prevented such a scheme in the past is the lack of time-synchronized measurements. There are two factors involved. First, it is necessary to measure the power low on two separate lines in order to determine the state of the power low. While SCADA systems can measure and sum power when slowly varying, as is the case during normal operation, they cannot provide accurate results during fast changes. It is precisely during these fast changes that the measurements are required. Second, a ixed-interval sampling scheme is required for estimating oscillation modes. Figure 3.7 shows a block diagram of the system. The time-stamped voltage and current at each important line are measured with a phasor measurement unit (PMU). A PMU is a type of device that makes time-synchronized measurements and conforms to the IEEE C37.118 standard. This standard includes requirements on iltering, measurement accuracy, and communication frame layout. In the Guatemalan system, the PMU is included as part of a protective relay. The advantage of including a PMU internal to the relay is the availability of the control function necessary to act based on the oscillation state. Each PMU is connected to a clock. The clock receives a GPS signal, extracts the time component, and sends it to the PMU. The result is that all measurements are based on a common, highly accurate time reference. The voltages and currents are measured at a rate of 60 per second. They are then streamed over an Ethernet connection to the AMM control center and received by a synchrophasor vector processor (SVP), an advanced PDC. The lower section of Figure 3.7 shows the SVP algorithm in detail. Starting with the block on the left, irst the measurements are time aligned. This is possible because of their precise time stamp, which is included with each measurement. Substation 1 Substation N PMU Clk PMU Clk Wide-area Ethernet provider SA Time alignment Calculate power SVP Modal analysis FIGURE 3.7 Block diagram of protection system. Control selection 48 Smart Grids The control device includes an IEEE 61131 logic engine. With this capability, the power on each line is calculated from the received voltages and currents. Each calculation is at a speciic time instant. * P ( tk ) + Q ( tk ) = V ( tk ) ⋅ I ( tk ) (3.11) The resulting power values for speciied lines are summed in order to determine the total low in the system. These are compared with a threshold. The power values are also sent to a modal analysis engine that computes the oscillatory mode frequency, amplitude, and damping. When an oscillation is detected with negative damping, and it persists in the system for a suficient duration, the control selection logic determines an isolation command. This is sent back through the network to open the appropriate breaker. Local protection that shares time-synchronized measurements for applications such as line-current differential schemes must act with latencies on the order of power system cycles, which are tens of milliseconds. For wide-area protection and control the allowed latencies may be slower because the instability evolves over periods from 100 ms, for transient stability, to seconds, for voltage stability or oscillatory stability. 3.8 SUMMARY TABLES Table 3.2 presents communication requirements in a quantitative form with scaling to indicate the level of dificulty on a range of 1–5 [47]. Included in Table 3.2 are values of the following data delivery requirements: required latency, required rate, the criticality level, the quantity of data that is delivered, and the geographical area over which the data travel. No applications with less than a high level of criticality are considered in this chapter. Table 3.3 quantiies each of the previously described applications based on the ranges shown in Table 3.2. These data delivery requirements are referenced in Chapter 4. For system analysis an input latency on the order of seconds or longer is acceptable because these are off-line applications. The data rate requirement, however, is high because resolution is important when investigating system behavior. The criticality of the data is moderately important. Of course, it is never good to lose data, but for off-line applications the impact is mitigated by the possibility of collecting the data at another time. The quantity of the data can be high and data are often collected over wide geographical areas. There are no output requirements for analysis. Situational awareness requires less latency than system analysis, but the effect of human operators in the loop limits the beneit of latencies that are below approximately 1 s. The required data rate is moderate. Rates of up to 60 messages per second are useful, but higher rates probably provide detail that an operator is unable to fully utilize when monitoring the power system in real time. The criticality level of the information received by operators is high. A fairly large quantity of data is Wide-Area Power Applications with Mission-Critical Data Delivery 49 TABLE 3.2 Normalized Values of Parameters Dificulty (5 hardest) 5 Latency (ms) 5–20 Rate (Hz) 240+ Criticality Ultra Quantity Very high 4 3 20–50 50–100 120–240 30–120 Very high High High Medium 2 1 100–1000 >1000 1–30 — — — Low Very low (serial) Geography Across grid or multiple ISOs Within an ISO/RTO Between a few utilities Within a single utility Within a substation Note: ISO, independent system operator; RTO, regional transmission organization. TABLE 3.3 Diversity of Data Delivery of Selected Power Applications Inputs Outputs Latency Rate Criticality Quantity Geography Latency Rate Criticality Quantity Geography System Analysis 1 3–4 3 5 2–5 — — — — — Situational Awareness 2 2–3 4 3–4 2–5 1 — 4–5 1–2 2–5 Stability Assessment 3 2–5 4 4 2–5 — — — — — Wide-Area Control 3–5 3–4 5 2–4 1–5 3–5 2–4 5 1–3 1–5 System Protection 4–5 4–5 5 2–4 1–4 5 — 5 1–3 1–3 received, and over a wide area. The quantity of data is limited by human factors and above a certain point it becomes dificult to effectively utilize a large amount of data. Too much information for an operator, especially during times when rapid response is required, can make assessing and responding to the situation more dificult. The output communication requirements for situational awareness are constrained by the fact that these commands are issued by human operators. Therefore, the tolerated latency is relatively high. The data rate aspect of the output is not relevant. Operators issue single commands, not streaming instructions. Operator commands are critical. The data quantity of the output commands is low. The geographical coverage of the commands is the same as for input data. For stability assessment, requirements are fairly similar to those of the situational awareness application. However, the latency is stricter and the data rate is higher because often the initial consumer of the data is a computer algorithm, 50 Smart Grids which is able to process a more detailed information level than is possible for a human. There are no output requirements because it is assumed that an operator acts based on the output of the assessment. These output requirements are covered under the situational awareness column. However, stability assessment is also an intermediate stage of wide-area control and system protection. In these cases, the assessment requirements are part of the data delivery requirements for those applications. It is important that wide-area control latency is low because latency negatively impacts the stability of feedback control loops. Similarly, the rate is fairly fast. All control schemes are critical for correct power system operation, and this necessitates not only a reliable communication network but also a distributed architecture when possible. The quantity of data is large, but typically not as large as required for system analysis or situational awareness. This is because control schemes often only require a few signals, such as positive sequence voltage or positive sequence current. The geographical span of the data is dependent on the speciics of each control application and can range from every location within a substation all the way up to an entire power system grid. The output data for control also require low latency and fairly high rates. Like the received measurements, output commands have a critical reliability need. System protection places the most demanding requirements on latency, data rate, and criticality of the application. The quantity of data may be lower because typically only a few key signals are monitored. The geographic scale can cover a wide area but is also limited by the required latency and reliability. Output signals for protection are single acting, and this is why there is no speciication for rate. However, a low latency is required. REFERENCES 1. Annual Energy Outlook 2013, with Projections to 2040, US Energy Information Administration, 2013. 2. V. Madani, D. Novosel, S. Horowitz, M. Adamiak, J. Amantegui, D. Karlsson, S. Imai, and A. Apostolov, IEEE PSRC report on global industry experiences with system integrity protection schemes (SIPS), IEEE Transactions on Power Delivery, 25(4), 2143– 2155, 2010. 3. F. C. Schweppe and J. Wildes, Power system static-state estimation, part I: Exact model, IEEE Transactions on Power Apparatus and Systems, PAS-89, 120–125, 1970. 4. J. S. Thorp, A. G. Phadke, and K. J. Karimi, Real time voltage-phasor measurements for static state estimation, IEEE Transactions on Power Apparatus and System, PAS104(11), 3098–3106, 1985. 5. G. Zweigle, D. Finney, and R. Moxley, Adding shaft angle measurement to generator protection and monitoring, in Western Protective Relaying Conference, October 2012, Spokane, WA. 6. R. Zivanovic and C. Cairns, Implementation of PMU technology in state estimation: An overview, in Proceedings of the IEEE 4th AFRICON, vol. 2, pp. 1006–1011, 24–27 September 1996, Stellenbosch, South Africa. 7. E. O. Schweitzer III and D. E. Whitehead, Real-world synchrophasor solutions, in Proceedings of the 35th Annual Western Protective Relay Conference, October 2008, Spokane, WA. Wide-Area Power Applications with Mission-Critical Data Delivery 51 8. E. O. Schweitzer III and D. E. Whitehead, Real-time power system control using synchrophasors, in Proceedings of the 34th Annual Western Protective Relay Conference, October 2007, Spokane, WA. 9. T. Yang, H. Sun, and A. Bose, Two-level PMU-based linear state estimator, in Proceedings of the IEEE PES Power Systems Conference and Exposition, pp. 1–6, 2009, Seattle, WA. 10. SEL-3373 Station Phasor Data Concentrator, Instruction Manual, Schweitzer Engineering Laboratories, Inc., 11 March 2013. 11. G. Zweigle, R. Moxley, B. Flerchinger, and J. Needs, Wide-area measurements to improve system models and system operation, in Proceedings of the 11th International Conference on Developments in Power System Protection, DPSP, 23–26 April 2012, Birmingham, UK. 12. D. Shi, D. Tylavsky, N. Logic, and K. Koellner, Identiication of short transmissionline parameters from synchrophasor measurements, in Proceedings of the 40th North American Power Symposium, 28–30 September 2008, Calgary, AB. 13. Z. Huang, P. Du, D. Kosterev, and S. Yang, Generator dynamic model validation and parameter calibration using phasor measurements at the point of connection, IEEE Transactions on Power Systems, 28(2), 1939–1949, 2013. 14. J. V. Espinoza, A. Guzman, F. Calero, M. Mynam, and E. Palma, Wide-area measurement and control scheme maintains Central America’s power system stability, in Western Protective Relay Conference, 16–18 October 2012, Spokane, WA. 15. R. Moxley, C. Petras, C. Anderson, and K. Fodero II, Display and analysis of transcontinental synchrophasors, in Proceedings of the 6th Annual Western Power Delivery and Automation Conference, April 2004, Spokane, WA. 16. D. J. Trudnowski, J. W. Pierre, N. Zhou, J. F. Hauer, and M. Parashar, Performance of three mode-meter block-processing algorithms for automated dynamic stability assessment, IEEE Transactions on Power Systems, 23(2), 680–690, 2008. 17. J. M. Ordagci, H. C. T. Santos, S. R. Morand, R. Cespedes, R. Mano, and D. Caceres, ONS—Brasil new control center architecture conceptual design, in Proceedings of the IEEE/PES Transmission and Distribution Conference and Exposition: Latin America, 13–15 August 2008, Bogota. 18. G. Zhang, S. Lee, R. Carroll, J. Zuo, L. Beard, and Y. Liu, Wide area power system visualization using real-time synchrophasor measurements, in Proceedings of IEEE Power and Energy Society General Meeting, 25–29 July 2010, Minneapolis, MN. 19. J. Wang, Y. Hu, A. Johnson, H. Tram, and R. Nasri, System requirements of visualization platform for wide area situation awareness system, in 2010 IEEE Power and Energy Society General Meeting, 25–29 July 2010, Minneapolis, MN. 20. SynchroWAVe Central Software Instruction Manual, Schweitzer Engineering Laboratories, Inc., 1 April 2013. 21. Arizona-Southern California Outages on September 8, 2011—Causes and Recommendations, Federal Energy Regulatory Commission and the North American Electric Reliability Corporation, April 2012. 22. P. Kundur, Power System Stability and Control, McGraw-Hill, New York, 1994. 23. K. Sun, S. Likhate, V. Vittal, V. S. Kolluri, and S. Mandal, An online dynamic security assessment scheme using phasor measurements and decision trees, IEEE Transactions on Power Systems, 22(4), 1935, 2007. 24. Z. Wang, A. Giris, V. Aravnthar, and E. Makram, Wide area power system transient stability assessment using catastrophe theory method and synchrophasors, in 2011 IEEE Power and Energy Society General Meeting, 24–29 July 2011, San Diego, CA. 25. C. Zheng, V. Malbasa, and M. Kezunovic, Regression tree for stability margin prediction using synchrophasor measurements, IEEE Transactions on Power Systems, 28(2), 1978–1987, 2013. 52 Smart Grids 26. M. Donolo, M. Venkatasubramanian, A. Guzman, and F. De Villiers, Mitigating voltage collapse problems in the Natal region of South Africa, in Proceedings of the IEEE Power Systems Conference and Exhibition, 15–18 March 2009, Seattle, WA. 27. M. Glavic and T. V. Cutsem, Wide-area detection of voltage instability from synchronized phasor measurements. Part I: Principle, IEEE Transactions on Power Systems, 24(3), 1408–1416, 2009. 28. M. Glavic and T. V. Cutsem, Wide-area detection of voltage instability from synchronized phasor measurements. Part II: Simulation results, IEEE Transactions on Power Systems, 24(3), 1417–1425, 2009. 29. A. Johnson, R. Tucker, T. Tran, D. Sullivan, C. Anderson, and D. E. Whitehead, Static var compensation controlled via synchrophasors, in Proceedings of the 34th Annual Western Protective Relay Conference, October 2007, Spokane, WA. 30. Y. Gong and A. Guzmán, Synchrophasor-based online modal analysis to mitigate power system inter-area oscillation, in Proceedings of the DistribuTech Conference and Exhibition, 2009, San Diego, CA. 31. J. Chow, J. Sanchez-Gasca, H. Ren, and S. Wang, Power system damping controller design using multiple input signals, IEEE Control Systems Magazine, 20(4), 82–90, 2000. 32. M. Mokhtari, F. Aminifar, D. Nazarpour, and S. Golshannavaz, Wide-area power oscillation damping with a fuzzy controller compensating the continuous communication delays, IEEE Transactions on Power Systems, 28(2), 1997–2005, 2013. 33. IEEE Application Guide for IEEE Std 1547, IEEE Standard for Interconnecting Distributed Resources with Electric Power Systems, 2009. 34. J. Mulhausen, J. Schaefer, M. Mynam, A. Guzmán, and M. Donolo, Anti-islanding today, successful islanding in the future, in Proceedings of the 36th Annual Western Protective Relay Conference, 20–22 October 2009, Spokane, WA. 35. M. Mills-Price, M. Scharf, S. Hummel, M. Ropp, D. Joshi, G. Zweigle, K. Ravikumar, R. Moxley, and B. Flerchinger, Interconnection control of distributed generation with time-synchronized phasors, in IEEE Power Systems Conference and Exposition, pp. 1–8, 20–23 March 2011, Phoenix, AZ. 36. E. O. Schweitzer III, D. E. Whitehead, G. C. Zweigle, and K. G. Ravikumar, Synchrophasor-based power system protection and control applications, in Proceedings of the 36th Annual Western Protective Relay Conference, October 2009, Spokane, WA. 37. SEL-421 Relay Protection and Automation System, Instruction Manual, Schweitzer Engineering Laboratories, Inc., 22 February 2013. 38. IEEE C37.118.1–2011, IEEE Standard for Synchrophasor Measurements for Power Systems, 2011. 39. Q. Yang, T. Bi, and J. Wu, WAMS implementation in China and the challenges for bulk power system protection, in IEEE Power Engineering Society General Meeting, 24–28 June 2007, Tampa, FL. 40. P. Wang, W. Deng, and Z. Qian, Method for the measurement of rotor voltage and current of power generator, IET Generation, Transmission, and Distribution, 4(7), 854– 860, 2010. 41. G. Zweigle and V. Venkatasubramanian, Model prediction based transient stability control, in IEEE PES Conference and Exhibition, 7–10 May 2012, Orlando, FL. 42. G. Zweigle and V. Venkatasubramanian, Wide-area optimal control of electric power systems with application to transient stability for higher order contingencies, IEEE Transactions on Power Systems, 28(3), 2313–2320, 2013. 43. A. Guzmán, V. Mynam, and G. Zweigle, Backup transmission line protection for ground faults and power swing detection using synchrophasors, in Proceedings of the 34th Annual Western Protective Relay Conference, October 2007, Spokane, WA. Wide-Area Power Applications with Mission-Critical Data Delivery 53 44. D. Hu and V. Venkatasubramanian, New wide-area algorithms for detection and mitigation of angle instability using synchrophasors, in IEEE Power Society General Meeting, 24–28 June 2007, Tampa, FL. 45. E. Martinez, N. Juárez, A. Guzmán, G. Zweigle, and J. León, Using synchronized phasor angle difference for wide-area protection and control, in Proceedings of the 33rd Annual Western Protective Relay Conference, October 2006, Spokane, WA. 46. C. Taylor, D. Erickson, K. Martin, R. E. Wilson, and V. Venkatasubramanian, WACS— Wide-area stability and voltage control system: R&D and online demonstration, Proceedings of the IEEE, 93(5), 892–906, 2005. 47. D. E. Bakken, A. Bose, C. H. Hauser, D. E. Whitehead, and G. C. Zweigle, Smart generation and transmission with coherent, real-time data, Proceedings of the IEEE, 99(6), 928–951, 2011. 4 High Availability, Low GridStat Latency, and Adaptive Sensor Data Delivery for Smart Generation and Transmission David E. Bakken, Harald Gjermundrød, and Ioanna Dionysiou CONTENTS 4.1 4.2 4.3 4.4 Introduction .................................................................................................... 56 Requirements and Guidelines for Coherent Real-Time Data Delivery .......... 57 4.2.1 System Model ..................................................................................... 59 4.2.2 Delivery Requirements for a WAMS-DD ..........................................60 4.2.3 Implementation Guidelines for a WAMS-DD ....................................64 4.2.3.1 Context .................................................................................64 4.2.3.2 Implementation Guidelines .................................................. 65 4.2.3.3 Analysis................................................................................ 70 Analysis of Existing Technologies for a WAMS-DD ..................................... 71 4.3.1 Technologies and Standards at the Traditional Network Layers ........ 71 4.3.1.1 Very Inadequate ................................................................... 71 4.3.1.2 Better Yet Incomplete .......................................................... 75 4.3.2 Middleware Technologies and Standards ........................................... 76 4.3.2.1 Deinition of Middleware ..................................................... 76 4.3.2.2 Broker-Based Publish-Subscribe Middleware ..................... 77 4.3.2.3 Peer-to-Peer Publish-Subscribe ........................................... 78 4.3.2.4 Other Middleware ................................................................ 79 4.3.3 Existing Electric Sector Communications-Related Technologies and Standards .....................................................................................80 4.3.4 Summary ............................................................................................ 82 NASPInet ........................................................................................................ 83 4.4.1 Architecture ........................................................................................ 83 4.4.2 Service Classes ...................................................................................84 55 56 Smart Grids 4.5 Why WAMS-DD Must Include Middleware, Not Solely Network Layers .... 85 4.5.1 Middleware in General ....................................................................... 86 4.5.2 Middleware and QoS+ ........................................................................ 86 4.5.3 Middleware and DRs and IGs ............................................................ 87 4.5.4 Middleware and Legacy Systems ....................................................... 88 4.5.5 Middleware and Avoiding QoS Stovepipe Systems ...........................90 4.6 Security and Trust for WAMS-DD ................................................................. 91 4.7 GridStat...........................................................................................................92 4.7.1 Overview............................................................................................. 93 4.7.1.1 Capabilities .......................................................................... 93 4.7.1.2 More Detail .......................................................................... 93 4.7.1.3 Performance and Scalability ................................................ 95 4.7.2 Rate Filtering for Per-Subscriber QoS+ ............................................. 96 4.7.2.1 Forwarding Algorithm .........................................................96 4.7.2.2 Example of the Forwarding Algorithm ...............................96 4.7.3 Condensation Functions......................................................................97 4.7.3.1 Condensation Function Architecture and Capabilities ........ 98 4.7.3.2 Condensation Function Development ..................................99 4.7.4 Operational Modes .............................................................................99 4.7.4.1 Mode Change Algorithms....................................................99 4.7.4.2 Hierarchical Mode Change Algorithm .............................. 101 4.7.4.3 Flooding Mode Change Algorithm.................................... 102 4.7.4.4 Overview of Performance .................................................. 102 4.7.5 Systematic Adaptation via Data Load Shedding with Modes .......... 102 4.7.6 Remote Procedure Call..................................................................... 103 4.7.6.1 Mechanism Details ............................................................ 104 4.7.6.2 The 2WoPS Protocol .......................................................... 104 4.7.6.3 The Ratatoskr RPC Mechanism ........................................ 105 4.7.7 Security and Trust in GridStat .......................................................... 106 4.8 A New World for Power Application Programmers and Researchers.......... 107 4.9 Conclusions ................................................................................................... 109 Acknowledgments.................................................................................................. 109 Disclaimer .............................................................................................................. 110 References .............................................................................................................. 110 4.1 INTRODUCTION Electric power grids around the world are becoming increasingly stressed by factors including inadequate transmission growth, system operators and power engineers retiring in large numbers, and integration of renewable sources of energy whose physics is different from well-known sources such as hydro and coal. All these factors, and more, can be mitigated by greatly increased sharing of sensor data between utilities (and independent system operators [ISOs]/regional transmission organizations [RTOs]) in a grid, something that is extremely limited in power grids today. This is an important step in helping make grids more resilient, because inadequate situational awareness resulting from little interutility data sharing has been a major GridStat 57 contributing factor in virtually all recent major blackouts. In such blackouts, there are always a few physical or operational root causes that are blamed. However, due to the poor situational awareness, the grids’ operators are unaware of emerging problems until it is too late to prevent a blackout. Additionally, closed-loop applications such as distributed control and distributed protection are being increasingly deployed. However, their data delivery requirements (DRs) are the most extreme of any infrastructure in the world. Thus, it is crucial to ensure that data delivery for power grids is appropriate for the task at hand. Conversely, adopting more generic data delivery infrastructures from other industries, such as air trafic control, which seem supericially similar, can cause blackouts. Similarly, using network mechanisms such as multiprotocol label switching (MPLS) that provide very weak statistical guarantees and very crude control over the network (e.g., MPLS has only eight trafic categories) is also unwise. In this chapter, we describe GridStat, which is in essence an overlay network providing strong guarantees and specialized semantics for wide-area monitoring system data delivery (WAMS-DD). In conigurations where GridStat’s specialized routers do not have 100% penetration, they can be augmented by other mechanisms providing very strong network-level quality of service (QoS) guarantees in order to support closed-loop applications over the wide area. GridStat can also overlay (and help better manage and integrate) other networking subsystems with weaker QoS guarantees in a critical infrastructure. This is in essence a software-deined network (SDN). However, it is not a generic one: it is an SDN infused with semantics from both middleware and power grid applications and sensors that allow it to exert ine-grained control over its trafic. This not only provides very low latency and extremely high availability, but also allows the SDN to change its delivery patterns in a small fraction of a second to adapt to power anomalies, benign information technology (IT) failures, and cyberattacks. The remainder of this chapter is organized as follows. Section 4.2 describes baseline DRs that any WAMS-DD system must meet. It then derives 20 implementation guidelines (IGs) that those requirements mandate any WAMS-DD to use. Section 4.3 then presents a detailed analysis of how well existing technologies at the network layers, at the middleware layers, and from the electricity sector meet these requirements and guidelines. Next, Section 4.4 overviews NASPInet, a WAMS-DD initiative in North America. After this, Section 4.5 explains why WAMS-DD must include the middleware layers. Section 4.6 then summarizes security and trust issues for WAMS-DD. Section 4.7 describes GridStat. Section 4.8 summarizes how GridStat enables a new way of thinking and programming for programmers and researchers of power applications. Finally, Section 4.9 concludes. 4.2 REQUIREMENTS AND GUIDELINES FOR COHERENT REAL-TIME DATA DELIVERY Data delivery in the power system today can be improved by reducing the use of hard-coded protocols, developing more reusable systems, and providing real end-to-end (E2E) performance guarantees. For example, in protection 58 Smart Grids applications, overprovisioning provides low latencies and high availability in the steady state but not necessarily in the face of IT failures, bugs in software or hardware that cause spurious trafic, or cyberattacks. As more applications that can exploit coherent, real-time data delivery emerge, such as those outlined in Chapter 3, using isolated networks may soon become unsustainable, as will designing a new communications system for each new application or application family. Fortunately, the state of the art in distributed computing, real-time systems, and fault-tolerant computing does support provision of strong guarantees with data delivered to many applications. If designed, implemented, and validated correctly, a stateof-the-art data delivery system can greatly lower the barrier to entry (in both time and money) and enable deployment of new power applications by simplifying the process of adding new sensors. If designed incorrectly, it will be dificult to maintain in the future because it will not be able to keep up with increasing demands. Further, these data delivery systems will have a long life, and no single network-level mechanism (for multicast, security, or QoS) can be assumed to be everywhere. It is crucial, therefore, that data delivery systems between the mission-critical peer-to-peer (P2P) automatic protection and control systems, and the power grid’s operations IT backbone have interoperability between different kinds of network mechanisms providing the same property, such as delay guarantees [1]. In this section, we examine how a WAMS-DD will be an enabling technology for the new and emerging power system. We irst overview the performance and reliability requirements that a WAMS-DD must meet. We then present IGs, based on best practices in other industries and in the ield of distributed computing systems, for achieving these requirements. Next, we compare how existing technologies meet these DRs and design guidelines when used in isolation without additional overlay networks. These include technologies and standards at the network layers (and below), at the middleware layer(s), and related ones from the power industry. We also discuss relevant research and development for wide-area middleware. After this, we discuss the emerging NASPInet effort and the GridStat data-delivery middleware. Finally, we conclude this section with a brief discussion of pertinent cybersecurity issues for next-generation data-delivery services for the electric power grid. Note that the following analysis focuses on coherent but asynchronous data delivery for operations, but the emerging communications infrastructure will provide the additional beneit of distributing the time signal required for timesynchronized measurements and control. Typically, time is received via a global positioning system (GPS) and distributed over a separate physical network using protocols such as the inter-range instrumentation group (IRIG). This results in a physical cable connection to the measuring devices, such as phasor measurement units (PMUs). Combining time distribution with the communications network provides advantages such as simplicity and reliability [2]. Furthermore, for many applications, such as a control scheme or system protection scheme, that use separate mission-critical P2P communications, operation can proceed even if global time is lost, as long as they maintain a local coherent time signal. The communications infrastructure can provide this locally common time signal when the primary GPS signal is unavailable. 59 GridStat 4.2.1 SYSTEM MODEL Figure 4.1 depicts the architecture of a WAMS-DD. Application programs or irmware that emit a stream of updates are called publishers, which are denoted as Pub1 through PubN in the diagram; Pub1, for example, outputs updates to variables X and Y. Applications that receive these updates are called subscribers, which are denoted as Sub1 through SubN. In the diagram, Sub1 subscribes to Y from Pub1 and to W from PubN. In the usual case in publish-subscribe (pub-sub) systems, neither publisher nor subscriber needs to know about the other; they are decoupled such that they only know about the variable they publish or subscribe to and how to contact the delivery system. In cases where the subscriber requires conirmation that the update came from its legitimate publisher—which may be common with a WAMS-DD— data-integrity techniques from the computer security ield can be used by the datadelivery system. Creating a pub-sub delivery path requires two steps. Publishers register their variables with the delivery system (only once per variable, not once per subscriber), and later subscribers request a subscription to a given variable. For both publishers and subscribers, the delivery system returns a handle to a piece of code called a proxy, which is generated at compile time by the data-delivery middleware. This proxy contains logic provided by the data-delivery service, which, besides doing the usual middleware proxy activities such as packaging of the parameters into a message, is also a place where data-delivery mechanisms may reside. In Figure 4.1, we denote a publisher-side proxy as Pub-Prx-Mech and the subscriber-side proxy as Sub-Prx-Mech. After the variable is registered and subscribed to, updates to variables low from publishers to subscribers, as shown by the gray arrows in Figure 4.1. To do this, they traverse what we call the WAMS-DD cloud. This is opaque because, as shown later in this section, it can be implemented in different ways resulting in different tradeoffs. For the purposes of our system model, the WAMS-DD cloud consists of a graph in which the edges are network links and the nodes contain forwarding mechanisms that can forward a message on its way toward a subscriber. X Y Pub1 Sensor app or firmware Sub1 WAMS-DD cloud Pub-Prx-Mech proxy X, Y updates Z W PubN Sensor app or firmware Pub-Prx-Mech proxy Z, W updates FIGURE 4.1 Architecture and system model of a WAMS-DD. Power app Y W Proxy Pub-Prx-Mech Y, W updates SubN Power app Z X Proxy Pub-Prx-Mech Z, X updates 60 Smart Grids We note that this WAMS-DD is in practice a virtual overlay implemented with a customer’s existing links. These existing links will have inherent trade-offs, so, considering the capabilities, most utilities/ISOs would conclude that their existing infrastructure will not support any closed-loop applications, at the very least. Thus, the existing links would in practice be augmented by adding key additional links from various sources, such as a Tier 1 iber provider, lighting up utility dark iber, or buying some new links with technologies providing very strong guarantees (see Table 4.2). This is discussed further in Section 4.5. Updates from a publisher of a sensor variable thus traverse one or more paths to be delivered to a given subscriber. Along a given path, an update may be delayed, so that its required delivery latency cannot be met, or the update may be dropped due to failures in a network link or forwarding node or due to a cyberattack. However, the probability of an update not meeting its DRs can be kept extremely low by carefully designing the WAMS-DD and by allocating multiple paths for important updates. That is, a WAMS-DD can be constructed so that the on-time delivery probability is very high, as long as its design constraints are met. Informally, these include forwarding capacity per node, maximum link trafic, number and kind of benign failures and cyberattacks, and so on. We now overview the DRs in Section 4.2.2; then, in Section 4.2.3, we describe IGs that can be used to meet these DRs with extremely high probabilities. These probabilities offer the potential to practice using dual isolated networks for critical protection applications, while at the same time supporting many more application families with thousands of update lows. However, such delivery technologies clearly need to be proven in the ield before any migration to them can begin to be contemplated. 4.2.2 DELIVERY REQUIREMENTS FOR A WAMS-DD The following DRs must be met by a WAMS-DD [3]; these do not include the details of cybersecurity-related requirements (which are overviewed in Sections 4.5.5 and 4.7.7). These DRs encompass the requirements of other network layers that the WAMS-DD layer builds upon. They are summarized in Table 4.1, along with their IGs, which will be overviewed in the next subsection. Requirement 1: Hard, E2E guarantees must be provided over an entire grid because protection and control applications depend on the data delivery. The guarantees must be deterministic: met unless the system’s design criteria have been violated (e.g., trafic amount, number of failures, and severity of cyberattack). Requirement 2: WAMS-DDs must have a long lifetime and thus must be designed with future prooing in mind. This is crucial in order to amortize costs over many projects, utilities, grids, and so on. The goal of NASPInet, for example, is to last at least 30 years. To achieve this, it is important not to have applications hardcoded directly using the application programming interfaces (APIs) of a single QoS+ mechanism for a given kind of resource (e.g., encryption, network delivery). Rather, wrappers or, better, more comprehensive middleware should be used. The system’s libraries can be updated to add new lower-level QoS mechanisms as they become available without having to recompile the application. GridStat 61 Requirement 3: One-to-many is the normal mode of communications, not pointto-point. Increasingly, a given sensor value is needed by multiple power applications. This is best done by pub-sub middleware, but in a worst case (unfortunately common in power grids) it is done by network-level multicast (e.g., Internet protocol [IP] multicast [IPMC]). Requirement 4: E2E guarantees must be provided for a wide range of QoS+. Data delivery for the power system is not “one size its all” [4], as shown in Chapter 3. For example, to provide very low latencies, very high rates, and very high criticality/availability to all applications would be prohibitively expensive. Fortunately, many applications do not require these stringent guarantees, but their less stringent requirements must nevertheless be met. Examples of the wide ranges that must be provided follow: 1. Latency and rate: Ten milliseconds or less, up to seconds (or hours or days for bulk transfer trafic), 001–720 Hz or more. 2. Criticality/availability: IntelliGrid [4] recommends ive levels of availability of data, from medium up to ultra (99.9999%), that is, six nines. 3. Cybersecurity: Support a range of trade-offs of strength of encryption and authentication compared with delay induced, resources consumed, and the exact lavor of authentication provided. Requirement 5: Some merging and future system integrity protection schemes (SIPS), transient stability, and control applications require ultralow latencies and one-way delivery on the order of a half or full power cycle (8–16 ms in the United States) over hundreds of miles [5]. Thus, any forwarding protocols should not add more than a millisecond or two of latency (through all forwarding hops) on top of the speed of light in the underlying communications medium. These latencies must be provided in a way that: 1. Is predictable and guaranteed for each update message. Each sensor update needs to arrive with an extremely high probability within its required guaranteed deadline, not a much weaker aggregate guarantee over longer periods of time, applications, and locations such as is provided by MPLS technology [6]. 2. Tolerates nonmalicious failures in the WAMS-DD infrastructure. No system can tolerate unlimited kinds and numbers of failures. However, much as the power system must continue in the face of one or more known contingencies, the IT infrastructure on which it increasingly depends must still provide these hard, E2E guarantees in the face of failures (up to design limits). 3. Tolerates malicious cyberattacks. Power systems are known to be subjects of extensive study and probing by multiple organizations that have signiicant information warfare capabilities, including nation-states, terrorist organizations, and organized crime. A WAMS-DD must adapt and continue to deliver data despite cyberattacks of a designed severity 62 TABLE 4.1 IGs and the DRs That Mandate Them DR1: Hard E2E WAN Guarantees DR2: Future Prooing DR3: DR4: One-ToWide Many Range of Comms QoS+ … 4A: Latency and Rate 4B: Criticality/ Availability 4C: Cybersecurity 5A: DR5: Per-Update 5B: 5C: Ultralow and Tolerating Tolerating DR6: High Latencies … Predictable Failures Cyberattacks Throughput IGx Prerequisites — — — — — — — X X — — — X — — — X — — — X X X X — — — X — — — — — — — — X — — — X — — — — — — — — X — — — — — — — — — — — — — 2, 3 X — — — — X — — X — — X — X X — — X X X — — — — — — X — — — — — — — X — — X — X — — — — — — — X X X — 8 X — — — — — — — X — — — — X — — — — — — — X — — — 8–10 IG1: Avoid posterror recovery mechanisms IG2: Optimize for rate-based sensors IG3: Provide per-subscriber QoS+ IG4: Provide eficient multicast IG5: Provide synchronized rate downsampling IG6: Do not depend on priority-based “guarantees” IG7: Provide end-to-end interoperability across different/new IT technologies (multicast, QoS+) IG8: Exploit a priori knowledge of trafic IG9: Have systematic, quick internal instrumentation IG10: Exploit smaller scale of the WAMS-DD IG11: Use static, not dynamic, routing Smart Grids X Summary of Implementation Guideline IGx — — — — — — — X X X — — IG12: Enforce complete perimeter control X — — — — — — — X X X X 12 IG13: Reject unauthorized messages quickly and locally — — — — — — — — — — — X 2, 8 IG14: Provide only simple subscription criteria — — — — — — — — X — — X 2 IG15: Support transient, not persistent, delivery — — — — — — — — X — — X — IG16: Do not overdesign consistency and (re)ordering — — — — — — — — — — — X 2, 8, 14–16 IG17: Minimize forwarding-time logic X X — — X X X — — — — — — IG18: Support multiple QoS+ mechanisms for different operating conditions — — — — — — — — X — — X 17 X — — — — — — — X — — X — IG19: Inspect only message header, not payload IG20: Manage aperiodic trafic GridStat X 63 64 Smart Grids (it should be possible to increase this threshold over the life of the system). Note that a bug in hardware or software that generates spurious trafic can have an effect similar to that of a cyberattack. Requirement 6: Extremely high throughput is required. Today’s synchrophasor applications are generally limited to 30 or 60 Hz in the United States, in part because the communications systems they use are not designed to support higher rates. Not to provide much higher sustainable throughput would greatly limit the number of new applications that can help the power system’s stability. Indeed, not just synchrophasors but digital fault recorders (DFRs) and intelligent electronic devices in substations provide a wealth of data. It is quite conceivable and likely that “if you build it, they will come” and there will be many thousands of synchrophasors, relays, DFRs, and other sources of sensor updates across a grid. These devices can output at 720 Hz and sample at 8 kHz, but their full output is not always used remotely due to communications limitations. If key relay or DFR data could be delivered from a set of devices across a grid at 720 Hz, many new opportunities would open up for transient protection without using expensive dedicated networks or “drilling down” into the root causes of an ongoing power contingency using additional contingencyspeciic data. We are not aware of any commercial or military market for a wide-area data delivery infrastructure that has the stringent requirements of a WAMS-DD, including the ability to enforce complete perimeter control, the ability to know the vast majority of the trafic ahead of time, and other factors incorporated into the IGs described next in this chapter. The reason is quite simple; electric power is the only market with such stringent requirements. However, these requirements are achievable using state-of-the-art distributed real-time embedded computing [7,8], as long as a careful E2E analysis is done [9] and the core data delivery mechanisms are not saddled with unnecessary features. Much broader reliability has been explored in the fault-tolerant distributed computing community, from where appropriate lessons, both good and bad, should be heeded [10,11]. 4.2.3 IMPLEMENTATION GUIDELINES FOR A WAMS-DD We now discuss practical issues that arise when trying to construct a WAMS-DD that meets the above requirements. 4.2.3.1 Context The requirements outlined in the previous section were kept to a bare minimum. In order to achieve them, however, we believe it will be necessary to utilize a number of IGs, many of which are quite different from what is provided in today’s best-effort Internet and what has been the conventional wisdom in networking research. In this section, we enumerate and explain these IGs. However, they are necessary (or at least highly advisable) in order to meet the stringent DRs. Some of the IGs below (e.g., IG4 and IG5) are actually deemed requirements for NASPInet [12], but we describe them here as IGs because it is possible to build a WAMS-DD without them (although we believe that would be inadvisable for an interutility backbone such as the proposed NASPInet; see Section 4.4). These IGs GridStat 65 are drawn from a number of sources, including our knowledge of what the state of the art in distributed computing has demonstrated is feasible, best practices in other industries, and decades of experience gained in Defense Advanced Research Projects Agency wide-area application and middleware projects, both our own and those of others. We note that these guidelines refer to best practices of how to build a WAMS-DD. Other guidelines (beyond the scope of this chapter) will apply on how to use one, and will need to be developed as best practices. For example, in our experience, many power engineers assume that, with synchrophasors, they should have phasor data concentrators (PDCs) inside their utility at many levels. However, this is a very bad idea for updates that need to be delivered with ultralow latencies (DR5). A PDC aggregates many PMU signals, performs error correction and angle computation, then outputs the collection of this information for a given PMU time slot (this is called time alignment). Such a PDC may have dozens of PMU signals coming into it, so performing time alignment means that the output has to wait until the slowest PMU update arrives. In this case, the updated sensor values will have suffered signiicant delays even before they leave the utility to be transported by a WAMS-DD such as NASPInet. Thus, for those updates that require ultralow delivery latency, any PDC or other time alignment should be placed as close to the subscribers as possible (ideally in their local area network), even at the cost of a small amount of either wasted bandwidth or duplication of PDCs. Similarly, data that are required with extremely low latency should not have a database in their path: they can be entered into a database after they are sent out, but the database must not slow down the fast delivery path. We also note that the scope of these IGs involves only the data-delivery system for WAMS-DD. It does not include the supporting services that will be required for coniguration, security, path allocation, resource management, and so on. It will be important for the WAMS-DD that the use of these tools avoids hard-coding choices, but rather allows them to be speciied in a high-level policy language (or at least a database) [13–15]. For an example of a hierarchical version of such services (a “management plane”), see [16,17]. 4.2.3.2 Implementation Guidelines Table 4.1 provides an overview of the IGs and the DRs that require the given IG. We now explain each of the IGs in turn. Guideline 1: Avoid posterror recovery mechanisms. Traditional protocols for the Internet in general, and reliable multicast protocols from the fault-tolerant computing research community, use posterror recovery. In these protocols the receiver sends either a positive acknowledgment (ACK) when it receives a message, or a negative acknowledgment (NACK) when it concludes that the message will not arrive. However, both add considerable latency when a message* is dropped: three one-way latencies are required, plus a timeout that is much greater than the average one-way message latency. * We use the term message rather than packet, because in many cases we are describing middlewarelayer mechanisms above the network and transport layers. 66 Smart Grids The better alternative is to send sensor updates (messages) proactively over multiple disjoint paths, each of which meets the latency and rate requirements [18,19]. Indeed, if multiple independent messages, each going over a QoS-managed path, cannot meet the delivery deadline, then sending ACKs or NACKs is very unlikely to help, and, indeed, will almost certainly only make things much worse. Guideline 2: Optimize for rate-based sensors. WAMS-DD can be made with higher throughput and robustness if they are not overengineered. General-purpose pub-sub systems offer a wide range of trafic types, because they are designed to support a wide range of applications. However, in a WAMS-DD, the vast majority of the trafic will be rate based. Design accordingly. Guideline 3: Provide per-subscriber QoS+. It is crucial that different subscribers to the same sensor variable are able to have different guarantees in terms of latency, rate, and criticality/availability. If not, then a lot of bandwidth will be wasted: all subscribers will have to be delivered that sensor’s updates at the most stringent QoS+ that any of its subscribers requires. Guideline 4: Provide rate-eficient multicast. In order to achieve the highest throughput possible, it is imperative to avoid unnecessary network trafic. Thus, never send an update over a link more than once. Also, as a sensor update is being forwarded through the network, if it is not needed downstream in the multicast tree (e.g., those subscribers require it at a lower rate than other subscribers), the update message should be dropped, which can best be implemented using an in-network rate down-sampling mechanism, as is done in GridStat [18,19]. These irst four guidelines add up to a need for multicast routing heuristics that provide multiple, disjoint paths to each subscriber, with each path meeting the subscriber’s latency requirement. A family of heuristics developed for this multicast routing problem [20,21] conirms the feasibility of the approach at the anticipated scale (see IG10) if routing decisions are made statically (IG11). Guideline 5: Provide synchronized rate downsampling. In providing rate downsampling, it is important not to downsample in a way that destroys the usefulness of some data. For example, synchrophasors are used to take a direct state measurement at a given microsecond. If some subscribers require only a small fraction of the updates for a set of synchrophasor sensors, it is important that the updates that reach the subscriber at each interval carry the same time stamp. For example, if a subscriber only requires one-tenth of the updates from two different variables, then it would not be meaningful to get updates {#1, #11, #21, …} from one synchrophasor and updates {#2, #12, #22} from another synchrophasor, because the given measurements (e.g., #1 vs. #2) do not correspond to the same time (they are not the same snapshot), which is the main point of synchrophasors. Guideline 6: Do not depend on priority-based “guarantees.” Pub-sub delivery systems typically offer a way to specify a priority, so if the trafic gets too heavy less important trafic can be dropped. However, this does not provide a hard E2E guarantee to subscribing applications, and even applications that are not of the highest criticality still need their DRs to be met. Instead of priorities, mechanisms must be used that exploit the characteristics of WAMS-DD (as outlined in these guidelines) to provide each subscriber with irm assurances that its guarantees will GridStat 67 be met as long as there are not more than the agreed-upon number of failures or severity of cyberattack. Guideline 7: Provide E2E interoperability across different/new IT technologies (providing multicast, latency, rate, etc.). A grid-wide WAMS-DD will ipso facto have to span many utility and network organizations. It is unlikely that the same mechanisms will be present across all these organizations. And, even if they are today, if the WAMS-DD gets locked into the lower-level APIs and semantics of a given multicast or QoS mechanism, it will be dificult to “ride the technology curve” and utilize new and better mechanisms that will inevitably become available over the long lifetime of the WAMS-DD. This is a stated goal of the GridWise community, for example [22]. Fortunately, it is possible to use middleware to span these different underlying technologies in order to provide guarantees that span this underlying diversity. Guideline 8: Exploit a priori knowledge of predictable trafic. Internet routers cannot make assumptions or optimizations based on the characteristics of the trafic that they will be subjected to, because they are intended to be general purpose and support a wide range of trafic types. WAMS-DDs, however, have trafic that is not just rate based, but is almost all known months ahead of time (e.g., when an engineering survey is carried out on a new power application). This common case can be optimized, as described in later IGs below. Guideline 9: Have systematic, quick internal instrumentation. In order to provide E2E guarantees across a wide area despite failures and cyberattacks, IG8 must be exploited to provide systematic and fast instrumentation of the WAMS-DD. This allows much quicker adaptation to anomalous trafic, whether accidental or malicious in origin. Guideline 10: Exploit smaller scale of the WAMS-DD. This is crucial if the challenging DRs are to be met over a wide area with reasonable cost. However, this requires rethinking the conventional wisdom in networking research and commercial middleware products. A WAMS-DD for even the largest electric grid will be orders of magnitude smaller in scale than the Internet at large,* so it is feasible for the entire coniguration to be stored in one location for the purposes of (mostly off-line) route selection. Additionally, academic computer science researchers historically consider something that is O(N2) for path calculation with N routers or forwarding engines (FEs) to be infeasible; see, for example, [17]. However, this assumption ignores two key factors for WAMS-DD. First, N is not in the neighborhood of 108, as in the Internet, but rather is more likely to be ~103, at least for the next 5–10 years; that is, even if O(N2) algorithms are feasible at this scale. Second, as a rule, power engineers do not decide that they need a given sensor’s values seconds before they really need them, due in part to the fact that today’s data-delivery infrastructure requires them to recode hard-coded socket programs and then recompile. Rather, power engineers plan their power contingencies (and what data they will need in them) months ahead of time with detailed engineering studies, and similarly for their monitoring, protection, * For example, in the entire United States there are approximately 3500 companies that participate in the grid. We thus believe that the number of router-like FEs that would be required for a NnDB backbone (at least in the case of broker-based pub-sub; deined later) is at most 10 4 and likely only around 103. 68 Smart Grids control, and visualization needs. Thus, the routing/forwarding decisions involved in path selection can be done off-line well ahead of time, while still allowing for handling a modest number of subscription requests at runtime. It is also feasible for router-like FEs to store state for each low. Having a router keep per-low state has long been considered a bane to networking researchers, because it is considered to be prohibitively unscalable. However, with the much smaller scale and the much more limited type of applications for a WAMS-DD, storing per-low state is not only feasible but is also a requirement for providing IG3 (per-subscriber QoS+) with IG4 (eficient multicast); this is something that our GridStat project has been advocating for many years [16]. However, recently networking researchers have realized the necessity of storing per-low state to provide any reasonable kind of QoS [23]. Guideline 11: Use static, not dynamic, routing and naming. Much stronger latency guarantees can be provided when using complete knowledge of topology coupled with static routing. Complete topology knowledge is a reasonable assumption in a managed NnDB, given that it will be a carefully managed critical infrastructure with complete admission control. Also, almost all of the sensors and power applications will be known well ahead of time, so optimizations for static (or slowly changing) naming can potentially be useful and can be done while still providing more lexible and dynamic discovery services at a much lower volume. We note that networking and security researchers generally assume that the membership of multicast groups (or a set of subscribers) may change rapidly; see, for example, [17]. However, as noted above, that is not the case with WAMS-DD. Guideline 12: Enforce complete perimeter control. All trafic put onto a WAMS-DD must pass admission control criteria (permissions based on both security and resource management) via a management system: the publisher registering a sensor variable (at a given rate) and the subscribers asking for a subscription with a given rate and E2E latency. This is essential to provide guarantees at a per-message granularity. It also enables quicker adaptations. This should ideally be done in the publishing application’s address space, via a middleware proxy, so that spurious trafic does not consume any network or router capacity. Guideline 13: Reject unauthorized messages quickly and locally. Messages that have gone around the admission control perimeter should be rejected as soon as possible, ideally at the next WAMS-DD FE, rather than going most or all the way across the WAMS-DD consuming resources along the way. Detection of such unauthorized packets is an indicator of anomalous trafic and evidence of a failure or cyberattack that needs to be reported to the management infrastructure. When suficient evidence is collected over suficient time, an appropriate adaptation can occur, but such rejection would be in practice reported to a management infrastructure, given that it is an anomaly. Guideline 14: Provide only simple subscription criteria. This is exactly the opposite of what is usually done with general-purpose pub-sub in either academic research or commercial products: both tend to favor complex subscription criteria, which are expensive to evaluate as each update is forwarded through the system (think of complex “topics”) [24]. For example, in GridStat, the subscription criteria are latency, rate, and number of paths, and, as noted below, the forwarding decision GridStat 69 is made completely based on rate, with static routing. Note also that the lower-level ID of a sensor variable could still be looked up through a complicated discovery service; this guideline is concerned with avoiding complex forwarding logic. Guideline 15: Support only transient delivery, not persistent delivery. Most pub-sub systems offer persistent delivery, whereby if an event cannot be immediately forwarded it is stored for some time and then the delivery is retried. This harms throughput, however, as well as potentially the per-packet predictability (because it requires storing the data). In our experience it is completely unnecessary for real-time visualization, control, and protection, due to the temporal redundancy inherent in rate-based update streams: the next update will be arriving very soon anyway, so the usefulness of a given update fades very quickly. Thus, it is inadvisable to complicate the critical paths of delivery mechanisms to support persistent delivery (though it can be provided “on the side” by other mechanisms). Furthermore, in the power grid, historian databases are already required for archiving data for regulatory reasons, so there is no reason to complicate the design or otherwise bog down the fastest and highest-availability mechanisms of WAMS-DD to deliver historical data.* Guideline 16: Do not overdesign for consistency and (re)ordering. Research in fault-tolerant multicast tends to provide different levels of ordering between updates from the same publisher, or between different clients of the same server, as well as consistency levels between different replicas or caches of a server. There is no need for anything like this in a WAMS-DD. Present data-delivery software provides no kind of consistency at all, so power applications assume nothing in terms of consistency and ordering. The only requirement for such consistency that we have found is relected in IG5 for synchrophasors, and the only ordering of any kind is where a PDC combines updates from different PMUs into one message to pass onward. With devices such as synchrophasors that have accurate GPS clocks, the order of events can be directly known and no delivery ordering mechanism is required other than that which is done by a PDC. Guideline 17: Minimize forwarding-time logic. In order to provide the highest throughput, the forwarding logic that decides how a packet or update is to be forwarded on should be kept as simple as possible. In the GridStat project, forwarding decisions are made based solely on the subscription rate of subscribers downstream in the multicast tree [16,19]. Given that the trafic is rate based (IG2) and known ahead of time (IG8), that subscription criteria are kept simple (IG14), that only transient delivery is supported (IG15), and that there are no consistency semantics (IG16), much logic can be pushed off to subscription set-up time or even off-line. This reduces the logic necessary when an update arrives at an FE (or P2P middleware mechanisms at an edge) and hence greatly increases throughput and decreases latency. Guideline 18: Support multiple QoS+ mechanisms for different runtime conditions, beyond just network resources, in a uniied, coherent way. A given mechanism * We note that such postevent historical data can be delivered by the same physical network links as the fast trafic with trafic isolation mechanisms; indeed, this is one of the main trafic categories for the emerging NASPInet. 70 Smart Grids that provides guarantees of latency and security, for example, will not be appropriate for all the runtime operating conditions in which a long-lived WAMS-DD may have to operate. This is because different implementations of a given QoS+ mechanism can require very different amounts of lower-level resources such as a central processing unit (CPU), memory, and bandwidth [25]. Also, in a very big WAMS-DD, many QoS+ mechanisms will not be present everywhere, or close, so spanning them in a coherent manner and such that applications do not have to be aware of the above limitations is important. This is discussed further in Section 4.5. Guideline 19: Inspect only packet header, not payload. In order to provide the highest throughput and lowest latency, ensure that subscription criteria and consistency semantics allow a forwarding decision to be based solely on a packet header. This is not possible for pub-sub middleware that has complicated subscription topics, as is typical of commercial and research systems. For them, data ields in the payload also have to be inspected. Guideline 20: Manage aperiodic trafic. Any trafic that is aperiodic (i.e., not based on rate but on a condition) must be isolated from rate-based periodic trafic and managed accordingly. This can be done deterministically, for example, with (OSI Layer 1) optical wave division multiplexing (OWDM) hardware. Further, aperiodic trafic should be aggregated intelligently—for example, based on updateable policies rather than hard-coded settings—instead of sending all alarms/alerts to the next level up for processing. 4.2.3.3 Analysis It is important to recognize that you cannot have the highest level of all the properties described in the DRs for every sensor variable. As noted in [1]: 1. Different properties inherently must be traded off against others. 2. Different mechanisms for a given property are appropriate for only some of the runtime operating conditions that an application may encounter (especially a long-lived one). 3. Different mechanisms for the same nonfunctional property can have different trade-offs of lower-level resources (CPU, bandwidth, storage). 4. Mechanisms most often cannot be combined in arbitrary ways. Even if you somehow could have them all at once, it would be prohibitively expensive. Given these realities, and the fact that application programmers are rarely experts in dealing with the above issues, middleware with QoS+ properties supported in a comprehensive and coherent way is a means to package up the handling of these issues and allow reuse across application families, organizations, and even industries. Finally, because of length constraints it is not possible in this chapter to fully discuss the cybersecurity issues that arise in a WAMS-DD. Clearly, a WAMS-DD providing universal connectivity creates cybersecurity challenges beyond those arising in a conventional, single-utility supervisory control and data acquisition (SCADA) system. Cybersecurity also interacts with DRs and IGs: for example, techniques used for message conidentiality and authentication must not impose too much additional latency, GridStat 71 yet the multicast requirement appears to limit use of symmetric-key cryptography for authentication. See Sections 4.5.5 and 4.7.7 for more information, and, of course, the references cited therein. 4.3 ANALYSIS OF EXISTING TECHNOLOGIES FOR A WAMS-DD We now analyze how existing technologies and standards meet the above DRs and IGs. We review, in turn, traditional network layer technologies, middleware technologies, and electric sector technologies. This section ends with Table 4.2, which very clearly depicts how almost all technologies are very inadequate for WAMS-DD, especially when closed-loop applications must be supported (or at least not designed out, either deliberately or inadvertently). 4.3.1 TECHNOLOGIES AND STANDARDS AT THE TRADITIONAL NETWORK LAYERS Almost all WAMS-DD products and research to date have been drawn from the “network” layers. We analyze representative technologies and categorize them under subsections 4.3.1.1 and 4.3.1.2. Their limitations are clearly summarized in Table 4.2. It is important to note that a number of these IGs have the appropriate networklevel mechanism to implement the given IG, but they cannot do it without help from higher layers: the middleware (or the application if middleware is not used) and also sometimes application management. This is to help parameterize the mechanisms with higher-level information outside the scope of the network level. This is discussed further in Section 4.5. 4.3.1.1 Very Inadequate Traditional network protocols, including the open systems interconnection (OSI)-2 “data link” layer (e.g., Ethernet), OSI-3 “network” layer (e.g., IP), and the OSI-4 “transport” layer (e.g., transmission control protocol [TCP], user datagram protocol [UDP], and stream control transmission protocol [SCTP]), do not provide E2E QoS+ guarantees or multicast [26,27]. This is because they are at lower networking layers and E2E functionality is not their intended use. All of these lower-layer protocols can be part of the E2E solution above which WAMS-DD sits. Nevertheless, some systems do apply them in ways that are nearly E2E in scope, and therefore we now examine these protocols and extensions to them to see how they would meet the requirements and guidelines if they were implemented as the complete E2E solution. We note that a major limitation of network-level solutions is that, based on our long investigations in the ield, there is no other application domain that has DRs anywhere near as stringent as those of the bulk power grid. IPv6 low labels [28] associate each “reservation” with an application-to-application network socket connection, which contains many different sensor-update streams with a wide range of required QoS+. Packets are processed in a low-speciic manner by the nodes that have been set up with a low-speciic state. The nature of the speciic treatment and the methods for the low state establishment are out of scope of the speciication. 72 TABLE 4.2 Coverage of DRs and IG by Existing Technologies Network Level Very Inadequate Better Yet Incomplete Middleware Electric Sector TCP, IPv6 VLANs BB P2P UDP, Flow IP and Open Int-Serv Net COTS COTS Streaming Military SCTP Labels Multicast MPLS VPNs ATM Flow Guara. Svcs Anagran Insight SeDAX Pub-Sub Pub-Sub SQL/CEP RT Apps SOA/Web Services IEC 61850, OPC UA DNP3, MMS Grid Stat ? — — — Y ? Y — Y — D D — D — — — Y — — S ? — Y — Y — ? — — S Y — — — Y S Y S Y Y Y Y Y D D Y Y Y S — — — — Y Y — — S — — — — — S Y — S ? — ? Y — — S D φ — φ φ D — S Y Y S — — Y Y Y Y Y Y Y — — — Y Y Y — — S — — — — — — Y Y Y — — — — — Y — — — Y — D D — D — — — Y φ φ Y — φ φ Y φ φ φ Y — φ φ Y — φ φ ? — ? — Y Y D D Y Y — — Y Y ? ? Y Y S Y Y Y Y ? Y — D ? Y ? — ? Y — — D Y D — S Y ? S — — — — — — — — — — — Y S Y Y — — — — — — — — — S — ? ? D ? — — — Y φ — — D — — S — — S — ? ? ? ? D — — Y φ φ S ? — — S S S S — ? D D ? — — — Y — — — — — — — — — — — — — D — — — — Y DR1: Hard E2E WAN guarantees DR2: Future prooing DR3: One-to-many (pub-sub or multicast) DR4: Wide range of QoS+: 4A: Latency and rate 4B: Criticality/availability 4C: Cybersecurity DR5: Ultralow latencies: 5A: Per-message and predictable 5B: Tolerating failures 5C: Tolerating cyberattacks DR6: High throughput IG1: Avoid posterror recovery mechanisms IG2: Optimize for rate-based sensors IG3: Provide per-subscriber QoS+ IG4: Provide rate-eficient multicast IG5: Provide synchronizedrate downsampling Smart Grids — Delivery Requirement or Implementation Guideline φ — — — — — — Y φ ? ? D ? — — — Y — — — — — — S φ φ S S Y Y S Y Y — — Y — ? — ? — — — — — S — ? ? L ? — — — Y — φ — ? — — — — ? — ? ? ? D D — — — S — — — — — — — — — — — ? ? D ? — — — Y — — — — — — S S S Y — ? ? D ? — — — Y — — — — — φ S φ φ — — — D — — — — Y — ? — — — S S S S — ? ? D ? S — — Y — Y — — — — φ φ φ φ — D D — D — — — Y — Y Y — — — Y Y Y Y — S S D D — — — Y Y Y Y Y Y Y Y Y Y Y Y ? ? D ? S — — Y Y Y Y Y Y Y Y Y Y Y — ? ? — ? — — — Y — — — — — — S — — — — ? ? — ? — — — Y Y Y Y Y Y Y Y Y Y Y ? — — — — — — — Y — — — — — — — — — — — D D — D — — — F IG6: Do not count on priority “guarantees” IG7: Provide E2E interoperability across different/ new IT technologies (multicast, QoS+) IG8: Exploit a priori knowledge of trafic IG9: Have systematic, quick internal instrument IG10: Exploit smaller scale of the WAMS-DD IG11: Use static, not dynamic, routing IG12: Enforce complete perimeter control IG13: Reject unauthorized packets quickly and locally IG14: Provide only simple subscription criteria IG15: Support transient, not persistent, delivery IG16: Do not provide unnecessary consistency IG17: Minimize forwarding-time logic IG18: Support multiple QoS+ mechanisms for different operating conditions IG19: Inspect only packet header, not payload IG20: Manage aperiodic trafic GridStat — φ Note: ‘—’ means the requirement is not met or the guideline not followed. Also note how few ‘Y’ entries there are in almost all columns. 73 74 Smart Grids IPMC provides eficient multicast for a single, nonreplicated low. However, if multiple IPMC addresses are used as a replication mechanism to attempt to provide redundant path delivery, there is no guarantee that the corresponding multicast trees will be disjoint, which is important not only for eficient multicast (IG4) but also for providing low latencies in the face of failures (DR5B). Further, IPMC addresses must be used carefully in even a mid-sized infrastructure; for example, due to the instabilities IPMC can cause, it has been banned from many cloud computing centers [29–31]. Finally, IPMC also does not, by itself, have other E2E capabilities that are necessary for a WAMS-DD, as clearly shown in Table 4.2. MPLS is designed to give Internet service providers (ISPs) a set of management tools for bandwidth provisioning, not to provide ine-grained (per-update) QoS [32]. Its guarantees are weak compared with the needs of a critical infrastructure. For example, it gives aggregate economic guarantees over user, location, and protocol, not hard guarantees (DR1) for each update (DR5A). As noted in [33], “DiffServ does not guarantee that lows classiied with a higher priority will really observe a better quality of service than lower priority ones.” Given that MPLS is a variant of DiffServ, this is directly applicable. It is worth noting that in the relatively tranquil steady state MPLS may give extremely good latency and availability. However, in power contingencies, IT failures, cyberattacks, or other situations that place stress on the WAMS-DD, the guarantees are nonexistent. Further, different ISPs can implement MPLS in different ways. There are no facilities for combining lows across different ISPs, as would be required in a WAMS-DD, or for predicting the E2E delays. MPLS also has severe limitations with respect to the granularity of its “guarantees” and adaptation. The MPLS header only has three bits for the class ield, so each low is assigned one class of only eight. This is a far cry from providing guarantees, management, adaptation, and so forth on a per-subscriber-per-publisher-each-update basis with hard guarantees. This is why MPLS meets few of the DRs and IGs, as shown in Table 4.2. MPLS has some fault-tolerance mechanisms, such as a fast reroute feature, detour merging, and E2E path protection. However, these mechanisms presently provide a minimum latency of about 50 ms, which is too long for the emerging SIPS and other applications described in Chapter 3. Finally, MPLS is a closed system; OpenFlow can match its performance while having an open and extensible system (though, as shown below, one at present far from adequate for closed-loop applications) [34]. Virtual local area networks (VLANs) and virtual private networks (VPNs) do not meet the DRs listed above because their purposes are orthogonal to the DRs. A VPN or VLAN could be part of a WAMS-DD, but VPN and VLAN technologies alone do not meet the requirements and can add to latency and decrease throughput. Pragmatic general multicast (PGM) is a transport-layer multicast protocol [35]. Implementation by Microsoft is known as reliably delivered messages (RDM). PGM runs over a datagram multicast protocol such as IPMC, to provide basic reliable delivery by use of NACKs. PGM uses a rate-based transmission strategy to constrain the bandwidth consumed. However, it does not provide real-time guarantees. For brevity, it is not summarized in Table 4.2, but its coverage of the DRs and IGs should be clear from the technologies that it depends upon, which are covered there. GridStat 75 Spread can be considered a high-level multicast protocol that provides a range of ordering strengths across a wide area network (WAN) [36]. It supports ordered delivery and the resulting consistency, even in the face of network partitions, and it is used largely for replicating databases. It has no real-time mechanisms. 4.3.1.2 Better Yet Incomplete Asynchronous transfer mode (ATM) and synchronous optical networking (SONET) are networking technologies sometimes employed in WANs. They offer strong latency guarantees on a per-message basis. ATM does not support multicast (DR3) or multiple disjoint paths (DR4B). Given ATM’s strong latency guarantees, however, at the right granularity, the ATM protocol can be part of a WAMS-DD that overlays ATM and other protocols. Of course, such an overlay network would by deinition be middleware, which is needed for WAMS-DD for other reasons discussed in Section 4.5. OpenFlow is a network communications protocol that gives remote access to the forwarding plane of a network switch or router [37,38]. It does this by abstracting common features of routers and switches to provide constructs such as low tables and an instruction set that allows remote access to those constructs. Many vendors have implemented OpenFlow on their products with varying degrees of completeness and consistency [39]. Finally, OpenFlow can implement all the features of MPLS (a closed system) and with similar performance in power settings [33]. One fundamental limitation of OpenFlow is that it is ipso facto a lowest-commondenominator (LCD) approach, having to be mapped to a wide range of devices. Very few of these have any mission-critical requirements, let alone anything close to their LCD. Worse, OpenFlow has no notion of global time. While it is possible to provide some degree of predictable performance by very carefully managing lows, providing strong guarantees with tight deadlines and low jitter under dificult conditions is problematic at best. Finally, OpenFlow has no primitives that can do anything like rate downsampling even on a single low (i.e., IG3 and IG4), let alone synchronized with another low (IG5). The standard Internet integrated services (int-serv) has a guaranteed service (GS) that strives to provide strong delivery guarantees. It does provide hard E2E guarantees (DR1). However, it guarantees a maximum latency, not an average, so it has to be very conservative in what it promises. Further, it is not designed to provide low jitter, and its delay can be quite signiicant compared with closed-loop requirements (it can easily be far longer than these requirements). Still, while it does not meet most of the DRs and IGs, using int-serv GS is signiicantly better than using MPLS. The signaling mechanism of int-serv, resource reservation protocol (RSVP), does perform admission control. However, it is done in the router, not in the application’s address space (as is simple to do with middleware proxies, which are in the application’s address space), so spurious trafic can still slow down routers. Also, using this admission control with a ine enough granularity, that is, to ensure that each publication from a given publisher applications does not go over rate, would be cumbersome with RSVP and take a large number of ports, potentially exhausting the 16-bit space of UDP ports. Anagran technology [23], which is based on earlier Caspian technology, has lowbased routers that keep per-low state to provide better QoS via low management. 76 Smart Grids It has not historically released much information on its router. The survey [34] notes that Anagran technology provides only weak statistical guarantees, not strong ones as required for closed-loop applications. It also has no per-sensor, per-subscriber granularity, or anything close: its lows use a 5-tuple in IPv4 (source and destination address and port, protocol) or an IPv6 low label (see above) if available. This is the opposite of what is needed in critical infrastructures: to be unfair to provide strong guarantees! Further, Anagran technology automatically classiies lows into a number of classes, rather than being provided with the QoS+ requirements for each publisher and subscriber. Anagran’s corporate life cycle may well be a lesson for the realities of inding QoS+ mechanisms from other industries that are appropriate for closed-loop applications: in the broader market, QoS+ mechanisms stronger than MPLS but weaker than what was needed for closed-loop applications did not ind a market big enough to sustain Anagran. It is worth noting that the company had a lot of deep technical expertise and credibility, given that it was founded by Larry Roberts, the person who envisioned and commissioned the Advanced Research Projects Agency Network (ARPANET) in the late 1960s, and it was very well funded by investors. Net Insight’s Nimbra platform [40] comes much closer than any other network-level technology to meeting the DRs and IGs that closed-loop applications mandate, and is actually the one counterexample we are aware of regarding the negative lessons from Anagran’s demise. Net Insight’s primary market has been broadcast television backhaul and distribution networks, but its technology is very applicable to closed-loop applications: Net Insight uses tight time synchronization (without requiring GPS) to provide both very strong latency guarantees and stronger security, even with multicast delivery. Nimbra also has wrappers for some network technologies such as MPLS, SONET, and so on, and presumably it also will for future networking technologies. 4.3.2 MIDDLEWARE TECHNOLOGIES AND STANDARDS There is a wide range of commercial, off-the-shelf (COTS) middleware frameworks providing different kinds of services with some relevance for a WAMS-DD. We irst consider middleware supporting the pub-sub paradigm. There are two distinct architectures for pub-sub middleware, each with advantages and disadvantages. 4.3.2.1 Deinition of Middleware Middleware is a layer of software above the network but below the application that provides a common programming abstraction across a distributed system (computers connected by networks and communicating only by message passing, or constructs built on top of the system) [41,42]. Middleware exists to help manage the complexity and heterogeneous mix of resources that are inherent in distributed systems. Middleware provides higher-level building blocks (these abstractions), which, compared with socket programming, can make code more portable, make the programmers more productive, and have fewer errors due to handling the lowerlevel communications details. Indeed, comparing programming with middleware with socket programming is much like comparing programming with a high-level language such as C# or Java with assembler programming: relatively few people do it any longer, and for very similar reasons. GridStat 77 Middleware masks the kinds of heterogeneity that are inherent in any wide-area distributed computing system in order to enhance interoperability and make it much simpler to program. All middleware masks heterogeneity in network technology and in CPU architecture. Almost all middleware masks heterogeneity across operating systems and programming languages. A notable exception to the latter is Java’s remote method invocation (RMI), which is the remote invocation system for Java (note that the Java language is not middleware; RMI is). Java RMI does this so it can exploit features in the Java programming language such as serialization and relection. A programmer can still use other middleware with Java, for example, the common object request broker architecture (CORBA) object-oriented middleware or the pub-sub data distribution service (DDS), both from the Object Management Group (OMG). Finally, some middleware does a very good job of masking heterogeneity across implementations of the same standard from different vendors; the OMG in particular does this well. Middleware can also shield programmers from many (but not all) of the dificulties inherent in programming a distributed computing system. For example, it can largely hide from the programmer the exact location of an object, the fact that an object may be replicated or may relocate, and so on. It is for these reasons, and more, that middleware has been considered “best practice” for 15–20 years in virtually any other industry that has applications spread across a network. It is our hope that it will become much more prevalent in the electricity sector. 4.3.2.2 Broker-Based Publish-Subscribe Middleware Broker-based (BB) pub-sub middleware systems rely on an infrastructure of broker nodes to forward messages toward subscribers. The data delivery plane (DDP) for a WAMS-DD, though not necessarily in many commercial systems, is a managed WAN because it implements IG12 (complete perimeter control). A BB pub-sub WAMS-DD is depicted in Figure 4.2. A node in the DDP is called an FE and is a device specialized for the particular BB pub-sub framework. We depict the mechanisms that a BB WAMS-DD system can exploit in gray; these consist of the FEs and the proxies in the same process as the applications (publishers and subscribers). In BB WAMS-DD systems intended for mission-critical applications, there is often a separate “plane” for managing the system* and providing services. This data management plane (DMP) is depicted by the dashed gray lines in Figure 4.2; it is shown here as a single entity, but it is often distributed. Publishers provide the DMP with basic QoS metadata (QMD) about their publications, for example, the rate at which they will output updates. Subscribers provide QoS requirements (QR), including rate and latency. The DMP then exerts control over the DDP (gray lines) in order to provide the delivery guarantees, for example, by updating a forwarding table for an FE. BB pub-sub systems require a broker/server infrastructure to be installed; you cannot just buy an IP router from Cisco or others. This can be a disadvantage, which often, for small and medium scales, cannot be amortized over enough applications to be justiied. BB pub-sub systems have an advantage, however, in that they place intelligence inside the network, not just at the edges. This enables, for example, eficient * In telecommunications parlance, this is often called the “control plane,” hence our use of the term plane. Telecommunications nomenclature also refers to WAMS-DD as the “data plane.” 78 Smart Grids Data management plane and services QoS metadata (QMD) X Y QoS requirements (QR) Data-delivery plane (managed WAN) Pub1 Sensor app or firmware Pub-Prx-Mech proxy X, Y updates Sub1 Power app FE FE Y W Proxy Pub-Prx-Mech Y, W updates FE Z W PubN Sensor app or firmware FE FE Pub-Prx-Mech proxy Z, W updates SubN FE Power app Z X Proxy Pub-Prx-Mech Z, X updates FE FIGURE 4.2 BB WAMS-DD. multicast (IG4) and rate downsampling throughout the data delivery system, not just at the edges. It also creates the potential to reject unauthorized packets at their next “hop” through the system (IG13). Additionally, BB systems can exploit mechanisms in the graph of FEs in order to meet more of the IGs. For example, such an FE can be used to provide per-subscriber QoS+ (IG3), provide synchronized rate downsampling (IG5), exploit a priori knowledge of trafic (IG8), and exploit the smaller scale of the WAMS-DD (IG10); that is, it can contain per-subscriber state, a forwarding table entry for every subscription for which it forwards updates. An example of a BB WAMS-DD is GridStat. 4.3.2.3 Peer-to-Peer Publish-Subscribe P2P pub-sub systems place mechanisms for reliability and iltering only at the edges of an infrastructure. A canonical architecture for a P2P pub-sub coniguration of a WAMS-DD is given in Figure 4.3. For the DDP, P2P systems typically rely on a Data-delivery plane (LAN/Intranet/WAN) Pub1 X Y Pub-Prx-Mech proxy X, Y updates Z W Sub1 Sensor app or firmware PubN Sensor app or firmware Pub-Prx-Mech proxy Z, W updates FIGURE 4.3 P2P WAMS-DD. Power app IP IP Y W Proxy Pub-Prx-Mech Y, W updates IP IP IP SubN IP IP Power app Z X Proxy Pub-Prx-Mech Z, X updates GridStat 79 combination of IPMC and Ethernet broadcast to be as eficient as possible. Note that in Figure 4.3 we omit the DMP, which is often not present as a separate core entity in P2P systems; the edge mechanisms collectively implement it. One other thing to note in Figure 4.3 is that the controllable mechanisms for affecting trafic lie at the edges, in the proxies. Certainly a P2P WAMS-DD will exploit IPMC as much as possible, but this has its limits, as described previously. Because its control mechanisms are at the edges, both QMD and QR are communicated to other proxies that collectively provide the delivery guarantees. Similarly, the only WAMSDD-speciic mechanisms are in the proxies, so control messages also must go there. In Figure 4.3, to help aid understanding, the dashed gray and the gray trafic lines are omitted. In practice, such trafic lines would be delivered via the DDP using IP routers. P2P pub-sub systems have an advantage in smaller- and medium-sized deployments, but, for larger scales, the lack of mechanisms in the backbone core for rate downsampling and quicker fault tolerance, adaptation, and instrumentation limit their abilities to achieve extremely low latencies in the presence of failures. A federated combination of P2P and BB pub-sub systems has the potential to offer much of the best of both worlds. Here, P2P pub-sub systems are employed near the edges, that is, within a single utility or sometimes within an ISO. Between utilities or ISOs, BB pub-sub systems are used in order to support higher throughputs and the lowest possible latencies over distance. A federated amalgamation of P2P systems would feature a globally unique namespace for variables and utilities, and could seamlessly pass messages with standardized wire and message formats [1]. 4.3.2.4 Other Middleware SEcure Data-centric Application eXtensible (SeDAX) is low-level middleware that provides both pub-sub and querying capabilities [43]. It is implemented with an overlay network on top of TCP/IP that is organized into a Delaunay Triangle network. SeDAX mainly targets the distribution side of the power grid, and as such aims to handle many more hosts than are present in a mission-critical WAMS-DD. Similarly, its availability requirements for lows are only 99.5%, which is far from the six nines required by DR4b (which comes from the Electric Power Research Institute in [4]). Finally, it does not provide subsecond latency, let alone strong latency guarantees. It is thus not appropriate for WAMS-DD. Another category is called streaming queries (also known as complex event processing). It consists of a network of computer nodes that manipulate data streams through continuous queries in order to selectively propagate data, merge streams with existing data, or store data in a distributed database. Such systems are not designed to provide hard E2E WAN guarantees (DR1) with per-message granularity (DR5A) while tolerating failures (DR5B). Given their intended application domain, they also do not follow most of the IGs. Finally, over the last decade or so, crude middleware based on web technologies such as hypertext transfer protocol (http), extensible markup language (XML), and “web services” has become popular in other industries, in large part because it is easy to get around irewalls with it. We note that scalability and throughput of such systems are highly questionable due to the many integration layers they typically add to make it possible to glue just about any application to another [44]. Ken Birman, 80 Smart Grids a leading expert in reliable distributed computing whose software has been widely ielded, notes in [45] (emphasis ours): It doesn’t take an oracle to see that the mania for web services, combined with such rampant online threats, contains the seeds of a future debacle. We’re poised to put airtrafic control, banking, military command and control, electronic medical records, and other vital systems into the hands of a profoundly insecure, untrustworthy platform cobbled together from complex legacy software components. Unfortunately, one can add the smart grid to this list: a number of utilities and organizations see web services as a key enabling technology for the smart grid (e.g., [12,46]). This is unfortunate because it inadvertently rules out building an extremely low-latency, extremely high-availability, extremely high-throughput WAMS-DD suitable for the more challenging closed-loop applications. This is generally done because power engineers, or IT staff at utilities (which tend to be understaffed), buy into the hype. But is it really wise to have the operations of increasingly stressed power grids depend on a technology whose raison d’être is to circumvent irewalls? We think not. Caveat emptor! 4.3.3 EXISTING ELECTRIC SECTOR COMMUNICATIONSRELATED TECHNOLOGIES AND STANDARDS Middleware is rarely used in today’s electric power systems, despite being considered best practice in many other industries for 15–20 years [1]. Further, the extreme conditions for closed-loop applications are far more stringent than in any other industry. Additionally, it is an unfortunate fact that almost all of the communications standards in the electric sector have been developed by power engineers or others who are not expert at building communications protocols, are unaware of the state of the art and practice, and so on. It is not surprising, then, that there seem to be no networking technologies developed for the power grid that meet all of the DRs above. Part of this limitation is because commonly used power technologies are intended for a substation scope, with the QoS+ “mechanism” being overprovisioning of bandwidth. When moving from a LAN to a WAN environment, there are implicit design decisions that cannot be solved by layering a new “WAN-appropriate” API over existing LAN-based protocols [25]. We now overview some of the more common power protocols and standards related to communications. Object linking and embedding for process control uniied architecture (OPC UA) [47] was designed for substations. It uses TCP, which was not designed for predictable latency and does not support multicast. Subscribers and publishers “ping” each other to verify whether the other is up, which not only does not scale but also ignores best practices for pub-sub systems. Institute of Electrical and Electronics Engineers (IEEE) C37.118 is a standard for synchrophasors that includes standard message formats. C37.118 is being revised to allow different data delivery mechanisms to be used. If successful, then C37.118 synchrophasor updates should easily be deliverable by any WAMS-DD transport. The IEC 61850 communications standard was also designed primarily for applications associated with a substation automation system (SAS). It was conceived and GridStat 81 created by protection providers in order to move data and information to, from, and among intelligent protection, control, and monitoring devices instead of legacy SCADA and remote terminal unit (RTU) methods. IEC 61850’s companion common information model (CIM), IEC 61970 Part 3, can potentially be of use in a WAMS-DD, especially when the harmonization with C37.118 is completed, in particular in helping automate default QoS+ settings and perhaps adaptation strategies for a wide variety of sensors and applications that use them. This would be signiicant value added. Further, if IEC 61850 manufacturing message speciication (MMS) and generic object-oriented substation events (GOOSE) APIs are successfully extended across the WAN, then IEC 61850 may well be able to successfully use a WAMS-DD transport. This would be of huge beneit to the many utilities that have a lot invested in IEC 61850 applications. However, this will only be true if the WAMS-DD transport is carefully designed to meet the DRs in Section 4.2.2. Unfortunately, this seems doubtful for a number of reasons, including the following: • Presently, IEC 61850-90-5 discusses delays of 50–500 ms, which completely rules out its use with closed-loop apps outside a single substation and also SIPS/remedial action schemes (RAS)/special protection systems (SPS), including those from Chapter 3. • Overall, the standard seems to many observers with a computer science or software engineering background far more complex than it has to be, given the problem it is tackling. Indeed, to the authors it seems more like something that a mechanical engineer in 1975 would have written, not a software engineer from 1995 (let alone today). • Vendors do not, perhaps without exception, implement the full standard. That makes interoperability between vendors very problematic. • There is no reference implementation and intervendor test suite, which makes true cross-vendor interoperability almost impossible because vendors have to test implementations pairwise, which is in turn problematic because, as noted above, few if any implement the full standard. • There are no tools for coniguring and otherwise using the standard that work well (or even fairly completely) across different vendors’ implementations. This, of course, is totally contrary to the goals (and hype) of the standard. • It takes twice the bandwidth to deliver synchrophasor data as C37.118 does, with no extra useful information. • Subscriber applications have to be able to detect missing and duplicate data. It would be much better if middleware did this, so that application programmers (many of whom are power engineers, or at least are not very familiar with distributed computing) did not have to do it themselves. • GOOSE authentication originally mandated the use of RSA signatures. However, they are too slow for closed-loop applications even with a highend PC, let alone a more modest embedded device in a substation (and one that cannot devote all its time to RSA calculations!) [48,49]. There is no silver bullet here: approaches that use a shared key for each publisher and its subscribers are vulnerable to a subscriber spooing a publisher if they do not 82 Smart Grids have other help from other mechanisms. The one solution we know of that works fast enough is described in [50]. It only adds 10 ms of latency, and has been integrated into GridStat, and thus is suitable for use with closed-loop applications such as SIPS. This is discussed in Section 4.7.7. • With 61850 in a substation, developers and deployers have to be very careful that the multicast (which is done there via Ethernet multicast) does not overload small embedded substation devices (which do not even need to receive most or almost all of the trafic). Over a WAN, that kind of spamming is much more likely to cause bandwidth problems than CPU problems. • It was designed by a committee with a “kitchen sink” approach before anyone had done close to a full implementation, which is very problematic. Indeed, Internet pioneer David Clark famously stated about the Internet’s development and the Internet Engineering Task Force (IETF) that deines its protocols [51]: We reject: kings, presidents, and voting. We believe in: rough consensus and running code. MMS also does not have data-delivery mechanisms. It can map onto the OSI protocol stack (which was not adopted in practice) and TCP/IP; see [26,27]. An information architecture for the power grid is proposed in [52], which contains an analysis of 162 disturbances between 1979 and 1995. This paper demonstrates that information systems have an impact on power grid reliability and points out major deiciencies in the current communications scheme. The paper contains proposals for different ways to structure interactions between control centers and substations, and it also contains reliability analyses of different schemes. However, it does not propose communication mechanisms and relies on off-the-shelf network technologies, which do not meet many of the DRs and IGs. Sadly, this is common, almost universal, practice in communications mechanisms and analysis in grids today. 4.3.4 SUMMARY The coverage of the DRs and IGs by various commercial and research technology is given in Table 4.2. The columns of this table are existing networking and middleware technologies, while the rows are the DRs and IGs outlined previously in this section. The table cells denote how well the given technology meets the given IG or DR. These have the following values: Y, yes; —, no; S, some; L, likely (but not conirmed); ?, unknown (documentation is not suficient); D, doubtful (but not conirmed, because the product has insuficient documentation or it is a broad product category but not mission critical); F, future plans (architected for this); and φ, not applicable (and thus, by deinition, does not provide). In some cases, it is not possible to discern the exact coverage of a DR or an IG due to lack of detail in a research paper, product documentation, and so on. However, rather than leaving them unspeciied, if we believe we understand the technology’s goals, environment, and so on enough to speculate, we grade them as “L” or “D” rather than simply “?”. This is possible because it is highly unlikely that the developers of the mechanisms inadvertently overdesigned them to meet WAMS-DD requirements that the designers were not even aware of! 83 GridStat 4.4 NASPINET The North American Synchrophasor Initiative (NASPI) is a government–industry consortium dedicated to effective deployment of synchrophasors in the United States. In this section, we describe its WAMS-DD initiative. Unfortunately, for a number of reasons, including budgetary, regulatory oversight, and technical culture, as of late 2013 there has been no deployment of anything resembling a full-features NASPInet, that is, one implementing all the DRs (and ideally most or all IGs) in Table 4.2. Nevertheless, it is an example worth studying. 4.4.1 ARCHITECTURE The NASPI network (NASPInet) concept is the only effort worldwide we are aware of that is dealing with E2E WAMS-DD issues at a more than supericial level. To support the use of synchrophasors, NASPI has been developing the notion of NASPInet, which has two main components: the data bus (NnDB) and the phasor gateway (NnPG). The NnDB is the electricity version of what is sometimes called an enterprise service bus (ESB), which provides communications services for business-to-business exchanges. NnDB satisies the DRs described earlier in this paper if it is implemented with appropriate technologies. The NnPG is the edge component of Nn, interfacing the utility or ISO to the NnDB. This is depicted in Figure 4.4. Note that, while NASPInet was originally conceived as being only between utilities (and ISOs/RTOs, of course), it is not limited to this. The same technologies would work in a large utility that wanted to have much more control over its network Monitoring center X Utility A Archive Apps PMU Phasor gateway Apps Utility C PMU PMU Phasor gateway PMU Historian NASPlnet data bus Utility B Apps Phasor gateway Historian Monitoring center Y Phasor gateway Apps PMU PMU PMU PMU FIGURE 4.4 NASPInet conceptual architecture. (Courtesy of NASPI.) Archive 84 Smart Grids and applications to provide WAMS-DD properties. This would be particularly important for closed-loop applications. 4.4.2 SERVICE CLASSES Five initial service classes have been identiied for the NnDB in recognition of the fact that different kinds of trafic with different DRs must be carried: 1. 2. 3. 4. 5. Feedback control (e.g., small signal stability) Feedforward control (e.g., enhancing state estimators with synchrophasors) Postevent (postmortem event analysis) Visualization (for operator visibility) Research (testing or R&D) Each class has associated qualitative requirements for such properties as low latency, availability, accuracy, time alignment, high message rate, and path redundancy. This is shown in Table 4.3. Distinguishing the classes in this way is an important irst step for a WAMS-DD system. The performance and availability requirements in Chapter 3, and the DRs and IGs in Section 4.2, can be considered a signiicant reinement of these NASPInet trafic classes. However, we note that the NnDB classes also consider the lowest required latency to be 100 ms, which is demonstrably insuficient for some applications. One issue with the class deinitions is that a customer, such as a utility, an ISO, an RTO, or the North American Electric Reliability Council (NERC), cannot specify only what it wants from a telecom provider, for example, a “Class A” network, and be done with it. This will not necessarily result in a WAMS-DD that meets TABLE 4.3 NASPInet Trafic Classes and Their Attributes Real-Time Streaming Data Historical Data NASPInet Trafic Attribute Class A Feedback Control Class B Feedforward Control Class C Visualization Class D Postevent Low latency Availability Accuracy Time alignment High message rate Path redundancy 4 4 4 4 4 4 3 2 2 4 2 4 2 1 1 2 2 2 1 3 4 1 4 1 Class E Research 1 1 1 1 1 1 Source: Hu, Y. et al., Data bus technical speciications for North American synchrophasor initiative network (NASPInet), May 2009. https://www.naspi.org/File.aspx?ileID=587. Note: 4, Critically important; 3, important; 2, somewhat important; 1, not very important. GridStat 85 the requirements across multiple trafic classes. For example, if too much trafic of “easier” classes is on the network, you will not get Class A guarantees. Rather, to provide the DRs identiied above, one needs to perform resource management within the data-delivery service. Application management and network management components must account for all trafic associated with each subscription using a given level of QoS+. This is part of enforcing a number of IGs, including IG8 (exploit trafic knowledge), IG9 (systematic, quick, internal instrumentation), IG12 (complete perimeter control), IG13 (reject unauthorized packets quickly and locally), and IG20 (manage aperiodic trafic). Further, implementing this requires middleware, as we discuss next. Finally, we note that, despite the US government spending more than $4 billion of federal iscal stimulus money on the power grid over the last 5 years, there has been no signiicant progress in deploying a full-featured NASPInet. There are a number of reasons for this. First of all, advanced communications are something that organizations are not bureaucratically set up to deal with. Second, power engineers and researchers almost always assume that communications are good enough. Third, through the imperfect and power-centric bureaucratic process, through no fault of its own the company hired to write a NASPInet speciication, while excellent at power issues, had zero experience with anything but the old hard-coded grid communications; nobody working for them has even a BS in computer science, or even an undergraduate minor, despite the fact that WAMS-DD is very advanced computer science. Further, even the most ambitious communications upgrade as part of this investment is not only not using pub-sub middleware, it is not even using IPMC. This was because the developers had so many problems getting vendor software running with 200 PMUs, rather than the 10 PMUs it was used to, that they had no time to signiicantly modify the communications. They know that with just point-to-point communications they will not scale to many more PMUs. Further, they are only using MPLS to provide QoS. As noted above, it is very deicient for WAMS-DD, and they are de facto ruling out any closed-loop applications. As demonstrated in this chapter, a full-featured WAMS-DD is very possible, but it has to be done by computer scientists leveraging the state of the art in real-time distributed computing, not by power researchers, or, for that matter, by software vendors repackaging an existing pub-sub infrastructure from other domains. However, the pressure is building: grids are becoming more stressed each year for a host of reasons that can be mitigated by a proper WAMS-DD. 4.5 WHY WAMS-DD MUST INCLUDE MIDDLEWARE, NOT SOLELY NETWORK LAYERS In this section, we explain how implementing a WAMS-DD requires the use of middleware. In general, according to the stated goals of the smart grid community, middleware is required [1] (this paper won a best paper award for demonstrating this proposition). Further, it is simply not possible to implement a WAMS-DD without information provided by middleware, sometimes augmented by application monitoring. In other cases, although it is possible to build some features and properties without middleware, with the help of middleware the lower-level mechanisms can 86 Smart Grids do a much better job. In general, the network level only knows about packets and IP addresses, not middleware-layer application variables (e.g., published sensor variable updates). It also knows nothing about the power applications that subscribe to these updates, and their importance in different situations. For example, GridStat’s systematic adaptation described in Section 4.7.5 cannot possibly be implemented without help above the network layers. In this section, we consider why middleware in general is required for WAMS-DD. We next describe how it is needed when QoS+ is required, as, of course, it is in any WAMS-DD. We next discuss how middleware is needed to properly implement some of the DRs and IGs. Finally, we discuss how middleware is very helpful in integrating legacy systems. 4.5.1 MIDDLEWARE IN GENERAL Middleware is required to accrue the beneits overviewed in Section 4.3.2. Additionally, in practice, in any system of moderate size and complexity there will be products from different vendors. While they may implement the same protocols, their coniguration and management may well be different. For example, even router vendors will admit that the way you conigure IPMC is different across vendors. Thus, if for no other reason, most such systems will have at least a thin layer above the network for applications to use. A utility, therefore, has three choices regarding this: 1. Have each application programmer recreate its own thin layer (these are sometimes called “application-level protocols” by network researchers) 2. “Roll your own” middleware in-house 3. Use commercial middleware Application programmers already have enough to deal with, so making them also handle Option 1 is highly inadvisable (but common practice in the electricity sector, regrettably). Rolling your own middleware is certainly better than Option 1 because it allows the implementation of the thin layer to be shared across application programs. However, utilities may not have sophisticated network programmers to do this, and if they do, most likely their time could be better spent on other tasks. This leaves us with Option 3: commercial middleware. Here experts in networking, middleware, QoS+ mechanisms, interoperability, and so on create middleware that is reused over many companies and projects. And, as noted in Section 4.3.2, programmers then are more productive and the code they write is more portable and manageable. 4.5.2 MIDDLEWARE AND QOS+ When QoS+ properties are required, as of course they are in a WAMS-DD, it is even more important to use middleware. Without it, application programmers have to combine lower-level QoS+ mechanisms in arbitrary ways. For example, how is an application programmer going to know how to combine network QoS from MPLS or int-serv GSs with Internet protocol security and nonrepudiation? These mechanisms GridStat 87 were designed independently of each other and interactions between them can be very subtle. It is thus a really bad idea to assume that application programmers can combine them in reasonable ways, knowing any inherent trade-offs or limitations of their combination. By far the better way to provide QoS+ properties in a WAMS-DD is for a middleware developer to combine multiple mechanisms, evaluate and measure their combination, document any trade-offs and limitations the application programmer should be aware of, and so on. In addition to providing QoS+ properties in the steady state, there are additional reasons why WAMS-DD requires middleware: coordinated adaptation driven from the top down. In the face of failures, cyberattacks, or other runtime IT problems, it is extremely undesirable to have different mechanisms adapting independently of each other and without any higher-level guidance. For example, having ACK and NACK messages clog the WAMS-DD is a really bad idea. It is much better to have coordinated, E2E adaptation driven by knowledge of the applications via knowledge captured in applications’ middleware. 4.5.3 MIDDLEWARE AND DRS AND IGS As mentioned previously, a number of the DRs and IGs cannot be implemented properly, or even at all, without the help of middleware. These include the following: DR2 (future prooing): Applications that use middleware can exploit new and better QoS+ mechanisms as they become available, because the applications are not locked into lower-level speciic QoS+ mechanisms. The applications do not need to be reprogrammed or recompiled, just to have a new version of their middleware library linked in. DR5B (ultralow latency tolerating benign failures): Providing this high level of fault tolerance without jeopardizing low-latency guarantees requires knowledge of the applications using the WAMS-DD, not just socket-level information on packets. IG2 (optimize for rate-based sensors): In order to optimize for rate-based sensor data delivery, it is necessary to know the rates that the different applications require. QoS-enabled middleware APIs capture this, while network-level mechanisms are generally unaware of application rates. IG3 (provide per-subscriber QoS): Similar reasons to IG2 above. IG4 (provide eficient multicast): Similar reasons to IG2 above. IG7 (provide E2E interoperability across different or new IT technologies— multicast, encryption, authentication, network-level QoS, etc.): Network-level mechanisms do not cover the entire space here, and, with few exceptions, cannot generally be composed with other kinds of network mechanisms providing the same property (e.g., latency) without a higher-level layer above them both, which is, of course, by deinition middleware. IG12 (enforce complete perimeter control): Network QoS mechanisms generally do not know much about the trafic delivered on them, including their rates. Further, if overrate trafic comes to a router that knows the rate, it is 88 Smart Grids still consuming resources on that router rejecting it. Middleware with QoS APIs can, of course, capture rate and other QoS parameters. Further, middleware generally uses a proxy object that is returned when the application connects to the middleware. The proxy object is then what the application (publisher or subscriber) uses to send and receive sensor updates, respectively. This proxy object can perform rate-based policing to ensure that no more than the guaranteed rate escapes the application and taxes the irst hop across the network. IG18 (support multiple QoS+ mechanisms for different operating conditions and conigurations): In a big WAMS-DD, no single mechanism for providing a QoS+ property such as latency will be present everywhere in the WAMS-DD. Sometimes the path from publisher to subscriber has to traverse multiple QoS+ mechanisms, for example, MPLS and Nimbra. Similarly to IG7, a layer is needed above the network layer to sensibly connect these different mechanisms providing the same property, and to know the limits and trade-offs inherent in their combination. IG20 (manage aperiodic trafic): Aperiodic trafic, also called event triggered or condition based, must obviously be isolated from rate-based trafic (also called time triggered or periodic). This can be done at the network layer. However, it is also necessary to manage that trafic with knowledge of the data and applications using the data. This is because in many circumstances there may be more data than can it in the bandwidth allocated to them. Fortunately, there is often a lot of redundancy in this trafic (e.g., a large number of alerts/alarms sounding from a single root cause, or fairly unimportant messages that can be logged and ignored, at least for some time, in some circumstances, or both). Carrying out this management requires middleware to capture the semantics of the data and application. It also requires higher-level management to intelligently aggregate the data using policies or other high-level mechanisms, but this requires the information that middleware provides. 4.5.4 MIDDLEWARE AND LEGACY SYSTEMS Middleware is often used not to write complete systems from scratch, but rather to integrate legacy systems that were not designed to interoperate with different kinds of systems. It does this by adding a middleware wrapper, which surrounds the legacy subsystem and maps down from a high-level middleware interface to the interface speciic to a given subsystem. When middleware wraps a legacy system in such a way, the APIs and also the management infrastructure can be written at a high level consistently across the different legacy subsystems, because the middleware shields them from the lower-level details of the individual legacy systems. This allows the subsystems to interact in a very high LCD (the middleware layer and its high-level abstractions), with strong type checking, and allowing the use of exception handling mechanisms across different programming languages. This is of great beneit. As an example, client–server-distributed object middleware such as CORBA is often used to wrap legacy servers. In this way a client written in Java or Python can 89 GridStat invoke a server object on a server written in C#, C, or FORTRAN, even though the last two languages are not object oriented. The previous example involved client–server interactions, but the technique of wrapping legacy subsystems with middleware also applies to pub-sub infrastructures such as WAMS-DD. It can be highly beneicial to do so, because communications infrastructures for utilities can be a jumbled mess of communications subsystems that were not written to work well together (not only due to the relative backwardness of the industry on IT matters, but also because many of the bigger vendors believe that intervendor interoperability can provide a path for them to lose their lucrative vendor lock-in). Such wrapping is a necessity, not just to be able to create a coherent WAMS-DD out of these legacy subsystems, but also because in the future no one mechanism for a given QoS+ property (such as latency) is ever going to “take over the world” and completely displace all others. Given the large investment in the legacy systems, and the fact that (at least if utilized correctly) they can still provide signiicant value, it will always be a case of trying to seamlessly interoperate across multiple subsystems. We give an example of this in Figure 4.5. Here an existing utility communications infrastructure includes a large number of legacy protocols such as MPLS, microwave links, and others (the “others” could include broadband over power lines/power line communications, satellite links, and even 4G cellular, for example, to remote substations where it is impractical to run iber or copper). This existing infrastructure has been strengthened by programming some of the existing routers using OpenFlow and by adding stronger guarantees with subsystems of GridStat FEs and Nimbra devices. This could in turn be managed as a coherent whole by GridStat’s management plane. In this scenario, non-GridStat subsystems look to GridStat as a single link between two FEs. However, the management plane can exhibit control over them to strengthen QoS via a control wrapper that has a high-level middleware control API, and the wrapper then maps down to the subsystem’s speciic control APIs (the control wrappers and communications Management plane MPLS Publisher 1 Publisher 2 C o m m o n M W Publisher N A P I Nimbra MPLS Subscriber 1 OpenFlow Legacy microwave Subscriber M GridStat forwarding engines = Cntl wrapper Nimbra = Data wrapper FIGURE 4.5 Middleware integrating legacy and new communications subsystems. 90 Smart Grids are shown in light gray). Finally, the data wrappers at the boundaries of the subsystem not only provide a common, high-level middleware API but could also be used to carry out rate-based policing, similarly as described above, using middleware proxies. Given such a high-level coherent WAMS-DD and a management plane such as GridStat’s, the different subsystems can be treated as a uniied whole, and route across different subsystems, taking into account not only the performance characteristics of the subsystem but also its reliability, and, of course, considering aggregate bandwidth. Further, the subsystems in Figure 4.5 may have fairly different failure characteristics and security vulnerabilities, which can greatly help the overall E2E data availability from publisher to subscriber. The reader may be curious why old microwave links, which have much lower bandwidth and often higher drop rates, would even be included. However, they have very different failure characteristics from most of the other subsystems, and this can add a lot of value, as long as their maintenance is not becoming cost-prohibitive. These links do have much less capacity than the other technologies, so only the most important trafic must be sent over them, perhaps at a lower rate. Of course, this requires middleware to capture these semantics for the management system. Indeed, it is crucial to note that, even if a WAMS-DD has no stronger QoS mechanisms than MPLS or even simple IPv4, it will still need the management and policing, wrappers, and so on that are shown on the edges below. If it does not have these, things may work well in the calm steady state, but when things get bad the predictability of the latency and other properties is extremely suspect. In short, when you need it the most you can count on it the least. Sadly, this is the situation at almost all utilities today. The only real QoS “mechanism” is massive overprovisioning of bandwidth. This works quite well most of the time, but when there are benign IT failures, let alone cyberattacks, the WAMS-DD is likely to perform quite unacceptably. This in itself could be a major vulnerability: terrorists can not only try to change actuator and relay settings to cause problems in the grid, but they may also attack the WAMS-DD so that it is much more dificult for utilities and ISOs/RTOs to understand what is going wrong and ix it before a large blackout is caused. Indeed, such attackers would, of course, attack when the grid is already severely stressed. The conditions under which a grid will be stressed, and what power assets can cause most problems if changed, are relatively easy for a determined adversary with a long-term view (e.g., Al Qaeda) to ascertain long ahead of time. 4.5.5 MIDDLEWARE AND AVOIDING QOS STOVEPIPE SYSTEMS Finally, we close this section with a system-level admonition that is crucial for the long-term beneits that WAMS-DD can provide, if done right. A stovepipe system is a legacy system that is comprised of different parts that are very dificult to combine or refactor. They were identiied as major problems in big systems 20–30+ years ago, and much software engineering and system architecting has been devoted to avoiding them on many newer systems (at least in other industries that are not as far behind in best practices as the electricity sector is). GridStat 91 But, given that WAMS-DD has to provide QoS+, and the mission-critical nature of its applications, stovepipe systems have another dimension. Bakken et al. [1] provide the following deinition: QoS stovepipe system (QSS): A system of systems whose subsystems are locked into low-level mechanisms for QoS and security such that 1. It cannot be deployed in many reasonable conigurations, or 2. Some programs cannot be combined because they use different lowerlevel QoS mechanisms for the same property (e.g., latency) that cannot be directly composed, or 3. It cannot be upgraded to “ride the technology” curve as better low-level QoS and security mechanisms become available Of course, it is crucial that WAMS-DD installations avoid QSS, and middleware, while no “silver bullet,” is a key enabling technology for this. It is for this reason (and others from this section) that many military programs in the United States and elsewhere require the use of QoS-enabled middleware; examples can be found in [1]. 4.6 SECURITY AND TRUST FOR WAMS-DD One of the United States Department of Homeland Security’s strategic plan objectives for the iscal years 2008–2013 has been to protect and strengthen the resilience of the nation’s critical infrastructure and key resources. WAMS-DD, being a middleware that could be deployed in a critical infrastructure such as the power grid, must provide support for key security services (authentication, conidentiality, integrity, and availability) not only at the system level, but at the data level as well. There are many points of interaction among a variety of participants, and a local change can have immediate impact everywhere. Thus, to cope with the uncertainties associated with security, there must be provision for a trust service that assesses the trustworthiness of the received data that come through nontrivial chains of processing. Trust, when used properly, has the potential to enable the entities that operate within critical infrastructures to share conidential, proprietary, and business-sensitive information with reliable partners. Even though WAMS is becoming one of the popular technologies for upgrading the electric power grid, still there is no consensus on a security standard or even a set of comprehensive security guidelines that will guarantee the secure generation, distribution, and consumption of data. It is a nontrivial and rather challenging task to provide trustworthy and secure interactions in such operating environments, where private, public, and national entities must collaborate in such a way that information is not compromised due to either malicious or accidental attacks. However, research has been done on the security aspects of an envisioned power grid communication infrastructure that could be used as a basis to form the security requirements for WAMS-DD. It is beyond the scope of this chapter to map the security spectrum for WAMS-DD. One of the irst research efforts to deine the security, trust, and QR in a datadelivery middleware tailored for the power grid is described in [53]. Starting with the application requirements for a variety of interactions (e.g., communication 92 Smart Grids between substations and control centers, and communication between control centers), the authors proceed to specify the implications of the requirements for security, trust, and QoS. The work in [54] complements the aforementioned speciications by presenting a conceptual lexible trust framework that models an entity’s trust as a relation whose state is updated as relevant conditions that affect trust change. Trust is expressed as a set of security, behavioral, and QoS properties. The novelty of the model is its ability to integrate trust reasoning and monitoring into a single model by extending the traditional concept of trust conditions into more expressive expectations, which include not only expected values for particular properties but also covering, aggregating, and triggering mechanisms that manipulate the observed value. In addition, and unlike other approaches, the model derives E2E trust assessments without using transitive indirect trust explicitly in the derivations. A recent work [55] presents a theoretical model based on Bayesian decision theory that automatically assesses the trustworthiness of entities in the electric power grid. One of the few published works on WAMS security is [56]. The authors interviewed the owners of two WAMS networks in order to identify the primary implementations of WAMS, and thus address security challenges in the light of each implementation. The three WAMS implementations are an isolated WAMS network, an isolated WAMS that interfaces with the control center network, and, inally, a WAMS fully integrated into the control center network infrastructure. An analysis of WAMS security concerns is presented, accompanied by a number of threat scenarios. The vulnerabilities identiied by the authors are addressed in terms of these three implementation approaches because their impact differs in each case. 4.7 GRIDSTAT GridStat is a data delivery service designed to support the DRs discussed in this chapter. Its research results have inluenced the shape of NASPInet. GridStat research started in 1999 by looking at the QoS+ requirements of innovative power applications being developed by power researchers and analyzing closely what the state of the art in applied distributed computing systems could support. After signiicant gaps were identiied, the detailed design and then development of the GridStat framework began in 2001. GridStat is a BB pub-sub system that meets all of the DRs from this chapter except 5C: Tolerating Cyberattacks, which it does to a modest degree (by redundant paths); covering this more completely is near-term future research. GridStat also implements all but three of the IGs, which have similarly been planned for and are also near-term future research. In this section, we irst provide an overview of GridStat. Next, we describe GridStat’s novel mechanisms that allow different subscribers to be delivered different QoS+, which is crucial for a WAMS-DD. We then describe condensation functions, a mechanism that allows arbitrary computations to be placed in a WAMS-DD coniguration. After this, we describe GridStat’s mechanisms for rapidly adapting its subscription base. Finally, we describe GridStat’s remote procedure call (RPC) mechanism and then overview its security and trust infrastructure. 93 GridStat 4.7.1 OVERVIEW 4.7.1.1 Capabilities The architecture of GridStat is given in Figure 4.6. The architecture is modular and it consists of two modules, the data plane and the management plane. Each of the modules is responsible for part of the overall functionality of the framework. GridStat’s data plane is responsible for delivering the sensor data. A single sensor datum is called a status variable, and can be subscribed to by multiple subscribers. The data plane consists of publishers, which generate the sensor data, subscribers, which consume the data, and a graph of FEs, which are specialized middleware-layer routers (“brokers” in pub-sub parlance) that deliver the status variables. GridStat’s management plane is responsible for managing the data plane, including making admission control decisions. It is organized into a hierarchy of QoS brokers (QB) that is designed to map directly onto the geographically based hierarchies into which the power grid and other critical infrastructures are normally organized. A leaf QB is the lowest level in this hierarchy and is responsible for, and directly connected to, a collection of FEs that are organized into a single administrative domain called a cloud. Brokers can host policies for resource allowance (how much bandwidth an organization or app can receive), security permissions (potentially default ones overridden in some cases), adaptation strategies, data aggregation, and specifying other coniguration and operational issues. 4.7.1.2 More Detail GridStat’s data-delivery semantics are novel in three broad ways in order to better support WAMS-DD. First, they are a specialization of the pub-sub paradigm Management plane QoS broker QoS broker QoS broker QoS broker QoS broker Data plane FE FE FE FE FE Subscriber FE FE FE FE FE FE FE FE Publisher FIGURE 4.6 GridStat architecture. FE FE FE FE Publisher Subscriber Publisher 94 Smart Grids tailored to leverage the semantics of periodic streams of data updates. For example, in a normal pub-sub system, an event that is being forwarded from a publisher to its subscribers cannot be arbitrarily dropped, because it is generally impossible to know how it will affect the application programs that subscribe to it. However, in GridStat the subscribers’ required rate, latency, and number of redundant (disjoint) paths are part of the API to subscribe to a given status variable. With this knowledge, the management plane conigures an FE such that it can discard updates, an operation we call rate iltering, when downstream subscribers have received an update recently enough to satisfy their latency and rate requirements. This can potentially save a large amount of bandwidth, because many sensors in the power grid produce updates with a hard-coded rate that is deliberately conservative, that is, at a rate above what its engineers believed any application would require. The absence of rate iltering means that each subscriber has to subscribe to a status variable at the highest rate that any subscriber has to subscribe at, which is, for example, what IPMC requires. With many subscribers, and only a few needing high-rate updates, this can be extremely wasteful in a WAMS-DD. The second novelty is that GridStat implements synchronized rate downsampling (IG5), a requirement that GridStat researchers irst described and that is necessary when performing rate iltering on synchrophasor data. The third way in which GridStat’s data delivery is novel is that different subscribers to a given status variable can be provided with different rate, latency, and number of redundant paths. We call this QoS+ heterogeneity. This allows a power device, such as a relay, that is physically very close to the publisher to be given a high rate and low latency, while remote subscribers that only need to loosely track the variable can be given a much lower rate and higher latency. If this lexibility were not supported, the remote subscribers would have to receive the same high rate of delivery as the device in the same substation, thus consuming unnecessary resources. This is what IPMC requires, because there is no rate iltering of any kind. These operations are depicted in Figure 4.7, where the management plane is not shown as a hierarchy for simplicity. When updates of any publication, such as Y, arrive at an FE, it knows Y’s publication rate and the rate at which each of its outgoing links needs Y. Consider the case where Y is being published by Pub1 at 120 Hz and where Sub1 requires it at 120 Hz but Sub2 requires it at 30 Hz. At the FE (in this example this could be FE4), where updates to Y split off on different links toward Sub1 and, Sub2, the former link will pass on all updates while the latter will drop three-quarters of them. Also, if Sub1 receives Y over two different paths, an update will traverse a different set of FEs to reach it. In the example above, the irst path would be FE0 → FE1 → FE2 and the second path would be FE0 → FE3 → FE4 → FE2. In a proxy (piece of middleware code beneath the application), duplicate updates for Y will be dropped. We are not aware of any system that has such broad coverage, or close to it, of the DRs and IGs, as should be clear from Table 4.2. Systems that provide either application-driven rate iltering or QoS+ heterogeneity are uncommon, and we know of no system that provides both. Yet, as is shown above, in Chapter 3, and in [57], leading-edge and emerging power applications can greatly beneit from having such capabilities. 95 GridStat Data management plane QoS metadata QoS requirements Control Pub1 Sensor app Sub1 X Y GridStat data-delivery plane (managed WAN) Sensor app X Y Proxy Proxy X, Y updates FE1 FE0 FE2 X, Y updates FE3 FE4 PubN Sensor app SubM FE5 FE6 Z FE7 Proxy Z updates FE8 Sensor app Y Z Proxy FE9 Y, Z updates FIGURE 4.7 GridStat FE operation. 4.7.1.3 Performance and Scalability On a 2010-era PC in Java, GridStat adds less than 0.1 ms per overlay hop. Because, in any real deployment of GridStat, this would be dominated by the speed of light, it is essentially free, adding no more than 1 ms over the speed of the underlying networks, because there would be no more than about 10 hops. This PC version can handle ~500 K forwards/s at each FE. This could be further improved if we made the assumption that rates that can be subscribed to are power-of-two multiples of a base rate; for example, 15, 30, 60, …, Hz. Further, given how simple the rate-based forwarding logic is and how ineficient Java is at high-performance input/output, a C version (which is planned for the very near term) should achieve 10–20 times the Java version, or roughly 5–10 M forwards/s. This is likely more than adequate for the backbone needs of major grids for at least 5 years, given the slow growth in interutility sharing. On 2003-era network processor hardware, it adds ~0.01 ms/hop and scales to a few million forwards/s [58]. Using 2013-era network processing hardware (which is at least 10 times more scalable than 2003 versions) and a small cluster, these results should easily be extendable to achieve 50–100 M forwards/s, while rejecting unauthenticated messages, monitoring trafic patterns, and checking for (and reporting) evidence of intrusions and cyberattacks. Custom hardware implementations in ieldprogrammable gate array or application-speciic integrated circuit would likely be highly parallelizable as well as “unrolling” the logic in the software FE’s two-deep loops. It is thus likely to support signiicantly higher than 100 M forwards/s, and likely up to line speeds on a 2.5 Gbps or faster iber link. This is possible because the logic in a GridStat FE is simpler than a general-purpose IP router: it is doing less deliberately! 96 Smart Grids 4.7.2 RATE FILTERING FOR PER-SUBSCRIBER QOS+ In order to get an operational view of a wide-area system, its monitoring and control system must be able to deliver a snapshot of the state of the entire system (or at least a snapshot of a set of interesting measurements). There are two requirements in order to achieve this. First, the measurements must be taken at the same time. Second, as described above, if any rate iltering of the measurements is performed by the communication system, the temporal relationship between the events must be preserved. GridStat’s forwarding logic provides deterministic multicast rate iltering of event streams for temporally related status variables. This feature is provided by the underlying data structure and the routing algorithm that the FEs are using; hence it is transparent to the end points and always activated for all events. In addition, there is no need for any coordination between the FEs or intervention from the management plane. The multicast feature is a side effect of the data structure used in the FE, and as a result no extra computational resources are needed. The computational resource usage by the iltering algorithm is further analyzed in [59]. 4.7.2.1 Forwarding Algorithm The forwarding algorithm is performed for each event that is received by an FE. The message format includes a header that stores the variable ID and the time stamp when the speciic measurement was taken. For each subscription that is established over a speciic FE, the following information is included during the subscription set-up command: variable ID, published interval, subscribed interval, and outgoing link. With this information the data structure for the forwarding algorithm is built (for more information see [59]). The three steps that are taken during the event forwarding are the following: 1. The irst step is to add to the event’s time stamp half of the publishing interval, that is, shift the event’s time stamp by half the publishing interval to the right. 2. From the result of Step 1, take the mod of a subscription interval. Note that this is done for all of the unique subscription intervals for this speciic variable. 3. If the result of Step 2 is less than the publisher interval, the event is forwarded. The three steps can be illustrated with the following pseudocode: if ((event.TS + ceil(pubInt/2)% subInt[0..i]) < pubInt) forward event else drop event 4.7.2.2 Example of the Forwarding Algorithm The example in Figure 4.8 illustrates how events from four variables (p1, p2, p3, p4) that publish events every 50 ms will be forwarded through an FE for a subscriber downstream that subscribed to receive the set every 100 ms. Due to various reasons, 97 GridStat Original event streams: 50 ts = 80 100 ts = 95 p1 150 ts = 105 ts = 120 p3 p2 p4 ts = 130 p1 Events in 100 ms Step 1: 50 ts = 145 p2 p3 150 ts = 105 p1 ts = 120 ts = 130 p2 p3 ts = 145 p1 p2 ts = 45 ts = 55 p3 p4 p1 p1 ts = 70 p2 ts = 80 ts = 95 p3 p4 50 ts = 20 p2 ts = 30 ts = 45 p3 p4 Events in 100 ms p3 p4 Events in 150 ms Step 3: 0 ts = 5 p2 ts = 180 ts = 195 100 ts = 30 Events in 100 ms ts = 170 Events in 150 ms 50 ts = 20 200 ts = 155 p1 p4 Events in 100 ms ts = 5 p4 Events in 150 ms 100 Step 2: 0 200 ts = 155 ts = 170 100 Events in 150 ms FIGURE 4.8 Illustration of the iltering steps. the events from the different publishers are published with slightly different time stamps. According to the algorithm, the following will occur in the FE: 1. The addition of 25 ms (publication interval divided by 2) to the time stamp of each of the events. After the addition, the time stamps that belong to the 100 ms group will have a time stamp of 100–149 ms. 2. The subscription interval is 100 ms, and that means the events in the group with time stamps 100, 200, 300, and so on should be forwarded. Taking the mod of the result of Step 1 results in events with time stamps 0–99. 3. After Step 2, there are two phases with events from the publishers, because subscription divided by publication interval equals 2. The framework will always pick the irst phase, so the result of Step 2 is checked to test whether it is less than the publication interval (the irst phase). If this is the case, the event is forwarded; if not, it is dropped. Events with original time-stamp values between 75 and 124 will be forwarded, while the others (time-stamp values between 125 and 174) will be dropped. 4.7.3 CONDENSATION FUNCTIONS In an environment where multiple entities must cooperate in order to maintain the stability of the system, there will be some common patterns that the various entities will be interested in. Without support from the control and monitoring system, 98 Smart Grids these patterns will have to be determined at the end points. If the determination for these patterns can be provided as a service by the monitoring and control application, a number of beneits can be observed. The most obvious beneit is the reuse of application logic, but also reduction of computational and network resource, as the computation will be performed only once, close to the source, and shared with all the interested parties. GridStat provides the condensation function mechanism, which allows an end user to specify a transformation of multiple raw event streams into a new stream during transmission. This transformation allows the end user to migrate application logic into the middleware layer, with great lexibility. 4.7.3.1 Condensation Function Architecture and Capabilities A condensation function consists of four modules, as shown in Figure 4.9: input ilter, input trigger, calculator, and output ilter, with input ilter and output ilter modules being optional. The input ilter module ilters status events from the input event streams. The iltering process compares the status event value with the upper or lower predeined thresholds, or both, and decides whether to drop the event or not. The trigger module is responsible for receiving events from the input status variables and initiating the calculation function. For each of the input variables, there is an entry in the data structure where the trigger module stores the latest value from each of the variables. The calculator module is initiated by one of the three built-in triggering mechanisms (others can be added). The built-in triggering mechanisms are: time triggered, number of received events, and number of alert events received. The calculator module must be provided by the end user. The user will have to overwrite two pure virtual methods, namely, the Initializer method (initializes the object) and the Calculator (transforms the input to an output). Lastly, the output ilter is similar to the input ilter, with the difference that it ilters the result produced by the calculator. The trigger module implements the same interface as an outgoing link. What this entails is that, during the routing process, an event routed to a condensation function has the same interface as if it were routed to any of the outgoing links. As a result, the condensation function does not interfere with the low of the routing process. Input filter Input trigger Calculator Output filter Filter the input data to either a lower or an upper threshold When should the calculation start, delay or number of input Applies a user-defined function to the arguments Filter the result to either a lower or an upper threshold FIGURE 4.9 Condensation function modules. GridStat 99 4.7.3.2 Condensation Function Development The GridStat framework provides a graphical user interface (GUI) tool that assists the end user in deining a condensation function. If the input stream or the produced output stream uses data types other than those built in, an interface deinition language (IDL) compiler will be used to unmarshal and marshal the events for iltering and calculation. The placement of the condensation function within the data plane is the responsibility of the management plane. The GUI tool sends the speciication for the condensation function to the leaf QB, through an FE. The leaf QB will then ind an appropriate FE and place the condensation there. In addition, it will set up all the subscription paths that are the input for the condensation function as well as registering the resulting event stream as a irst-class variable, just like all the other publications. More detailed information on the condensation function can be found in [60]. 4.7.4 OPERATIONAL MODES Routing algorithms that disjoint path routing with additive constraints (latency) are computationally very complex. It would in practice be very bad if there were thousands of new subscription requests in a crisis: that would likely be an inadvertent but effective denial of service attack on the subscription infrastructure. Fortunately, power grid operators do not choose random sensor variables to start observing during a power contingency. Rather, they have (paper) checklists for each contingency, which were created months or years ago by engineering studies. GridStat’s mode mechanism exploits this fact to enable its subscription base to react very rapidly. It allows multiple forwarding tables to be preloaded in its FEs, then rapidly switched when appropriate [61]. This action is called a mode change. Depending on the GridStat deployment, FEs can utilize several forwarding tables corresponding to the operating modes, while inactive forwarding tables lie dormant until their corresponding mode is activated. GridStat also supports a hierarchy of modes. QBs also inherit mode sets from their ancestors in the hierarchy. A QB will have one set of forwarding tables for each mode set, and will have exactly one active mode in each mode set. 4.7.4.1 Mode Change Algorithms A mode deinition consists of an ID, a name, and a set of data plane subscriptions (a forwarding table), and is owned by a single QB in the management hierarchy. Modes deined and owned by a QB constitute a mode set, and exactly one of the modes in a mode set is active at any time; that is, the operating mode. This means that every QB always has exactly one of its modes active, which could be a default and unnamed mode if no modes are deined. A QB that operates in a mode assures that all +subscriptions (and parameterizations thereof) contained in the subscription set of that mode are active in the data plane by using the forwarding table corresponding to the active mode. A QB can quickly switch forwarding tables in the data plane through modes. Figure 4.10 shows how operating modes are used in both the management plane and the data plane. In this example, QB-A is the parent of both QB-B and QB-C. 100 Smart Grids Management plane QoS broker A Mode sets B: (Stable, unstable) Active modes: Stable QoS broker B Mode sets A: (Green, yellow, red) Active modes: Green QoS broker C Mode sets C: (Normal, waming, critical) Active modes: Normal Data plane B1 B2 Publisher B3 Publisher C4 B4 Subscriber C2 C1 Mode sets: A, B Active modes: (A.Green, B.Stable) C3 Mode sets: A, C Active modes: (A.Green, C.Normal) Subscriber FIGURE 4.10 Mode example. QB-A deines a mode set that contains {Green,Yellow,Red}. However, it has no cloud directly under its control, so this mode set is used to inluence forwarding in its children, B and C. The active mode at this time for QB-A is Green. If it later changes its mode, FEs belonging to its descendants will change forwarding tables. QB-B deines the mode set with labels {Stable,Unstable}, and its active mode presently is Stable. Thus, QB-B will utilize forwarding tables for both A.Green and B.Stable. QB-C deines the mode set with labels {Normal,Warning,Critical}. Its active mode is Normal. Thus, QB-C will utilize forwarding tables for both A.Green and C.Normal. A mode change operation that switches the forwarding tables in all of the involved FEs at the expected time is called a consistent mode change operation. Otherwise, the operation is called an inconsistent mode change operation, and additional recovery mechanisms must be utilized in order to restore the operating modes on the FEs that are considered inconsistent. An inconsistent mode change operation will most likely result in some subscribers not receiving all the events that they have subscribed to, or receiving a few updates to a new mode while still receiving updates that should only have come in the old mode. However, these glitches will only happen for the duration of the mode change, which is very small. The recovery mechanism is provided in order to tolerate some degree of network failures and to eventually ensure consistent mode change operations. The recovery mechanism is triggered by the QB, which detects missing mode change acknowledgments, and attempts to resolve these situations when possible. Two mode change algorithms have been implemented in GridStat: the hierarchical mode change and the looding mode change. The hierarchical mode change algorithm uses the management hierarchy to disseminate mode change operations and gather acknowledgments from FEs and QBs. The hierarchical mode change algorithm enables all subscriptions registered to operate in the coordinator’s current and GridStat 101 new mode to low during the entire mode change. Thus, subscribers with subscriptions registered in both the current and new modes continue to receive status events during hierarchical mode change operations that switch between those modes. The looding mode change algorithm disseminates mode change operations directly out on the data plane by using the limited looding mechanism in GridStat. When an FE receives the operation, it forwards the operation on all outgoing event channels, except the event channel from which it received the operation. As FEs may receive multiple copies of the same operation, redundant copies are discarded. FEs are informed to change from the current mode to the new mode at some future time stamp. The two mode change algorithms provide different trade-offs. The hierarchical mode change algorithm is a resource-intensive algorithm that is split into ive message phases in order to enable transferred subscriptions present in both modes in a mode change, for example, from Green to Yellow, to low. A message phase indicates that the coordinator has to initiate and propagate a mode change phase (message) down to all FEs in its hierarchical scope, and the next phase cannot be initiated until the previous phase has completed. The coordinator must receive aggregated acknowledgments from all FEs and QBs. The looding mode change algorithm, on the other hand, is a best-effort algorithm and disseminates mode change operations directly out on the data plane, where FEs are informed to switch at a predetermined future time. The looding mode change algorithm is eficient, in terms of resource usage and performance, but does not guarantee any subscriptions to low during the mode change operation. The looding mode change algorithm relies on the FEs’ ability to switch modes at the exact same time and therefore requires all FEs to be time synchronized. We now provide more detail on both of the mode change algorithms. 4.7.4.2 Hierarchical Mode Change Algorithm As mentioned earlier, the hierarchical mode change algorithm is divided into ive distinct phases that enable transferred subscriptions to be present in both modes of a mode change operation. Furthermore, the ive phases eliminate any FE overload scenarios during a hierarchical mode change operation. The following steps show how the hierarchical algorithm affects the FEs in a mode change from Green to Yellow: 1. The inform phase: The FEs that have subscribers attached to them inform the subscribers about the upcoming mode change. 2. The prepare phase: The FEs that have publishers connected to them switch to the temporary forwarding table Green ∩ Yellow. The highest subscription interval (lowest rate) of transferred subscriptions is issued in this phase in order to reduce the load on downstream FEs. This phase ensures that subscription trafic that belongs in both modes (Green and Yellow) is forwarded through the FE network. This step only eliminates subscriptions, and hence cloud resources cannot become overloaded. 3. The internal change phase: FEs that have neither publishers nor subscribers connected to them switch to Yellow’s forwarding table. Since all other 102 Smart Grids FEs operate in a temporary forwarding table and only forward a smaller set of subscriptions (in mode Green and Yellow), the FEs can safely switch to mode Yellow without overloading any FEs downstream. 4. The edge change phase: The FEs that are currently in the temporary forwarding table Green ∩ Yellow will switch to Yellow’s forwarding table. 5. The commit phase: The FEs inform their subscribers about the completed mode change. 4.7.4.3 Flooding Mode Change Algorithm The looding mode change algorithm disseminates mode change operations directly out on the data plane by using the limited looding mechanism in GridStat [62]. The leaf QB that is responsible for the speciic cloud initiates the looding by using its embedded publisher. When an FE receives the operation it forwards the operation on all outgoing event channels, except the event channel from which it received the operation. As FEs may receive multiple copies of the same operation, redundant copies are discarded. FEs are informed to change from the current mode to the new mode at some future time stamp. Upon receiving the mode change command, an FE immediately responds with an acknowledgment to its leaf QB. It will then activate the new mode at the destined future time stamp. The looding mode change algorithm offers better statistical delivery guarantees to the data plane than the hierarchical mode change algorithm. This is because of the number of redundant paths in the data plane, and it is thus more resilient to network failures than the hierarchical mode change algorithm. While the hierarchical mode change algorithm attempts to preserve the subscriptions registered in the two involved modes, the looding mode change algorithm switches directly to the new mode, even though some of the FEs have not acknowledged that they will perform the switch. 4.7.4.4 Overview of Performance The two algorithms presented above are quite different with respect to their complexity. This is also relected in their performance. In an experimental setting with three layers of QBs and ive clouds with ive FEs in each cloud, with the packet loss varied from 0% to 8%, and the minimum consecutive packet loss varied from 0 to 3 and the maximum consecutive packet loss varied from 0 to 5, the following observations were made. The looded mode change algorithm reached all FEs in approximately 45 ms, while the hierarchical algorithm required 1200–3200 ms, depending on the link loss and burstiness setting. This was observed when the link latency was set to a very conservative 8 ms for each of the links in order to emulate a wide-area power grid (the speed of light in iber or copper is roughly 125 mi/ms); this includes the links between the QBs as well as the links between the FEs. More detailed information about the performance of the mode change algorithms can be found in [61]. 4.7.5 SYSTEMATIC ADAPTATION VIA DATA LOAD SHEDDING WITH MODES The mode change algorithms described above, coupled with other information that can easily be captured by middleware and an application management system, GridStat 103 provide powerful capabilities in helping a WAMS-DD not only to adapt to IT problems, but also to optimize its beneit for the power problems being dealt with at any given time, not just the (usually calm) steady state. Electric utilities perform load shedding, whereby they cut off power to some customers as a last resort to try to avoid a blackout. We call this power load shedding. There is a communications analog for WAMS-DD: data load shedding. For example, GridStat’s QoS+ APIs are for rate, latency, and number of paths. However, they require not only the desired amount (which would be provided in the steady state) but also the worst case that the application can tolerate. This implicitly gives permission to throttle back some lows from their desired QoS+ toward the worst case if necessary, by reducing the rate or the number of paths, or both. This can help the WAMS-DD adapt to fewer functioning resources being available due to benign IT failures or malicious cyberattacks. But it can also help free up capacity in order to add new subscriptions speciic to a given contingency. For example, dozens of new synchrophasor sensors could be subscribed to, the rates of some existing subscriptions could be increased, and a few feeds from high-rate DFRs could be added (at 720 Hz or more). This could help operators have a much better understanding of what is happening in the grid at that moment. This could be done very rapidly: in a fraction of a second, not only are the data feeds changed, but also visualization windows pop up to alert operators and give them quick insight into the ongoing power problems. Given GridStat’s semantics, such data load shedding can easily be done in a systematic way that dynamically optimizes (or at least greatly improves) the value of the WAMS-DD to the utility. A very simple scheme would be one in which different subscriber applications are tagged with their importance (think of a number from 0 to 10, for example), not only for the steady state but for that application’s importance in different contingencies. These data can then be used to decide which subscriptions are throttled back toward their worst case in any of a number of ways. This adaptation could be further optimized using more complex beneit functions.* With beneit functions, an application (or system integrator) would specify how much beneit a given level of QoS+ provided to the application; again, for simplicity, think of a number from 0 to 10 here. Examples of a beneit function speciication, where individual list items are (rate:beneit), might be {(<30:0), (30:3), (60:7), (≥120:10)}. This means that there is no beneit to the application from a rate less than 30 Hz and no added beneit from going above 120 Hz, and the values between have the beneits listed. 4.7.6 REMOTE PROCEDURE CALL GridStat’s one-way delivery mechanisms are suficient for delivering updated values from remote sensors. However, they are inadequate for a round-trip RPC, for example, to an actuator in the power grid or between control centers. To overcome * Beneit functions are more commonly called utility functions, but we use the former term in the electricity sector to avoid confusion. 104 Smart Grids this limitation of the one-way delivery mechanism, the Ratatoskr round-trip RPC mechanism is built over GridStat [63]. It is a highly tunable RPC mechanism tailored to the needs of the power grid. The RPC mechanism implemented is a simple invocation of remote subroutines without any object-oriented semantics, which provides a familiar programming environment and lends itself well to serial programming. More importantly, the mechanism proves the feasibility of a two-way control system on top of the GridStat framework. 4.7.6.1 Mechanism Details Ratatoskr is a two-level protocol stack that is built on top of the GridStat network (see Figure 4.11). The bottom level is the 2 Way over Pub-Sub (2WoPS) protocol that provides for two-way communication, while the top level provides an RPC mechanism (Ratatoskr RPC [RRPC]). The 2WoPS protocol utilizes the QoS semantics provided by the GridStat network (spatial redundancy), and it introduces two additional redundancy techniques: temporally redundant sends and ACK/retries. RRPC provides extensive customization of the redundancy levels for each call, which provides for various trade-offs. Further, pre- and postconditions built into the call semantics provide additional safety mechanisms for application designers. These are described after we irst explain the details of the RPC protocol. 4.7.6.2 The 2WoPS Protocol To achieve a two-way communication style, needed for RPC signaling, Ratatoskr uses two separate GridStat pub-sub connections. Each RPC host utilizes both a publisher and a subscriber—the publisher for sending data as published status variables and the subscriber for receiving data. Currently, the 2WoPS protocol supports only communication between two hosts, but the GridStat framework has support for multicast of status events, so future versions could add one-to-many and group communication styles. Control center Substation Control center system Actuator Ratatoskr RPC Legacy control 2WoPS GridStat publisher Sensor state monitor GridStat subscriber Substation sensors Ratatoskr RPC 2WoPS GridStat publisher GridStat data plane Network: UDP-IP, ATM, network processors, . . . FIGURE 4.11 RPC module stack. Legacy control Protection scheme GridStat subscriber GridStat 105 The protocol aims to provide a high degree of fault tolerance while retaining predictable delivery times. To this end, three redundancy techniques are employed to strengthen delivery guarantees. These are: 1. Spatial redundancy: GridStat provides QoS that a subscription stream should be forwarded through multiple disjoint network paths. The 2WoPS protocol uses this feature to provide spatial redundancy for its communication. 2. Temporal redundancy: While spatial redundancy provides protection from a range of network failures, the network costs and delay properties of a large number of paths could prove prohibitive in cases where a very high degree of fault tolerance is required or where the network topology is unsuited for spatial redundancy. Ratatoskr offers temporal redundancy by allowing each data unit sent by the 2WoPS protocol to be repeatedly published without waiting for a round trip. A data unit published n times has a temporal redundancy level of n and tolerates n−1 losses. 3. ACK/resend: While spatial and temporal redundancy provide a high degree of fault tolerance, they do not provide any feedback to the application about successful delivery, and do not provide protection from failures that affect the whole network, such as periods of heavy congestion. Ratatoskr provides a third technique for redundancy by having message receivers ACK successfully received 2WoPS messages, and allowing the user to specify that messages that do not produce an ACK should be resent up to n times. A message that is set to be resent up to n times when ACKs are missing has an ACK/retry level of n. 4.7.6.3 The Ratatoskr RPC Mechanism RRPC is an RPC protocol for power grid control communication built on top of the 2WoPS protocol. It draws extensively on the features provided by the 2WoPS protocol. Delivery guarantees for calls are achieved using GridStat’s QoS-enabled network communications. In addition to increased safety through fault tolerance, pre- and postconditions on calls are built into the RPC semantics. Preconditions are conditional expressions over GridStat status variables that are evaluated at the server (a power grid actuator) before execution of an RRPC call, and, if the expression is false, the call to the physical actuator is not executed. Postconditions are similar expressions evaluated after the call has executed, and negative results are reported back to the application via exception. This can provide a physical conirmation that the actuator succeeded even if all of the redundant reply messages are lost. This helps circumvent a wellknown limitation in distributed computing—that in an asynchronous system, even with only one failure, a client cannot know whether its request was processed by the server or not [64]—by using a physical feedback loop. It also handles the case of a failed actuator, when the irmware returns a successful return code but the physical device is stuck open or closed. An example of precondition in power grid operation could be the control of an isolator. Isolators are actuators that connect and disconnect de-energized power circuits. 106 Smart Grids A precondition could be to verify that a line is de-energized before attempting isolation. An example of a postcondition could be the control of a load tap changer. Load tap changers are components of certain transformers that allow adjustment of output voltage. A postcondition could be to verify the new voltage level after load tap changer operation, or even to generate a status report from all connected devices and send it back to the client. 4.7.7 SECURITY AND TRUST IN GRIDSTAT As discussed in Section 4.6, security and trust are important for the operation of any information infrastructure supporting the power grid, and there are multiple dimensions upon which they must be addressed. GridStat research has led to implementation within the framework of mechanisms for securing and authenticating published messages [65] as well as mutual authentication between components of the GridStat system [66]. Both mechanisms are built using an architecture for composable and replaceable security modules under the control of security managers operating in cooperation with QBs. This architecture was motivated by two principal facts. First, the long lifetime of power grid IT infrastructure strongly suggests that security implementations developed today will not remain secure over the decades that the system will be deployed. Second, no single existing protocol, particularly for message authentication, simultaneously provides both adequate security and adequate latency to meet the needs of all potential applications for the GridStat system: it is necessary to make trade-offs between the level of security provided and the performance, and these trade-offs will be different for different applications and data streams [67]. The security module architecture addresses both needs. A number of different encryption, authentication, and obfuscation modules have been built for GridStat. It is relatively easy to build modules to explore new cryptographic protocols having different performance trade-offs. For example, recent work on protocols based on one-time public-key signature protocols has been incorporated as a new module choice offering a favorable latency/ security trade-off for some situations [50,68]. The module replacement mechanism is designed to allow the modules in use for a particular stream or node authentication purpose to be securely replaced during operation, even if the algorithm used by the replaced module has been compromised. The scheme for this relies on sharing a collection of symmetric-key material between a security manager node and the nodes that it manages at the time that the nodes are commissioned. This preshared key material is slowly consumed over the life of the system when module replacements occur. Another aspect concerns securing the nodes of the security management system themselves. Compromising the security management node associated with a particular QB potentially compromises communications of the management subtree that the QB anchors. To address this, a distributed security service has been designed using threshold cryptographic techniques to ensure that compromising even several nodes does not compromise operation of the system [69]. As yet, the GridStat framework does not contain any speciic support for trust decision making, though the work in [55] is designed to support GridStat or any GridStat 107 other critical-infrastructure pub-sub system. A utility operator’s decisions rely on data received and inferred. In an adversarial world, it must be possible for operators to assess how likely it is that the data they are provided truly relect the state of the system or whether they have been manipulated or corrupted. Security mechanisms such as those described above are extremely important for reducing the likelihood of tampering, but decisions should not be taken on the blind assumption that the security mechanisms are completely effective in all cases (e.g., data could be manipulated by tampering with sensors, ahead of the point where a digital signature is applied). Research undertaken so far addresses what such mechanisms should look like [70] and how they can be used to support utilities’ operational decision making [55,71]. 4.8 A NEW WORLD FOR POWER APPLICATION PROGRAMMERS AND RESEARCHERS The capabilities provided by GridStat, or any WAMS-DD that meets the DRs and IGs and implements security appropriate for WAMs-DD as outlined above, are a quantum leap in data delivery for electric power grids and other critical infrastructures. However, they also open up new ways of thinking for power researchers and engineers in terms of their power applications. We summarize these in this section. As a baseline, power applications can be written assuming that the WAMS-DD will be: • Extremely fast: less than 1 ms or so added over the speed of light in the underlying network layers • Extremely available: and providing the performance guarantees even in the face of failures • Appropriately cybersecure: security mechanisms appropriate for longlived embedded devices and as lightweight as possible so as not to preclude closed-loop applications • Tightly managed: on a per-subscriber-per-publication basis to provide the above strong guarantees • Adaptive: can change its subscription base very quickly In order to exploit this much-improved baseline, power researchers and engineers need to ask themselves a number of questions, including: • What {rate, latency, redundancy} does my application really need for its remote data? • Why fundamentally does it need this level of QoS+? • How sensitive is my application to occasional longer delays, period drops (perhaps a few in a row), or data blackouts for longer periods of time? • How can I formulate and test hypotheses for the above? 108 Smart Grids Further, note that the above is only for the steady state. However, the questions should also be asked for different power contingencies. For example: • How important is my application in a given contingency (e.g., simply assign a number 0–10)? • How much lower {rate, latency, redundancy} can I live with in a contingency (here, even at a lower rate, note that these are still strong guarantees)? • Can I program my application to run at different rates (assuming, if necessary, that antialiasing is done appropriately for that rate), or is there a fundamental reason why my application has to run at a given rate? • What extra data feeds (or existing ones at higher rates) could I use in a given contingency, assuming that they could be available within a second of the contingency being identiied? How can I best use these for either operator situational awareness or closed-loop responses? From the above, it may not be clear that power researchers need to work closely with computer scientists in order to help the power grid be more resilient. The status quo today is that power researchers tend to ignore communications issues and computer scientists do not have nearly enough useful information about the power grid’s applications, as well as its pragmatic IT constraints, in order to offer significant help. However, if they work together, much can be achieved. For example, power researchers can assume much “better” communications (strong guarantees, systematic adaptation, etc.) while computer scientists can come up with even better WAMS-DD mechanisms if they know not only emerging “killer apps” for power but also the trade-offs that power applications can live with. Finally, note that the above only involves communications. The computational resources available to make power grids more stable should not be assumed to be static, in particular thanks to cloud computing [72]. Indeed, the GridCloud project is combining GridStat researchers with Cornell researchers who are developing advanced cloud computing mechanisms to provide a comprehensive, E2E solution for mission-critical cloud computing. Such a system will keep the same high throughput, as generic, commercial cloud computing mechanisms improve consistency and predictability. The consistency is improved by removing the deliberate (and, as it turns out, unnecessary) trade-off to allow occasional inconsistencies in cloud data to preserve high throughput. In terms of performance, GridCloud is also implementing ways to make the aggregate performance and ramp-up time of cloud computations much better. Given such an improved computational infrastructure, then, questions for power researchers include how they could use • Hundreds of processors in steady state • Thousands or tens of thousands of processors when the grid is approaching a blackout • Data from all participants in a grid enabled quickly when a blackout is being approached Such improved computational support will likely have a huge impact on power grid operations. GridStat 109 The questions in this section represent a completely new way of thinking for power researchers. However, they have the potential to make the power grid much more resilient than simply using today’s applications (or ones planned for tomorrow) and assuming that communications and computations are limited and nonadaptive. 4.9 CONCLUSIONS This chapter has described the requirements for building WAMS-DD, and IGs that the authors believe are necessary for a proper WAMS-DD, or at least highly advisable. It then analyzed how completely these requirements and guidelines are met by network-level technologies, middleware-level technologies, and technologies developed in the electricity sector. Following this, the chapter overviewed the NASPInet initiative, and explained how NASPInet and other WAMS-DDs must employ middleware (as well as, of course, a network-level mechanism with middleware layers on top). Finally, we described GridStat in detail, and overviewed the completely new way of thinking of new power application programs that it enables. ACKNOWLEDGMENTS The GridStat project has been going on since 1999, so there are many people to acknowledge for their support and guidance besides the authors, and we apologize in advance for any omissions. GridStat co-principal investigator Professor Carl Hauser has been an invaluable partner in the development of GridStat. Professor Anjan Bose has been an invaluable collaborator from the power grid side. Dr. Thoshitha Gamage has contributed heavily to GridStat in the last few years. Graduate students Erlend Viddal, Stian Abelsen, Venkata Irava, Erik Solum, Ryan Johnston, Joel Helkey, Kim Swenson, Sunil Muthuswamy, Supreeth Sheshadri, Ping Jiang, Rasika Chakravarthy, Michael Carter, and Punit Agrawal have all contributed heavily. Power researchers who have helped inluence GridStat’s applications and requirements include graduate student Tao Yang and power researchers Dr. Chuanlin Zhao and Professor Mani Venkatasubramanian. Lead GridStat staff programmer David Anderson has contributed immensely to GridStat’s implementation as a full-time programmer for the last 4 years. Undergraduate staff programmers include Andrew Lytle, Raeann Marks, Loren Hoffman, Nathan Schubkegel, Alex Schuldberg, Ryan Hoffman, Kyle Wilcox, Travis Gomez, Chad Selph, Constance Bell, Simon Ngyuen, Nick Newsome, and DeAndre Blackwell. Suggestions, questions, and advice from a large number of people have helped inluence GridStat. These are too numerous to list, but include Jeff Dagle, Greg Zweigle, Phil Overholt, Patricia Hoffman, Dave Bertagnolli, Rich Stephenson, Jeff Gooding, Mahendra Patel, Ritchie Carroll, Ken Martin, Floyd Galvan, Lisa Beard, Neeraj Suri, Deb Frincke, Jim McNierney, Dan Brancaccio, Dick Wilson, Dan Lutter, and Don Kopczynski. Many funding agencies have provided inancial support for GridStat. These include US Department of Energy grant DE-OE0000097 (TCIPG), US Department of Energy grant DE-OE0000032 (GridSim), ARPA-E grant DE-AR0000230 (GridCloud), grants CNS 05-24695 (CT-CS: Trustworthy Cyber Infrastructure for the Power Grid (TCIP)) and CCR-0326006 from the US National Science Foundation; 110 Smart Grids and by grant #60NANB1D0016 (Critical Infrastructure Protection Program) from the National Institute of Standards and Technology, in a subcontract to Schweitzer Engineering Labs Inc. We thank the US Department of Energy and the Department of Homeland Security for their part in funding the National Science Foundation TCIP center. We thank Schweitzer Engineering Labs for their donation of equipment and Paciic Northwest National Laboratory for their loan of equipment. DISCLAIMER This chapter was prepared as an account of work sponsored by an agency of the US government. Neither the US government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any speciic commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the US government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or relect those of the US government or any agency thereof. REFERENCES 1. D. E. Bakken, R. E. Schantz, and R. D. Tucker, Smart grid communications: QoS stovepipes or QoS interoperability, in Proceedings of the 2009 Grid-Interop, GridWise Architecture Council, Denver, CO, 2009. Available at: http://www.gridstat.net/ publications/TR-GS-013.pdf. 2. K. Fodero, C. Huntley, D. E. Whitehead, and B. Kasztenny, A novel scheme for widearea time synchronization, in Proceedings of the 10th IET International Conference on Developments in Power System Protection, Managing the Change, pp. 1–5, 29 March–1 April, Manchester, 2010. 3. D. E. Bakken, C. H. Hauser, H. Gjermundrød, and A. Bose, Towards more lexible and robust data delivery for monitoring and control of the electric power grid, Washington State University, Pullman, WA, Technical Report EECS-GS-009, May 2007. 4. Electric Power Research Institute (EPRI), The integrated energy and communication systems architecture, vol. IV, Technical Analysis, 2004. 5. S. Horowitz, D. Novosel, V. Madani, and M. Adamiak, System-wide protection, IEEE Power and Energy Magazine, 6(5), 34–42, 2008. 6. E. Rosen, A. Vishanathan, and R. Callon, RFC-3031: Multiprotocol label switching architecture, The Internet Society, 2001. 7. Y. Krishnamurthy, V. Kachroo, D. A. Karr, C. Rodrigues, J. P. Loyall, R. E. Schantz, and D. C. Schmidt, Integration of QoS-enabled distributed object computing middleware for developing next-generation distributed applications, in Proceedings of the 2001 ACM SIGPLAN Workshop on Optimization of Middleware and Distributed Systems, vol. 36, no. 8, pp. 230–237, 18 June, Snowbird, UT, 2001. 8. R. E. Schantz and D. C. Schmidt, Research advances in middleware for distributed systems: State of the art, in Proceedings of the 2002 IFIP World Computer Congress, 25–30 August, Montreal, QC, 2002. GridStat 111 9. J. Saltzer, D. Reed, and D. Clark, End-to-end arguments in system design, Transactions on Computer Systems, Association of Computing Machinery, 2(4), 277–288, 1984. 10. G. V. Chockler, I. Keidar, and R. Vitenberg, Group communication speciications: A comprehensive study, ACM Computing Surveys, 33(4), 1–43, 2001. 11. X. Défago, A. Schiper, and P. Urbán, Totally ordered broadcast and multicast algorithms: Taxonomy and survey, ACM Computing Surveys, 36(4), 372–421, 2004. 12. Y. Hu, Data bus technical speciications for North American synchro-phasor initiative network (NASPInet), US Department of Energy, May 2009. https://www.naspi.org/File. aspx?ileID=587. 13. R. E. Schantz, J. P. Loyall, M. Atighetch, and P. P. Pal, Packaging quality of service control behaviors for reuse, in Proceedings of the 2002 5th IEEE International Symposium on Object-Oriented Real-Time Distributed Computing, 29 April–1 May, Washington, DC, 2002. 14. P. P. Pal, J. P. Loyall, R. E. Schantz, J. A. Zinky, R. Shapiro, and J. Megquier, Using QDL to specify QoS aware distributed (QuO) application coniguration, in Proceedings of the 2000 3rd IEEE International Symposium on Object-Oriented Real-Time Distributed Computing, pp. 310–319, 15–17 March, Newport, CA, 2000. 15. D. E. Bakken, Quality of service design considerations for NASPInet, Presentation to the North American Synchrophasor Initiative (NASPI) Work Group Meeting, 4 February, 2009. Available at: http://www.naspi.org. 16. K. H. Gjermundrød, I. Dionysiou, C. H. Hauser, D. E. Bakken, and A. Bose, Flexible and robust status dissemination middleware for the electronic power grid, School of Electrical Engineering and Computer Science, Washington State University, Pullman, WA, Technical Report EECS-GS-003, September 2003. 17. R. Bobba, E. Heine, H. Khurana, and T. Yardley, Exploring a tiered architecture for NASPInet, in Proceedings of the 2010 IEEE Conference on Innovative Smart Grid Technologies, pp. 1–8, 19–21 January, Gaithersburg, MD, 2010. 18. K. Tomsovic, D. E. Bakken, M. Venkatasubramanian, and A. Bose, Designing the next generation of real-time control, communication and computations for large power systems, Proceedings of the IEEE Special Issue on Energy Infrastructure Systems, 93(5), 965–979, 2005. 19. K. H. Gjermundrød, D. E. Bakken, C. H. Hauser, and A. Bose, GridStat: A lexible QoS-managed data dissemination framework for the power grid, IEEE Transactions on Power Delivery, 4(1), 136–143, 2009. 20. V. S. Irava and C. H. Hauser, Survivable low-cost low-delay multicast trees, in Proceedings of the IEEE Global Telecommunications Conference, vol. 1, pp. 6–115, 2 December, St Louis, MO, 2005. 21. V. S. Irava, Low-cost delay-constrained multicast routing heuristics and their evaluation, PhD Dissertation, Washington State University, Pullman, WA, August 2006. 22. GridWise Architecture Council, Interoperability Constitution Whitepaper (v1.1), December 2006. 23. L. Roberts, A radical new router, Proceedings of the IEEE Spectrum, 46(7), 35–39, 2009. 24. P. T. Eugster, P. A. Felber, R. Guerraoui, and A. M. Kermarrec, The many faces of publish-subscribe, ACM Computing Surveys, 35(2), 114–131, 2003. 25. J. Zinky, D. E. Bakken, and R. Schantz, Architectural support for quality of service for CORBA objects, Theory and Practice of Object Systems, 3(1), 55–73, 1997. 26. K. P. Birman, J. Chen, K. M. Hopkinson, R. J. Thomas, J. S. Thorp, R. van Renesse, and W. Vogels, Overcoming communications challenges in software for monitoring and controlling power systems, Proceedings of the IEEE, 93(5), 1028–1041, 2005. 112 Smart Grids 27. K. Hopkinson, G. Roberts, X. Wang, and J. Thorp, Quality of service considerations in utility communication networks, IEEE Transactions on Power Delivery, 24(3), 1465– 1474, 2009. 28. J. Rajahalme, A. Conta, B. Carpenter, and S. Deering, RFC3697: IPv6 low label speciication, The Internet Society, 2004. Available at: http://www.faqs.org/rfcs/rfc3697.html. 29. Y. Vigfusson, H. Abu-Libdeh, M. Balakrishnan, K. Birman, R. Burgess, H. Li, G. Chockler, and Y. Tock, Dr. Multicast: Rx for data center communication scalability, in Proceedings of the 5th European Conference on Computer Systems Eurosys, pp. 349–362, April, ACM SIGOPS, Paris, France, 2010. 30. D. Basin, K. Birman, I. Keidar, and Y. Vigfusson, Sources of instability in data center multicast, in Proceedings of the 4th International Workshop on Large Scale Distributed Systems and Middleware (LADIS 2010), August, ACM Principles of Distributed Computing (PODC), 25–28 July, Zurich, 2010. 31. K. Birman, Guide to Reliable Distributed Systems: Building High-Assurance Applications and Cloud-Hosted Services, Springer, London, 2012. 32. A. Farrel (ed.), Network Quality of Service Know It All, Elsevier, Boston, 2009. 33. R. Wojcik and A. Jajszczyk, Flow oriented approaches to QoS assurance, ACM Computing Surveys, 44(1), 1–37, 2012. 34. A. Sydney, J. Nutaro, C. Scoglio, D. Gruenbacher, and N. Schulz, Simulative comparison of multiprotocol label switching and openFlow network technologies for transmission operations, IEEE Transactions on Smart Grid, 4(2), 763–770, 2013. 35. PGM Reliable Transport Protocol Speciication, RFC 3208, IETF, 2001. 36. Y. Amir, C. Danilov, and J. Stanton, A low latency, loss tolerant architecture and protocol for wide area group communication, in Proceedings of the 2000 IEEE First International Conference on Dependable Systems and Networking, pp. 327–336, 25–28 June, New York, 2000. 37. N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson, J. Rexford, S. Shenker, and J. Turner, OpenFlow: Enabling innovation in campus networks, ACM SIGCOMM Computer Communication Review, 38(2), 69–74, 2007. 38. Open Network Foundation, OpenFlow Switch Speciication 1.3.2, April 25, 2013. 39. S. McGillicuddy, Not all OpenFlow hardware is created equal: Understanding the options, Open Network Foundation, 25 September 2013. Available at: http://www. opennetworking.org. 40. Net Insight, The Nimbra platform, 2013. http://www.netinsight.net/Global/Documents/ Products/Brochures/Nimbra_Brochure.pdf. 41. K. Geihs, Middleware challenges ahead, IEEE Computer, 34(6), 24–31, 2001. 42. D. Bakken, Middleware, unpublished article, 2001. http://www.eecs.wsu.edu/~bakken/ middleware.pdf. 43. Y.-J. Kim, J. Lee, G. Atkinson, H. Kim, and M. Thottan. SeDAX: A scalable, resilient, and secure platform for smart grid communications, IEEE Journal on Selected Areas in Communications, 30(6), 1119–1136, 2012. 44. K. Birman, Like it or not, web services are distributed objects! Communications of the ACM, 47(12), 60–62, 2004. 45. K. Birman, The untrustworthy web services revolution, IEEE Computer, 39(2), 98–100, 2006. 46. PJM, PJM 2007 Strategic Report, 2 April, 2007. http://www.pjm.com/~/media/Images/ ctc-display/modules/timeline/2007-10-yr-pdf.ashx. 47. W. Mahnke, S. Leitner, and M. Damm, OPC Uniied Architecture, Springer, Berlin, 2009. 48. J. Zhang and C. A. Gunter, Application-aware secure multicast for power grid communications, in Proceedings of the IEEE International Conference on Smart Grid Communications (SmartGridComm), pp. 339–344, 4–6 October, Gaithersburg, MD, 2010. GridStat 113 49. C. Hauser, T. Manivannan, and D. Bakken, Evaluating multicast message authentication protocols for use in wide area power grid data delivery services, in Proceedings of the 45th Hawaii International Conference on System Sciences (HICSS), pp. 2151–2158, 4–7 January, IEEE, Maui, Hawaii, 2012. 50. K. Cairns, C. Hauser, and T. Gamage, Flexible data authentication evaluated for the smart grid, in Proceedings of the 4th Annual IEEE SmartGrid Communications Conference, 21–24 October, Vancouver, BC, IEEE. Forthcoming. 51. D. Clark, A cloudy crystal ball—Visions of the future, in Proceedings of the TwentyFourth Meeting of the Internet Engineering Task Force (IETF), pp. 539–543, July, Cambridge, MA, 1992. 52. Z. Xie, M. Govindarasu, V. Vittal, A. Phadke, and V. Centeno, An information architecture for future power systems and its reliability analysis, IEEE Transactions on Power Systems, 17(3), 857–863, 2002. 53. C. H. Hauser, D. E. Bakken, I. Dionysiou, K. H. Gjermundrød, V. Irava, J. Helkey, and A. Bose, Security, trust and QoS in next-generation control and communication for large power systems, International Journal of Critical Infrastructures, 4(1–2), 3–16, 2008. 54. I. Dionysiou, D. Frincke, C. Hauser, and D. Bakken, An approach to trust management challenges for critical infrastructures, in Proceedings of the 2nd International Workshop on Critical Information Infrastructures Security (CRITIS07), Lecture Notes in Computer Science Series volume 5141, pp. 173–184, 2–5 October, Springer Berlin, Malaga, Spain, 2007. 55. Y. Wang and C. Hauser, An evidence-based trust assessment framework for critical infrastructure decision making, in Proceedings of the 5th IFIP Working Group 11.10 International Conference on Critical Infrastructure Protection (ICCIP), pp. 125–135, 23–25 March, Hanover, NH, 2011. 56. M. D. Hadley, J. B. McBride, T. W. Edgar, L. R. O’Neil, and J. D. Johnson, Securing wide area measurement systems, Paciic Northwest National Laboratory, Technical Report PNNL-17116, 2007. 57. D. E. Bakken, A. Bose, C. H. Hauser, D. E. Whitehead, and G. C. Zweigle, Smart generation and transmission with coherent, real-time data, Proceedings of the IEEE, 99(6), 928–951, 2011. 58. K. Swenson, Exploiting network processors for low latency, high throughput, rate-based sensor updated delivery, MS Thesis, Washington State University, Pullman, WA, 2009. 59. H. Gjermundrod, C. Hauser, and D. Bakken, Scalable wide-area multicast with temporal rate iltering distribution framework, in Proceedings of the 2011 IEEE 11th International Conference on Computer and Information Technology (CIT), pp.1–8, 31 August–2 September, Washington, DC, 2011. 60. H. Gjermundrød, D. E. Bakken, and C. H. Hauser, Integrating an event pattern mechanism in a status dissemination middleware, in Proceedings of 1st International Conference on Pervasive Patterns and Applications (PATTERN’09), pp. 259–264, 15–20 November, IEEE, Athens, Greece, 2009. 61. S. F. Abelsen, H. Gjermundrød, D. E. Bakken, and C. H. Hauser, Adaptive data stream mechanism for control and monitoring applications, in Proceedings of 1st International Conference on Adaptive and Self-Adaptive Systems and Applications (ADAPTIVE’09), pp. 86–91, 15–20 November, IEEE, Athens, Greece, 2009. 62. K. H. Gjermundrød, Flexible QoS-managed status dissemination middleware framework for the electric power grid, PhD Dissertation, Washington State University, Pullman, WA, August, 2006. 63. E. S. Viddal, D. E. Bakken, H. Gjermundrød, and C. H. Hauser, Wide-area actuator RPC over GridStat with timeliness, redundancy, and safety, in Proceedings of the International Conference on Complex, Intelligent and Software Intensive Systems (CISIS’2010), pp. 17–24, 15–18 February, Krakow, Poland, 2010. 114 Smart Grids 64. M. J. Fischer, N. A. Lynch, and M. S. Patterson, Impossibility of distributed consensus with one faulty process, Journal of the ACM, 32, 374–382, 1985. 65. E. Solum, C. H. Hauser, and R. Chakravarthy, Modular over-the-wire conigurable security for long-lived critical infrastructure monitoring systems, in A. Gokhale and D. C. Schmidt (eds), Proceedings of the 3rd ACM International Conference on Distributed Event-Based Systems, pp. 1–9, ACM, New York, 2009. 66. R. Chakravarthy, C. H. Hauser, and D. E. Bakken, Long-lived authentication protocols for critical infrastructure process control systems, in Proceedings of the Fourth IFIP WG 11.10 International Conference on Critical Infrastructure Protection, March, 2010. 67. C. H. Hauser, T. Manivannan, and D. E. Bakken, Evaluating multicast message authentication protocols for use in wide area power grid data delivery services, in Proceedings of the 45th Hawaii International Conference on System Science (HICSS), pp. 2151–2158, 4–7 January, Maui, HI, 2012. 68. K. Cairns, T. Gamage, and C. Hauser, Eficient targeted key subset retrieval in fractal hash sequences, in Proceedings of the 20th ACM Conference on Computer and Communications Security, pp. 1273–1284, 4–8 November, Berlin, Germany, 2013. 69. P. Agrawal, A fault-tolerant key storage service with proactive recovery. MS Thesis, School of Electrical Engineering and Computer Science, Washington State University, August, 2012. Available at: http://www.dissertations.wsu.edu/Thesis/Summer2012/p_ agrawal_091412.pdf. 70. I. Dionysiou, Dynamic and composable trust for indirect interactions, PhD Dissertation, Washington State University, Pullman, WA, August, 2006. 71. C. H. Hauser, Trust research to address uncertainty in security for the smart grid, in Proceedings of the IEEE PES Innovative Smart Grid Technologies (ISGT), pp. 1–2, 16–20 January, Washington, DC, 2012. 72. K. P. Birman, L. Ganesh, and R. van Renesse, Running smart grid control software on cloud computing architectures, in Workshop on Computational Needs for the Next Generation Electric Grid, 19–20 April, Cornell University, Ithaca, NY, 2011. 5 A Distributed Framework for Smart Grid Modeling, Monitoring, and Control Alfredo Vaccaro and Eugenio Zimeo CONTENTS 5.1 5.2 5.3 Introduction .................................................................................................. 115 Related Works ............................................................................................... 118 Web Services–Based Framework for SG Management................................ 120 5.3.1 GISinterfaceWS ................................................................................ 122 5.3.2 DataAcquisitionWS........................................................................... 122 5.3.3 eAssessmentWS ................................................................................ 122 5.3.4 ComputationalWS............................................................................. 123 5.3.5 DataStorageWS ................................................................................. 124 5.3.6 Worklow Enactor ............................................................................. 125 5.4 Experimental Setting .................................................................................... 126 5.4.1 Data Acquisition Web Service .......................................................... 126 5.4.2 Job Submission Web Service ............................................................ 127 5.4.3 Computational Web Service ............................................................. 127 5.4.4 Data Storage Web Service ................................................................ 129 5.4.5 Worklow Enactor ............................................................................. 129 5.5 Conclusions ................................................................................................... 130 References .............................................................................................................. 131 5.1 INTRODUCTION Electrical power systems were traditionally characterized by the presence of numerous utilities, heterogeneous standards, overlapping territories, and a general lack of integration. Recently, under pressure from deregulated electricity markets and new environmental policies, the main limitations intrinsic to this environment have become clear and the structures of modern power systems have evolved to accommodate large changes induced by these new energy policy trends. In this emerging scenario, the large-scale deployment of the smart grid (SG) paradigm could play a strategic role in supporting the evolution of conventional electrical grids toward active, lexible, and self-healing web energy networks composed of distributed and cooperative energy resources. 115 116 Smart Grids From a conceptual point of view, the SG is the convergence of information and operational technologies applied to the electric grid, providing sustainable options to customers and improved security [1]. SG technologies include advanced sensing systems, two-way high-speed communications, monitoring and enterprise analysis software, and related services to get location-speciic and real-time actionable data in order to provide enhanced services for both system operators (i.e., distribution automation, asset management, advanced metering infrastructure) and end users (i.e., demand-side management, demand response) [2,3]. Advances in research on SG could increase the eficiency of modern electrical power systems by (i) supporting large-scale penetration of small-scale distributed energy resources (DERs); (ii) facilitating the integration of renewable energy sources; (iii) reducing system losses and greenhouse gas emissions; and (iv) increasing the reliability of electricity supply to customers. From this perspective, a crucial issue is how to increase the penetration of DERs into the SG by properly coordinating their operation and by restricting their negative impact on grid operations and control. It is well known that the integration of DERs into running distribution systems affects the power lows and voltage conditions of customer and utility equipment by inducing a number of side effects (e.g., bidirectional power lows, increased fault current levels, and a need for new concepts of network protection [4]). This increases the complexity of the control, protection, and maintenance of distribution systems, which are typically designed to operate radially without any generation at the distribution line or customer side [5]. This requires extensive analysis aimed at (i) identiication of the optimal generation schedule that minimizes production costs and balances the demand and supply that come from both micro sources and the distribution feeder; (ii) online assessment of the SG security and reliability levels; and (iii) identiication of proper preventive and corrective measures that mitigate the effects of critical contingencies. The main limitations that should be addressed in order to overcome these issues originate mostly from the following intrinsic features characterizing the SG operation: • The rising level of DERs, which makes the distribution systems more vulnerable with respect to dynamic perturbations • The increasing number of smaller geographically dispersed generators, which could signiicantly increase the number of power transactions • The dificulties arising in predicting and modeling market operators’ behavior, governed mainly by unpredictable economic dynamics, which introduce considerable uncertainty into short-term SG operation • The need for more detailed security studies assessing the impact of multiple contingencies on the distribution system (i.e., N-2 security criteria) • The high penetration of generation units powered by renewable energy sources, which induces considerable uncertainty in SG operation • The low level of upgrade and interoperability of existing distribution management systems that are typically based on low-scalable and proprietary A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 117 software platforms characterized by different information technologies and standards for the data exchange • The need for standard interfaces between the distributed tools and components characterizing the power system’s operation environment; these interfaces must be standardized so as to work in a “plug-and-play” concept similar to the hardware used in substation automation In this context the adoption of web services, a new paradigm for enterprise application integration built on the foundation of open standard and common infrastructure could contribute to overcoming some of these limiting aspects. A web service is an interface that describes a collection of modular, self-contained applications that are network-accessible through standardized extensible markup language (XML) messaging. The interface hides the implementation details of the service, allowing it to be used independently of the hardware or software platform on which it is implemented and also independently of the programming language in which it is written [6,7]. The use of standard XML protocols makes web services technology platform-, language-, and vendor-independent, and they are thus an ideal candidate for supporting the deployment of the SG paradigm [8]. In agreement with these arguments, this chapter outlines the design of a prototype web services–based framework for integrated SG modeling, monitoring, and control. The architecture of the proposed framework is based on the following web services: 1. GISinterfaceWS acquires georeferential information on the current network topology and the control devices asset by interacting with a geographical information system (GIS). 2. DataAcquisitionWS acquires the network ield data interacting with the entire set of supervisory control and data acquisition (SCADA) systems and ield electrical measurements (FEM) by standardized XML messaging. 3. eAssessmentWS develops the distributed generation/storage unit capability assessment by interacting with a network of intelligent units. 4. ComputationalWS employs a parallel solution engine, based on a computational grid, to develop the calculations needed to support the entire set of SG services in terms of state estimation, network analysis, online security assessment, and so on. 5. DataStorageWS is used to permanently store historical data and alarm conditions that can be analyzed off-line using a web browser. The service can be remotely accessed by using data access middleware. In order to prove the effectiveness of the proposed framework, the results of ield activities developed on an experimental test bed are presented and discussed. The experimental results obtained show that the web service technologies in the SG are an effective tool to integrate monolithic and hard-to-customize power system control and monitoring software tools, allowing system operators to easily manage proprietary hardware instrumentation and multiple software modules. 118 5.2 Smart Grids RELATED WORKS The control and operational issues of SGs are presented and discussed in [9,10]. The analysis in this chapter shows that the required control and operational strategies of an SG can be observably and conceptually different from those of conventional power systems. These differences will depend on the type and depth of penetration of DER units, load characteristics, power quality constraints, and market participation strategies. Consequently, novel control strategies and ad hoc energy management systems should be deployed for this new application. These facilities would support such applications as DER and demand response dispatch, distribution automation and substation automation, adaptive relaying, energy management, market pricing, grid modeling, operator displays, and advanced visualization systems [9]. SGs could play an important role in deregulated energy markets, as outlined in [11]. In this context, SGs could be operated in order to serve the total load demand, using local production as much as possible, without exporting power to the upstream transmission grid. This strategy represents a beneit for the transmission system operator, since the SG relieves possible network congestion during peak demand by partly or fully supplying the system’s energy needs. As outlined in [12], operational control and energy management systems for the SG should be implemented through the cooperation of various controllers, located at all these levels, on the basis of the communication and collection of information about distributed energy systems and control commands. This could be deployed according to a centralized or decentralized control paradigm. In a centralized control paradigm, the SG controllers optimize the power exchanged between the distributed grid system (DGS) and the main grid, thus maximizing local production depending on market prices and security constraints. This is achieved by issuing control set points to DERs and controllable loads in order to optimize local energy production and power exchanges with the main distribution grid [12]. The decentralized control paradigm is intended to provide maximum autonomy for the DER units and loads. This implies that local controllers are intelligent and can communicate with each other to form a larger intelligent entity. In a decentralized control framework, the main task of each controller is not to maximize the revenue of the corresponding unit but to improve the overall performance of the SG. Thus, the architecture must be able to include economic functions, environmental factors, and technical requirements, for example, black start. These features indicate that a multiagent system (MAS) can be a prime candidate for decentralized SG control [13–15]. Armed with such a vision, [14,15] present a general framework for the control of DERs organized in SGs. The proposed architecture is based on the agent technology and aims at integrating several functionalities, as well as being adaptable to the complexity and size of the SG. Another important issue to address for effective SG operational control is dealing with the problem of security assessment in both steady-state and dynamic scenarios [10]. The main steady-state security concerns for the operation of SGs are satisfying voltage constraints and maintaining power lows within thermal limits. Dynamic A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 119 security concerns the SG operation under a number of contingencies both within and above it (i.e., during a transition between the grid-connected and islanded modes of operation). Effective SG security assessment requires detailed analysis of phenomena that can compromise SG operation. In this connection, it is expected that online time-domain-based analysis such as voltage stability and transient angular stability can be carried out in real time [16]. All of these functions require the design of reliable, resilient, secure, and manageable standards-based open communication systems [12]. These infrastructures represent a key issue in SG deployment, as they would provide the fundamental backbone for connecting the SG elements, the data providers, and the decision-making entities in an open and interoperable framework. They should support ubiquitous connectivity between decision-making points and dispersed and heterogeneous data sources characterized by varying degrees of transport, security, and reliability requirements [12,17]. They must ensure the capability of SG control systems to transmit data to and from distributed intelligent controllers; moreover, the prospect of the development of innovative energy market policies imposes the need for the realization of eficient and reliable widespread bidirectional communication links between SG operators and customer premises. In the light of these needs, IEC 61850 provides a standardized framework for substation automation integration that speciies the following: the communication requirements, the functional characteristics, the data structure within devices, how applications interact with and control other devices, and how conformity to the standard should be tested [18,19]. In the IEC 61850 standard, part 7-420, the information exchange between DER units and monitoring and control devices [20,21] is considered. The standard also deines some measures for the success of an integrated mobile gateway (MG) communication network design [22]: • Flexibility to adjust and grow the system topology as requirements change • Performance, especially quality of service (QoS), to enable effective prioritization among competing applications and to meet critical requirements of the most important protection and control functions • Reliability, for critical protection systems, but also because so many different systems are relying on the same infrastructure The deployment of a low-cost and high-performance communication infrastructure allowing the cooperation of distributed controllers on the basis of communication and collection of information about distributed energy systems and control commands is a challenge that must be faced by the emerging technologies [12]. The literature analysis discussed above reveals that the present power system operation environment is primarily composed of many distributed tools and components, each addressing only a speciic functionality (i.e., energy management, voltage control, network modeling, security assessment, data acquisition, market interface). They do not deal with the deinition of an integrated platform capable of easily executing very complex applications, built by composing required functionalities in a 120 Smart Grids standardized, easy-to-use, and well-deined way. To address this problem, serviceoriented architectures (SOAs), in particular those based on the standard web service technologies, seem particularly suitable for SG control and monitoring due to their ease of application programming, portability, maintenance, and integration with legacy services [6]. 5.3 WEB SERVICES–BASED FRAMEWORK FOR SG MANAGEMENT The proposed framework for SG control, modelling, and monitoring is based on well-deined interfaces and contracts between services, according to the SOA principles. The interface is deined in a neutral manner that should be independent of the hardware platform, the operating system, and the programming language in which the service is implemented. This allows services built on a variety of such systems to interact with each other in a uniform and universal manner. The main beneits of an SOA system are the improvement of interoperability, the integration of new and legacy applications, the ability to survive evolutionary changes in structure, and the implementation of the internals of each service. Even though many technologies can be adopted to implement an SOA, for example, the common object request broker architecture (CORBA) and message-oriented middleware systems, web services technologies are emerging as particularly suited for ensuring a high degree of integration among existing or newly created services available on the web. Web services, in fact, extend the advantages of software components, making it possible to employ an existing low-level middleware infrastructure based on web servers and hypertext transfer protocol (http). From a programming point of view, a web service is a set of operations that can be easily accessed independently of the service deployment details. The World Wide Web Consortium (W3C) is standardizing many protocols and languages for supporting web services technologies. At low level, simple object access protocol (SOAP) represents the exchange format based on XML for data and speciic data types in a speciic web service call. This protocol is used over the http to issue a call to a remote service whose interface is described through web service description language (WSDL), an XML-based language that is more dynamic and lexible than interface description language (IDL), which characterizes remote method invocation (RMI)-based middleware such as CORBA. Figure 5.1 shows the overall distributed meta-architecture for SG monitoring and control. The core component of the meta-architecture is the SG engine, which is responsible for the execution of SG control, modeling, and monitoring functions in a geographically distributed scenario. It includes high-level components, mainly the submission service, the operation service, and the notiication service. The submission service is responsible for the handling of user submission requests and is designed to simplify the submission phase performed by a nonexpert network operator. The SG operation service is responsible for the execution of SG control and monitoring. Finally, the notiication service is responsible for the asynchronous management of output data of an application able to notify speciic events. A basic and fundamental component of the SG engine is the worklow enactor, which manages the execution of the SG service and adds some functionalities A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 121 Intelligent electronic device 1 Field energy meter n Intelligent electronic device k SCADA 1 Field energy meter 1 SCADA m ..... ..... ..... WAN Utility j Utility 1 GISinterfaceWS DataAcquisitionWS DataStorageWS System operator http/IP interactions Field data acquisition State estimation eAssessmentWS RBDMS Security analysis Data storage Workflow engine Advanced dispatching XML-based database Monitoring Computational engine Smart grid engine RBDMS ComputationalWS Computational grid FIGURE 5.1 Overall web services–based architecture. related to the speciic monitoring application. Here the worklow enactor is traditionally described using its business logic description, and it is written in a certain worklow language. The WSDL interfaces of the three kinds of web services deined above are the following: 1. GISinterfaceWS: for georeferential information acquisition 2. DataAcquisitionWS: for real-time data acquisition 3. eAssessmentWS: for distributed generation/storage unit capability assessment 4. ComputationalWS: to process the mathematical computations required by the SG control and monitoring functions 5. DataStorageWS: related to the data storage functionalities In Figure 5.1, GISInterfaceWS and DataAcquisitionWS are abstracted by the generic DataExchangeWS interface, whereas eAssessmentWS can be bound to the eAssessment module shown in the igure. 122 Smart Grids 5.3.1 gIsInterFaceWs GISinterfaceWS was deined for the acquisition of georeferentiated information on the SG. It delivers the following services: • GetNetworkdata: acquires SG structural data concerning the geographical map, the network topology, and the sources of power supply • GetControlDevicesAsset: type, location, and characteristic parameters of the available control and monitoring devices • GetElectricalNetworkDetails: line parameters, power transformer rating and numbers, impedance values, bus bar scheme, and circuit breaker type and installation • GetOperationalParameters: substation equipment status, feeder breakdowns, failure of distribution transformers, tripping on feeders/lines, consumer outages 5.3.2 dataacquIsItIonWs DataAcquisitionWS was deined for the real-time acquisition of the entire set of measurable SG ield data. This comprises, in particular, the SG topology, the available measurements of the electrical nodes parameters, and the actual state of the generator/storage units computed locally by intelligent electronic devices (IEDs). All possible kinds of measurable parameters are well deined; for each kind of parameter, a code and a basic measurement unit are deined according to IEC 61850-7-420 [20,21]. This standard deines the communication and control interfaces for all SG devices, and develops DER object models. The main beneits of using this standard are: • • • • Consistent data models Easy maintenance of data models Sharing interoperability of communication based on IEC 61850 Seamless integration into the station automation and the power control system Each measurement value returned by the DataAcquisitionWS is correlated to a time stamp, which indicates the time at which the value was acquired. The returned type by the described portType is a string that corresponds to the uniform resource locator (URL) of the ile containing the required information, which can be downloaded at a convenient time. 5.3.3 eassessmentWs A reliable assessment of the actual and future generator/storage unit capabilities appears to be crucial for reliable SG operation. This demands the design of mathematical models able to predict—given the actual unit state and the forecasted environmental conditions—the evolution of the supply A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 123 capabilities [23]. These predictive models should also exhibit adaptive features, to deal with the time-varying phenomena affecting the generation/storage unit performances (i.e., aging, parameter drifts) and low computational requirements in order to ensure effective hardware implementation. To address this problem, the proposed framework integrates a network of distributed IEDs to dynamically assess the supply capability of the main SG components. These IEDs are fully managed by a dedicated web service. This web service delivers the getCapabilitydata portType, which acquires the georeferentiated input data for the capability calculations by interfacing with dedicated web-based public services (forecasted environmental variables), and the submitCapabilitydata portType, which submits the right input data to the distributed IEDs for the capability calculations. 5.3.4 comPutatIonalWs To support the SG’s monitoring and control functions, intensive online network computations are required. These comprise system state estimation, optimal power low studies (i.e., voltage control), security assessment (i.e., contingency analysis), and economic analysis (i.e., the generators must be dispatched based on an economic assessment of fuel cost, electric power cost, weather conditions, and anticipated process operation). The online solution of these tasks is computer-intensive activity, especially in the presence of a large number of DGSs, the extension of electrical networks, and the depth of analysis. So, to address these computationally demanding analyses in useful execution times, parallel algorithms to be executed in a distributed environment have to be adopted [8,24]. In this context, the computational engine has to ensure high reliability, lexibility, and scalability. A practical solution is to exploit the interesting features offered by cloud computing providers. A cloud system is able to provide several services at different abstraction levels: infrastructure as a service (IaaS), platform as a service (PaaS), and software as a service (SaaS). By exploiting these services, customers can satisfy every kind of information and communications technology (ICT) need by simply exploiting a high-bandwidth network connection. In particular, the availability of a large number of resources in cloud systems and sophisticated software infrastructures for handling the available hardware makes it possible to ensure high levels of scalability, reliability, and performance. Therefore, unlike the approaches used in the past for solving similar problems, such as middleware for clusters of computers or geographically dispersed computers networked through the Internet, by using cloud systems, computational resources can be handled by exploiting the features of the PaaS level of the cloud stack, where speciic middleware can beneit from the security, scalability, reliability, and performance of the overall infrastructure. This kind of middleware is abstracted in our approach by the speciic implementation of the JobSubmissionWS interface, can be adopted by the SG engine to execute the computations supporting the monitoring services, and so can be easily integrated into the overall architecture. The JobSubmissionWS interface exports a set of brokering functionalities that are related to the interaction with a computational engine and are necessary for 124 Smart Grids the submission of an application and the management of execution results in both asynchronous and synchronous ways. In particular, following the grid broker pattern proposed in [25], the following portTypes are deined: 1. setJob. This is used to deine all the necessary information for the application execution (which includes information on the execution code and on the input data ile), information on the application structure and its execution requirements in relation to the resource coniguration (such as the operating system, a particular run-time environment, library, or transport communication layer, etc.), to QoS parameters, such as the maximum execution time, and, in the case of an economics-driven grid system, the maximum budget that can be spent. A call to the setJob portType returns an application ID, which can be used successively to perform actions on it as required. For generalization purposes, we considered the application setup made through a compressed ile, structured in a well-deined way, and understandable by the speciic underlying grid-computing middleware. This ile will typically contain source and executable iles or both, input data iles application, and QoS requirement description iles. 2. submitJob. This is used to verify that the computational engine will be effectively able to execute the application, respecting all the necessary hardware, software, and QoS requirements. 3. discardJob. When the submitter does not judge the execution offer of the computational engine to be suficient, he or she can remove it from the list of submitted applications. 4. executeJob. This is used to request the execution start of the application. 5. statusMonitor. This is used to continuously verify online the status of the application execution, such as information on the state of the overall execution (ready, running, waiting, ended, suspended, etc.). 6. collectResults. This is used for synchronous waiting for the normal execution termination, in order to receive contextual information on the output data. 7. terminateJob. This is used to abruptly terminate the application execution. 5.3.5 datastorageWs Storage systems are required to permanently store alarm conditions that can be analyzed off-line, for example, by using a web browser or a dedicated application. Since they can be adopted as a reference in network planning or as a knowledge base for expert systems based on decision support systems, these historical data are extremely useful in power system analysis. In order to be compatible with the rest of the framework, these storage systems have to export a DataStorageWS WSDL interface. So, a standardized and convenient method of accessing and managing distributed and heterogeneous data storage resources, such as a remotely accessed realtime database management system (RDBMS) or XML repository, must be deined. A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 125 5.3.6 WorkFloW enactor The worklow enactor composes and integrates web services that deliver different functionalities, related mainly to (1) real-time electrical data acquisition from different sources; (2) supply of the high computational power to develop SG monitoring and control functions; and (3) storage functionalities to handle huge amounts of data. The worklow enactor handles the execution request of the monitoring/control service described in a certain worklow language, which describes the temporal composition of the web services delivering the basic functionalities, submitted by means of the submission service. Figure 5.2 shows the action sequence performed by the worklow, referring to a single execution of the monitoring/control application [8]. Typically, the monitoring/control application has to be executed continuously, requiring for each execution the acquisition of current data on the SG’s state, particularly the following: • Phase 1. getNetworkTopology and 2. getControlAsset are performed to obtain the necessary SG description and the characteristic data of the available control and monitoring devices. • Phase 3. getCapabilitydata and 4. submitCapabilitydata are performed to acquire the input data for the generator/storage supply capability calculations, to arrange these data for each IED as a function of its geographical position, and to submit the corresponding information to the distributed IEDs. • Phase 5. getNodesParameters is performed to obtain the real-time data of each node and is made to interact with the service implementing the DataAcquisitionWS interface, typically delivered by the electrical network for analysis. DataAcquisitionWS Operator Submission Workflow engine Monitoring Computational engine 3. setJob 4. Submit 5. collectResults JobSubmissionWS FIGURE 5.2 Main actions of the worklow enactor. 6. Update DataStorageWS Sensors data acquisition Notification Data storage Submit Internet-http/IP interactions 1. getNetworkTopology 2. getNodesParameters 126 Smart Grids • Phase 6. setJob, 7. submit, and 8. collectResults are performed to request the execution of the control and monitoring computations by a computational engine, and, in particular, are made to interact with a service implementing the JobSubmissionWS interface exported by a grid computing system. These computations comprise: (1) the SG’s state estimation, computed by processing the network data acquired and solving the system state equations; (2) the checking of the distribution network operational constraints; (3) the assessment of the system’s security and reliability levels; and (4) the control and regulation functions of the local SG (i.e., power factor, frequency and voltage control, and load management). Finally, Phase 9, update, is executed to save the output data of the analysis in a data storage system interfaced through a service implementing the DataStorageWS interface. 5.4 EXPERIMENTAL SETTING In the following section, the implementation details of the basic web services and the worklow enactor for SG modeling and control are presented. 5.4.1 data acquIsItIon WeB servIce Field data can be acquired by means of geographically dispersed resources (i.e., SCADA systems, IED, and ield energy meters [FEM]). A prototype web service implementing the DataAcquisitionWS interface considering the FEM units located on each SG unit (load/generation/storage), and the direct interaction with them for real-time acquisition of the electrical parameters, was developed. In particular, a network of distributed FEMs based on ION 7330-7600 units was considered. They can measure demand on any instantaneous value and record peak (maximum) and minimum demand with a date and time stamp per second. The units are equipped with an onboard web server, which supports the XML format and the protocols ION, Modbus transmission control protocol (TCP), and Telnet, for their full remote control. The FEMs were physically connected to a fast ethernet local area network (LAN), conigured with an Internet protocol (IP) address. The authors prototyped the interaction through the http protocol to access, in particular, an XML ile, called realtime.xml, which contains the most recent measured ield data. To assess the supply capability curves of the SG components, a prototype version of an IED based on the ZWORLD-BL2100 unit (an advanced single-board computer that incorporates the Rabbit 2000 microprocessor) was adopted. These IEDs are installed directly on the SG components. They allow us to continuously monitor their state and to dynamically predict the corresponding supply capability curves as a function of the forecasted climatic variable proiles. These functionalities are fully available via an IP-based connection [23]. The authors are currently developing another implementation of the DataAcquisitionWS interface, considering the interaction with proprietary SCADAs. This new functionality will allow the integration of all the information A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 127 coming from existing distributed measurement systems, and so decrease the response time of the service. 5.4.2 JoB suBmIssIon WeB servIce The irst implementation of the JobSubmissionWS was based on the hierarchical metacomputer middleware (HiMM), version 1.1, developed at the University of Sannio [26]. The main features of HiMM that are useful in this context are (1) the virtual hierarchical topology, which, allowing applications to exploit dedicated networks or clusters that are not directly accessible by the Internet, makes the system highly scalable; and (2) the support for the object-oriented programming paradigm and, in particular, for the Java language, which simpliies application development and easy integration with web technologies. In addition, HiMM supports a grid broker service, which allows an application to exploit computing resources in a simpliied way, hiding the complexity of the underlying system. The grid broker delivers a speciic service to easily submit a master/slave application by using its nearly sequential version and some information necessary for deployment. It simpliies distributed programming, since it supports the separation of concerns related to functional (algorithmic issues, such as the creation of high-level domain-dependent abstractions in the form of objects and classes) and system (lower-level tasks such as object distribution, mapping, and load balancing) aspects of programming. The grid broker is responsible in particular for resource discovery and resource selection on the basis of QoS parameters speciied by a user (such as the desired deadline and the maximum budget), task mapping on the selected resources, and task scheduling according to the hierarchical master/ slave parallel model. In order to submit an application, the following information is required: (1) application code; (2) input data; and (3) application description and QoS requirements. For (3), descriptors based on XML called job description format (JDF) and user requirements description format (URDF) are to be deined. 5.4.3 comPutatIonal WeB servIce The computational grid used for performing the online computations was deployed on a test bed organized according to a logical hierarchical and heterogeneous topology of two levels, as shown in Figure 5.3 [27]. The irst level consists of a set of directly accessible heterogeneous resources composed of eight computers, and two front-end computers that were used as masters to coordinate the computations of two pools of computers located in two different buildings and representing the second level of the network. The irst pool consists of a cluster of workstations (COW) composed of eight homogeneous computers interconnected by a fast ethernet LAN through a hub 3Com TP16C. The second pool consists of a network of eight heterogeneous workstations (NOW), interconnected by a fast ethernet LAN through a Switch HP Procurve 2524. The irst-level machines are interconnected by a fast ethernet LAN. The mathematical solution engine adopted to support the online computations is based on a JavaRuntime Environment equipped with speciic power systems analysis 128 Smart Grids Servlet container HiMM broker (R2) Client Level 1 Fast ethernet LAN COW front end (R2) NOW front end (R2) COW NOW Node (R6) Node (R2) Node (R1) Node (R1) Node (R6) Node (R3) Node (R1) Node (R1) Node (R6) Node (R2) Node (R6) Node (R3) Node (R8) Node (R4) Node (R1) Node (R7) Node (R1) Node (R4) Node (R8) Node (R5) Node (R1) Node (R1) Node (R4) Node (R8) Level 2 FIGURE 5.3 Computational grid adopted for online MG monitoring and control. features. It allows system operators to investigate steady-state and dynamic issues in electrical networks by performing load low analysis, time domain simulations, and optimal economic dispatch. Various control and monitoring functions could be executed in parallel by the computational web service. It adopts a brokering service, based on an economydriven model, to satisfy the QoS constraints speciied by the user (i.e., a time deadline to solve a speciic computation). By using an economy-driven model, the broker is able to (1) identify the optimal asset of the computational resources needed to satisfy the user requirements (i.e., resource discovery and resource selection process); (2) transparently split a sequential object-oriented task (i.e., a set of contingencies) into subtasks (i.e., a subset of contingencies) according to a hierarchical master/slave computing model; and (3) automatically distribute subtasks to the set of selected computational resources. In order to evaluate the parallel processing performances of the computational web service, many experimental tests were executed by varying the number of machines used for the computational engine in order to evaluate its speedup factor. The computation time and the speedup obtained by increasing the slaves in powers of two (two, four, and eight slaves) were evaluated. Each computation was performed several times in order to use the average. The speedup obtained by using a two-level architecture is nearly linearly related to the number of slave machines and is very close to the ideal case. A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 129 5.4.4 data storage WeB servIce In this irst phase, the authors used a relation database, installed on a Linux workstation and based on the Postgres RDBMS. 5.4.5 WorkFloW enactor The main implementation issue related to the worklow enactor is the choice of a worklow language and its related technologies, which can be usefully adopted to describe the wireless asset monitoring (WAM) application in the context of a web services–based approach. Many proposals have been made in the literature, especially by the major web services software providers, and many of them are continuously evolving. In these proposals the application worklow is usually described with a textual scripting language (typically the XML format) that is used to provide the “glue” that links involved services together. Among the possible solutions, we chose business process execution language (BPEL). BPEL is a programming abstraction that allows developers to compose multiple discrete web services into an end-to-end process low. At the present time it could be considered the most complete, and has risen in popularity since many vendors are developing their toolkits, many of which are also becoming commercial. Among the available BPEL engines, we chose to integrate in our framework the worklow enactor delivered by BPWS4J, since it is free and suficiently stable. The BPWS4J toolkit includes a worklow execution platform, a tool that validates BPEL4J documents, and an Eclipse plug-in that provides a simple editor for creating and modifying BPEL4J iles. In particular, the worklow enactor of our platform is based on the worklow execution of BPEL4J integrated in Apache Tomcat, version 5.0.24, and is specialized to process the online power systems security analysis (OPSSA) worklow. For each process, the BPWS4J engine takes in a BPEL document that describes the process worklow to be executed, a WSDL document (without binding information) that describes the interface that the process will present to clients, and WSDL documents that describe the services that the process may or will invoke during its execution. From this information, the process is made available as a web service with a SOAP interface. The SG worklow describes the composition of the web services involved, where each service is deined as a <partner>. Service composition in the SG worklow is deined using a sequence tag, which deines how the partners will be sequentially executed. The speciication of a sequence includes deinitions of the input message (the receive tag), of the service invocation (the invoke tag), and of the output message (the reply tag). The assign tag establishes the relationship between the output message of a service and the input message of the service in the worklow. Similarly to the individual web services, which are described through the corresponding WSDL ile, the SG web service is also described in a WSDL ile (Figure 5.4). Messages, portTypes, and operations are also deined, and there is an additional section (serviceLinkType tag) to deine the role of each service during their interaction. 130 Smart Grids <definitions …. <message name="StringMessageType"> <part name="echoString" type="xsd:string"/> </message> <wsdl:message name="OPSSARequest"> <wsdl:part name="in0" type="xsd:string"/> </wsdl:message> <wsdl:message name="OPSSAResponse"> <wsdl:part name="setJobReturn" type="xsd:int"/> </wsdl:message> <portType name="OPSSAService"> <operation name="execOPSSA" parameterOrder="in0"> <input name="OPSSARequest" message="tns:StringMessageType"/> <output name="OPSSAResponse” message="tns:StringMessageType"/> </operation> </portType> .….. </definitions> FIGURE 5.4 WAM WSDL ile. The web services implemented, in particular the data acquisition and the job submission web services, were written using the Java language and the Apache Axis toolkit. The Axis toolkit essentially included an SOAP engine, that is, a framework for constructing SOAP processors such as clients and servers, a simple stand-alone server, a server that plugs into servlet engines such as Apache Tomcat, and extensive support for WSDL. 5.5 CONCLUSIONS The chapter proposed an advanced web services–based framework for integrated SG control, modelling, and monitoring. The core component of the proposed framework is the SG engine, which is responsible for the execution of the SG management functions in a geographically distributed scenario. It includes a network of remotely controlled units distributed in the most critical SG sections for ield data acquisition and advanced protective functionalities; a GIS interface for the acquisition of georeferentiated information on the SG; a grid computing–based solution engine for the online computations; and a web-based interface for graphical synoptic and reporting development. The web services–based framework has been deployed on a computational grid. The irst experimental results have emphasized the important role of web services technology in SG control, modelling, and monitoring. The application of this technology is very promising, since it makes the proposed framework platform-, language-, and vendor-independent, and thus an ideal candidate for an effective integration in existing energy and distribution management systems (EMS/DMS). The future directions of our research are oriented toward the conceptualization of a fully decentralized nonhierarchical architecture for SG monitoring and control that A Distributed Framework for Smart Grid Modeling, Monitoring, and Control 131 will move away from the older centralized paradigm to a system distributed in the ield with an increasing pervasion of intelligence devices (smart sensors). The adoption of smart sensors is expected to lead to a more eficient task distribution among the monitoring system resources and, consequently, to signiicantly less demand on centralized computing resources. REFERENCES 1. S. M. Amin and B. F. Wollenberg, Toward a smart grid, IEEE Power and Energy Magazine, 3(5), 34–38, 2005. 2. A. Schroder, I. Laresgoiti, K. Werlen, B. Schowe-von der Brelie and A. Schnettler, Intelligent self-describing power grids, in Proceedings of the 20th International Conference and Exhibition on Electricity Distribution (CIRED), 8–11 June 2009, Prague, Czech Republic. 3. R. L. King, Information services for smart grids, in Proceedings of 2008 IEEE Power and Energy Society General Meeting—Conversion and Delivery of Electrical Energy in the 21st Century, 20–24 July 2008, Pittsburgh, PA. 4. N. Hadjsaid, J.-F. Canard and F. Dumas, Dispersed generation impact on distribution networks, IEEE Computer Applications in Power, 12(2), 22–28, 1999. 5. R. Comfort, A. Mansoor and A. Sundaram, Power quality impact of distributed generation: Effect on steady state voltage regulation. Presented at the PQA 2001 North America Conference, 2001, Pittsburgh, PA. 6. J. Zhu, Web services provide the power to integrate, IEEE Power and Energy Magazine, 1(6), 40–49, 2003. 7. Q. Liu, J. Wen and H. Sun, Implementing power system management via semantic web services composition, in Lecture Notes in Computer Science, Springer, Berlin, 2009. 8. Q. Morante, N. Ranaldo and E. Zimeo, Web services worklow for power system security assessment, in Proceedings of IEEE International Conference on e-Technology, e-Commerce, e-Service, IEEE Computer Society Press, 29 March–1 April 2005, Hong Kong. 9. J. R. Roncero, Integration is key to smart grid management, in IET-CIRED Seminar on SmartGrids for Distribution, 23–24 June 2008, Frankfurt. 10. H. Johal, W. Ren, Y. Pan and M. Krok, An integrated approach for controlling and optimizing the operation of a power distribution system, in 2010 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT Europe), Institute of Electrical and Electronics Engineers, 11–13 October 2010, Gothenburg. 11. B. Ramachandran, S. K. Srivastava, D. A. Cartes and C. S. Edrington, Distributed energy resource management in a smart grid by risk based auction strategy for proit maximization, in Power and Energy Society General Meeting, IEEE, 25–29 July 2010, Minneapolis, MN. 12. A. Mahsoom, O. Olusola and L. Jonathan, Exploring control systems to implement SMART GRID solutions, in Proceedings of 10th IET International Conference on Developments in Power System Protection (DPSP 2010), 29 March–1 April 2010, Manchester, UK. 13. T. M. Overman and R. W. Sackman, High assurance smart grid: Smart grid control systems communications architecture, in 2010 First IEEE International Conference on Smart Grid Communications (SmartGridComm), 4–6 October 2010, Gaithersburg, MD. 14. M. Pipattanasomporn, H. Feroze and S. Rahman, Multi-agent systems in a distributed smart grid: Design and implementation, in IEEE/PES Power Systems Conference and Exposition, PSCE ‘09, 15–18 March 2009, Seattle, WA. 132 Smart Grids 15. V. M. Catterson, E. M. Davidson and S. D. J. McArthur, Agents for active network management and condition monitoring in the smart grid, in First International Workshop on Agent Technologies for Energy Systems, 11 May 2010, Toronto, ON. 16. F. Li, P. Zhang and N. Bhatt, Next generation monitoring and control functions for future control centers, in IEEE PES General Meeting, 26–30 July 2009, Calgary, AB. 17. J. G. Cupp and M. E. Beehler, Implementing smart grid communications, TECHBriefs, No. 4, 5–8, 2008. 18. D. Baigent, M. Adamiak and R. Mackiewicz, IEC 61850 Communication Networks and Systems in Substations: An Overview for Users, SIPSEP, Monterrey, Mexico, 2004. 19. A. Apostolov, C. Brunner and K. Clinard, Use of IEC 61850 object models for power systems quality/security data exchange, in Quality and Security of Electric Power Delivery Systems. CIGRE/IEEE PES International Symposium, pp. 155–164, 8–10 October, Montreal, Canada, 2003. 20. M. C. Janssen, C. A. Andersen, C. Brunner and J. Dall, Modelling of DER schedules using IEC 61850, in Proceedings of 20th International Conference on Electricity Distribution, 8–11 June 2009, Prague, Czech Republic. 21. C. Brunner, IEC 61850 for power system communication, in Proceedings of IEEE/PES Transmission and Distribution Conference and Exposition, 21–24 April 2008, Chicago, IL. 22. P. Lloret, J. L. Velasquez, L. Molas-Balada, R. Villafaila, A. Sumper and S. GalceranArellano, IEC 61850 as a lexible tool for electrical systems monitoring, in Proceedings of the 9th International Conference on Electrical Power Quality and Utilisation, EPQU, 9–11 October 2007, Barcelona, Spain. 23. A. Vaccaro and D. Villacci, Prototyping of a distributed JAVA based client/server architecture for e-Assessment of power equipments loading capability, in Proceedings of IEEE-Powertech, 22–26 June 2003, Bologna, Italy. 24. M. Di Santo, N. Ranaldo, A. Vaccaro and E. Zimeo, Java-based distributed architectures for intensive computations in electrical grids, in Proceedings of International Workshop on Java for Parallel and Distributed Computing, April 2004, Santa Fe, NM. 25. M. Irving, G. Taylor and P. Hobson, Plug in to grid computing, IEEE Power and Energy Magazine, 2(2), 40–44, 2004. 26. M. Di Santo, N. Ranaldo and E. Zimeo, A broker architecture for object-oriented master/slave computing in a hierarchical grid system, in Parallel Computing, September 2003, Dresden, Germany. 27. Q. Morante, N. Ranaldo, A. Vaccaro and E. Zimeo, Pervasive grid for large-scale power systems contingency analysis, IEEE Transactions on Industrial Informatics, 2(3), 165–175, 2006. 6 Role of PLC Technology in Smart Grid Communication Networks Angeliki M. Sarai, Artemis C. Voulkidis, Spiros Livieratos, and Panayotis G. Cottis CONTENTS 6.1 6.2 Introduction .................................................................................................. 133 Overview of Smart Grid Networks ............................................................... 135 6.2.1 Energy-Related Applications/Services ............................................. 136 6.2.2 Applications/Services Related to Smart Grid Communications ...... 136 6.2.3 IT Applications/Services .................................................................. 138 6.3 Role of PLC in the Smart Grid Communications Infrastructure ................. 139 6.3.1 PLC in CPNs..................................................................................... 140 6.3.2 PLC in NANs.................................................................................... 141 6.3.3 PLC in FANs .................................................................................... 142 6.3.4 PLC in SNs ....................................................................................... 143 6.4 Aggregation and Power Control in SGC Networks Employing PLC ........... 143 6.4.1 Aggregation of PLC Trafic .............................................................. 143 6.4.2 Power Control at PLC Information Generators ................................ 144 6.5 Design Guidelines for Smart Grid Implementations .................................... 145 6.6 Conclusions ................................................................................................... 146 References .............................................................................................................. 147 6.1 INTRODUCTION Although the demand for electric power is increasingly changing, the dimensioning and speciications of electric power systems are greatly inluenced by statistical studies and historical proiles of energy demand. The biggest part of the power grid was engineered several decades ago based on the operational conditions and the technological options of that time. Since then, decisive changes in both the size and proile of electricity demand have taken place that necessitate the upgrading of major operational modules or parts of the equipment and, also, changes in the switching functionalities of the power grid with regard to load redistribution and balancing. 133 134 Smart Grids These important issues can be addressed either by applying proper strategies to manage power demand or by appropriately incorporating distributed energy resources (DERs). Furthermore, the mandatory use of smart energy meters combined with lexible pricing tools can mitigate the cost of congestion by limiting the peak load where the electric energy is delivered. At the same time, demand control from the consumption side should be enabled. According to the European Union’s 20-20-20 smart grid (SG) target toward the low-carbon economy, a 20% reduction in CO2 emissions, an increase of the share of renewable sources to 20%, and 20% lower energy consumption are the targets for 2020 [1]. Thus, a major challenge for utilities is to provide reliable and eficient SG services that will enable seamless integration of DERs and support of smart metering applications, while enhancing power quality through sophisticated distribution automation (DA) services. The SG constitutes the modern approach to monitoring and controlling the power grid. As such, the SG can be regarded as an information and communication technology (ICT)-enhanced electric power network, which can intelligently integrate/incorporate the actions of all related stakeholders, namely, the generators, distributors, and consumers of energy. As the main target of the SG is to deliver electrical power in a sustainable, economical, and secure way, the development of reliable and eficient data transmission schemes is imperative. Recent developments in SG communications (SGCs) technologies address the crucial objectives of reliability, state awareness, and inancial viability of the SG stakeholders, producers/operators/customers, and society as a whole. These technologies employ wireless and wired options that are usually integrated and operate interchangeably to form hybrid communications platforms that serve the SG. The SGC networks (SGNs) support the two-way transfer of SG data employing a wide range of telecommunication technologies. Indicative examples encountered in SG implementations are cellular technologies [2], power line communication (PLC) options [3], Wi-Fi [4], and optical technology [5]. The resulting hybrid communications platform, usually consisting of multiple, sometimes superposed, networks, should have suficient capacity to accommodate the quality of service (QoS)-enabled SG trafic and guarantee an all-Internet protocol (IP), secure, two-way, end-to-end transfer of the SG data. A critical question concerning SGCs is which transmission medium should be employed each time. As the SG is composed of a variety of constituent networks, namely, high-speed enterprise networks, ield area networks (FANs), substation networks (SNs), and customer premises networks (CPNs), the answer is not unique. The present chapter explores the role of the PLC transmission medium in the hybrid communications platform employed by SGNs and proposes relevant techniques. The recently standardized broadband over power lines (BB-PLC) networks [6,7] aim at effectively serving QoS-enabled trafic. The ubiquitous presence of the power grid, along with the easy and scalable deployment of the PLC SGNs over the power grid [8], renders BB-PLC the only broadband access technology readily available to remote or rural areas [9]. The above considerations, together with the independence of utilities from a third-party communications provider, render PLC technology a very promising SG option, especially in the medium-voltage (MV) grid. 135 Role of PLC Technology in Smart Grid Communication Networks The rest of the chapter is organized as follows. In Section 6.2, SG networks are outlined, considering the relevant energy, communications, and information technology (IT) aspects. In Section 6.3, the perspectives of SGCs are analyzed and the role of the PLC is examined in the framework of the various networking options arising in SGNs. In Section 6.4, PLC transmission techniques are examined in their attempt to handle QoS-enabled SG data. The relevant design guidelines are given in Section 6.5. Finally, conclusions are drawn in Section 6.6. 6.2 OVERVIEW OF SMART GRID NETWORKS There have been several attempts in the literature to present the fundamental elements of SGs from various aspects. The authors in [10] present the systems constituting the SG, namely, the smart infrastructure system, the smart management system, and the smart protection system. The authors in [11–13] follow the National Institute of Standards and Technology (NIST) conceptual model [14], depicted in Figure 6.1. Also, the authors in [15] consider SGs from the IT perspective, presenting reference models related to the application layer. Finally, a layered approach is proposed in [16,17]. This section adopts the Institute of Electrical and Electronics Engineers (IEEE) 2030 [18] interoperability reference model to classify the SGs under three perspectives, Conceptual model Operations Markets Service provider Bulk generation Transmission Distribution Customer Secure communication interface Electrical interface Domain FIGURE 6.1 NIST conceptual model of the smart grid. (National Institute of Standards and Technology, NIST framework and roadmap for smart grid interoperability standards, release 1.0, January 2010. Available at: http://www.nist.gov/public_affairs/releases/upload/ smartgrid_interoperability_inal.pdf [accessed March 2013].) 136 Smart Grids namely, (i) the energy perspective, (ii) the communications perspective, and (iii) the IT perspective, and presents the related fundamental applications/services. 6.2.1 energy-related aPPlIcatIons/servIces As depicted in Figure 6.1, systems interconnected over the power grid include bulk generation, the transmission and distribution networks, and the customer premises. The seamless integration of such entities requires a reliable, two-way low of both electric power and information. The following challenges should be addressed: • Bulk generation: A successful power grid operation depends strongly on the equilibrium between energy production and energy consumption. The main challenge on the power generation side is to incorporate intermittent DERs to increase microgrid penetration and enhance the sustainability of the power grid. However, the generation of DERs exhibits severe luctuations that prohibit their seamless integration into the power grid [19] and disrupt the equilibrium between energy generation and consumption. It is expected that the formulation of appropriate prediction models will signiicantly increase the reliability of DERs and, consequently, their use in generation. • Transmission and DA: To increase power quality and redundancy, the smart infrastructure should enable DA services employing sensors and actuators throughout the grid to monitor power production and energy demand and detect faults and component failures. Also, though the substations are already supported by supervisory control and data acquisition (SCADA) automation services, there is still the need to incorporate digital devices that monitor and control the operation of the power grid in an autonomous way. • Customer premises: The end users can interact with the SG in various ways. First, real-time monitoring of power consumption is enabled through automatic meter reading/advanced metering infrastructure (AMR/AMI) systems. The operator of the power grid provides demand-response (DR) applications for both direct and indirect load control, where demand-side management (DSM) applications notify the customers about excessive energy demand and provide incentives to reduce consumption in an attempt to reduce peak load. Moreover, SG customers are potential prosumers; that is, they not only consume but may also produce energy. In addition to DERs, new load paradigms such as plug-in hybrid electric vehicles (PHEVs) [20] are present in the SG, as PHEVs consume energy while charging and return energy to the power grid when they are fully charged [10]. 6.2.2 aPPlIcatIons/servIces related to smart grId communIcatIons From the SGC point of view, a variety of access technologies and protocols should be engaged to support a two-way information low to/from the entities appearing in Figure 6.1. In Figure 6.2, the reference communications model for the SG deined in the IEEE 2030 standard [18] shows how SGNs act as an intermediate layer between 137 Role of PLC Technology in Smart Grid Communication Networks End-to-end smart grid communications model Communications network security layer Communications network management layer Utility local area network (LAN) Public Internet Regional/ metropolitan area networks (MAN) Backbone/core network Substation network Wide area networks (WAN) Bulk generation Transmission Backhaul Customer premises Last mile Distribution substation network HAN: Home area “hot spots” network NAN/AMI: Neighborhood BAN: Business/building area network/advanced area network metering infrastructure IAN: Industrial area EAN/AMI: Extended area Backhaul network network/advanced metering infrastructure FAN: Field area network Smart Distribution meter Customer HAN Nonrenewable BAN Substation Renewable IAN Distributed generation/microgrids Renewable/ microgrids Smart grid power and electric system layer FIGURE 6.2 The IEEE 2030-2011 smart grid communication layered model. (IEEE guide for smart grid interoperability of energy technology and information technology operation with the electric power system [EPS], end-use applications, and loads, IEEE Std 2030-2011, pp. 1–126, 10 September 2011.) the power grid and the network management layer. As shown in Figure 6.2, multiple networks are employed to support the various QoS requirements of SG services. Speciically, the SG services are associated with the following types of networks: • CPNs: CPNs vary according to their size and to the number of serving devices and can be classiied into home area networks (HANs), building area networks (BANs), and industrial area networks (IANs). Regardless of the CPN type, local area network (LAN) technologies are used to provide connectivity within CPNs that are employed to support applications such as AMR, remote load control, monitoring and control of DERs, energy consumption management through DSM, and charging of PHEVs. • Neighborhood area networks (NANs): NANs are employed to collect information generated by CPNs. Usually, a NAN comprises the AMI installed to collect the AMR data. NANs constitute the last mile of the SGNs, as they are intermediate between the customer and the distribution grid. The signiicant components of NANs are their distribution access points (DAPs) operating as gateways for the trafic collected via the AMI. The DAPs decide on data aggregation and routing, based on the type and signiicance of the transferred SG trafic. 138 Smart Grids • FANs: FANs are deployed over the distribution grid to monitor and control various ield devices, such as insulators, feeders, and transformers. These devices are dispersed over the grid to enable DA services and services related to the seamless integration and management of DERs. As FANs cover the whole distribution grid, they may also serve for utilities communications, offering communication and monitoring services to the technical personnel responsible for grid maintenance and day-to-day operation. • SNs: SNs are highly reliable enterprise LANs used for protection, monitoring, and automation services. Among them, SG protection and SCADA services require a maximum delay of 10 ms [21]. Hence, to meet the QoS requirements of teleprotection services, SNs must usually have a one-hop point-to-point link to the utility operations and control center. Table 6.1 provides an overview of the available transmission media along with their advantages and disadvantages. The options tabulated in Table 6.1 may be employed in different segments of SGNs. 6.2.3 It aPPlIcatIons/servIces The IT applications involved in the SG primarily support the management and control of various procedures and devices dispersed over the grid. The data related to SG operation come from a plethora of sources either internal or external to the grid. Internal sources include: (i) devices for substation monitoring, (ii) smart units attached to various grid nodes such as AMR devices and smart sensors, and (iii) databases employed by the utilities as well as information originating from human intervention. Next, the applications and services supported by various segments of the grid are presented: • Bulk generation: Applications related to the management of DER generation are supported. Also, generation scheduling is enabled and relevant data are sent to the various stakeholders of the electric energy market. • Transmission and distribution grids: Data acquisition and collection from devices installed both in the ield and in substations are supported all over the SG. An overview of various QoS requirements for SCADA automation applications can be found in [21–23] and in the IEEE standard for substation automation, IEEE 1646 [24]. Also, maintenance and day-to-day activities that process data collected by geographic information services are performed to guarantee power availability. Regarding the distribution grid, applications related to DER management and integration are offered. Finally, parts of the grid that connect end users should support data aggregation for better collection and management of billing data. • Customer applications/services: Applications related to the seamless integration of DERs must be provided. Also, various energy management systems and home automation applications may offer real-time monitoring of energy consumption. Finally, applications accessed through portals may provide data and services related to energy usage, billing, and customer authorizations. Role of PLC Technology in Smart Grid Communication Networks 139 TABLE 6.1 Smart Grid Communications Media Overview Network Access Technology Cellular networks (GPRS/3G/4G) Wi-Fi (IEEE 802.11) Ethernet Data Rates (Maximum Values) GPRS 56–115 Kbps 3G 2 Mbps 4G 50–100 Mbps ~100 Mbps Advantages Widely adopted for certain types of smart grid applications Easy installation Reduced cost 100 Mbps–10 Gbps Easy installation Reduced cost WiMAX (IEEE WiMAX (8062.16 m) Wide coverage 802.16) and LTE 300 Mbps Scalability LTE 100 Mbps Fiber optics 40–1600 Gbps (when High capacity (SONET/SDH WDM is employed) Reliability E/GPON) Security DSL 20 Mbps Easy installation Scalability PLC (NB/BB/ BB-PLC The transmission medium BPL) 500 Mbps on CPNs already exists Third-party independence WSN/WPAN ZigBee Reliability (IEEE 802.15.4 250 Kbps Low delay ZigBee, DASH7 DASH7) 200 Kbps 6.3 Disadvantages Offers only periodic service monitoring Dependence on third parties Limited scalability and reliability Limited scalability Cost related to leasing spectrum High cost Not available in rural areas Public access network Nonuniversal legislation Adverse transmission medium Short coverage ROLE OF PLC IN THE SMART GRID COMMUNICATIONS INFRASTRUCTURE Although various communication options are available to support SGC, PLC is preferable on many occasions for the following reasons: • The deployment of the PLC segments of hybrid SGC solutions is immediate, simple, and scalable. • As far as the PLC segments of SGNs are concerned, data security oficers (DSOs) do not have to procure telecommunication services from third parties. • The privacy and independent processing of SG data are guaranteed, as the DSOs are the owners of the PLC part of SGNs. • As PLC technology employs the power lines themselves as the physical transmission medium, it is the only SGC option that offers monitoring 140 Smart Grids of the operational status of the electric grid by appropriately processing inherent noise patterns and spurious signals collected by the PLC nodes. • Alternative technologies to PLC can serve only part of the desired applications, especially in the MV SG. In any case, the PLC nodes may usually embed any related technology in order to (i) increase their functional and communication capabilities and (ii) provide compatibility with alternative communication options or existing or future automation technologies. • Direct and selective scalable integration of all application requirements of the SG can be implemented at the level of the grid nodes. The rest of the section explores the role of the PLC in SGNs of the SG. 6.3.1 Plc In cPns The CPNs constitute a multiprotocol, multivendor environment [18], in which many technologies are involved, with PLC and Wi-Fi being the most mature. Other solutions intended for the SG employ Global System for Mobile Communications/ General Packet Radio Service (GSM/GPRS) and ZigBee technologies. An overview of the communication technologies used in CPNs, along with their respective level of maturity, is presented in Table 6.2 [25]. In most cases, BB-PLC seems preferable to narrowband (NB)-PLC as it may provide a variety of broadband applications requiring fast Internet access. NB-PLC may be the solution for transferring low-rate data over large distances along the MV grid in remote areas. The main efforts to standardize BB-PLC access resulted in the International Telecommunication Union Telecommunication Standardization Sector (ITU-T) G.hn [7] and IEEE 1901 [6] protocols summarized in Table 6.3. The operational principles of the two standards are presented in [26–30]. The large amount of trafic moving through an SGN, along with the necessity to prioritize data with regard to their importance and/or bandwidth requirements, renders the time division TABLE 6.2 Communication Technologies Employed in CPNs Communication Technologies PLC ZigBee Wi-Fi WiMAX GSM/GPRS DASH7 HAN IAN AMR PHEV √ √ √ ~ √ Δ √ √ √ ~ √ Δ √ √ √ Δ Δ Δ ~ √ Δ Δ √ Δ Source: Usman, A. and Shami, S.D., Renewable and Sustainable Energy Reviews, 19, 191–199, 2013. Notes: √, in use, some mature solutions available; Δ, not currently in use, solutions can be developed; ~, ongoing research, some solutions available but under testing. Role of PLC Technology in Smart Grid Communication Networks 141 TABLE 6.3 BB-PLC Technologies Employed in CPNs IEEE 1901 Frequency band Maximum supported data rate OFDM FEC ITU-T G.hn FFT-PHY Wavelet-PHY 2–25 MHz 2–50 MHz 2–100 MHz 1 Gbps 2–30 MHz 2–48 MHz 2–60 MHz 545 Mbps 2–28 MHz 2–60 MHz Windowed 2048, 4096 carriers FFT 3072, 6144 carriers Turbo convolutional code BPSK, QPSK, 8-, 16-, 64-, 256-, 1024-, 4096-QAM CSMA/CA (4 priority levels) TDMA Wavelet 512, 1024 carriers RS, RS-CC, LDPC QC-LDPC Modulation BPSK, QPSK, 8-, 16-, 64-, 256-, 1024-, 4096-QAM Contention-based MAC scheme Contention-free MAC scheme CSMA/CA (4 priority levels) TDMA 544 Mbps BPSK, 4-, 8-, 16-, 32-PAM CSMA/CA (8 priority levels) TDMA Source: Rahman, M.M., Hong, C.S., Lee, S., Lee, J., Razzaque, M.A., and Kim, J.H., IEEE Communications Magazine, 49, 183–191, 2011. multiple access (TDMA) contention-free medium access control (MAC) option preferable to the carrier sense multiple access (CSMA) contention-based option when reliable SG services are required. On the other hand, noncritical broadband services employ contention-based access. 6.3.2 Plc In nans NANs incorporate the AMI components to collect the trafic generated in CPNs. Figure 6.3 shows an indicative network architecture, where the NANs employ PLC links. The trafic generated in the CPNs is routed to the network gateways, namely, the NAN DAPs, which may also serve for data aggregation. Speciically, critical data are usually routed to the utility premises without being aggregated, whereas data aggregation is usually applied to noncritical data. Even though only the PLC option is employed to implement the NANs in Figure 6.3, strict QoS constraints may necessitate a redundant SGC infrastructure in order to support time-critical applications. The PLC links encountered in CPNs are short, so that BB-PLC may be effectively used. This is not the case in NANs, where longer links are encountered. Therefore, due to the lower attenuation in lower frequencies and the low-rate requirements in 142 Smart Grids Utility operation and control Substation network Distribution network Neighborhood area network (NAN) Field area network (FAN) Customer premises network AMR DA AMR DER AMR DA DA High-speed wireless backbone DER DAP PLC link Wireless link FIGURE 6.3 Indicative heterogeneous SGN architecture. NANs, NB-PLC is preferable for NANs, especially to support DR applications, as argued in [31], where the PLC options available are compared with regard to their applicability in NANs. As to the standardization of PLC access networks, there is an ongoing process by the IEEE P1901.2 Task Group to standardize NB-PLC [32]. With regard to BB-PLC, although many PLC options are currently available for CPNs, this is not the case for NANs, since the only broadband option supporting access networks is IEEE 1901. The BB-PLC NANs may effectively support scalability and redundancy, and, when combined with node clustering approaches [33,34], they may successfully meet the various QoS constraints of SG applications. Also, the use of repeaters and the proper design of BB-PLC networks may render the BB-PLC option capable of supporting the trafic moving through the NANs [9]. In addition to PLC, other options may be employed to support the NAN infrastructure, provided that they guarantee reliable and secure data transmission. As to private wide area network (WAN) solutions, there is active research on licensed wireless solutions in an attempt to serve delay-sensitive applications [35–37]. Other mature solutions include Wi-Fi and GPRS/GSM infrastructures [38]. 6.3.3 Plc In Fans FANs are usually incorporated into NANs, since they are both SGNs deployed over the MV grid. However, while NANs usually interface with metering devices dispersed over CPNs, FANs interface with wireless mesh options, that is, IEEE 802.15.4 protocol, necessitating a different communications platform. The authors in [39] classify FANs as typical cases of low-power and lossy networks (LLNs), since they exhibit connection losses and unpredictable errors caused by the severe voltage luctuations and noise caused by the operation of ield devices. Thus, FANs usually require different QoS handling, employing network layer protocols such as IPv6 over wireless personal area networks [40,41] to address the issues of end-to-end connectivity and reliability. Role of PLC Technology in Smart Grid Communication Networks 143 The low-rate applications supported by FANs render NB-PLC appropriate for FAN transmission, as claimed by the upcoming IEEE P1901.2 standard. Also, interoperability with the wireless IEEE 802.15.4 parts is imperative. Relevant key applications that require sophisticated network approaches to handle the heterogeneous nature of FANs are related to DA [31]. Also, PLC may effectively be used to support the detection and localization of faults by monitoring the signal-to-noise ratio (SNR) of the PLC signals [42]. 6.3.4 Plc In sns The SNs are usually the irst SGNs to implement, since they support the most critical applications realized in the SG. Substation automation services include the remote monitoring and control of SCADA systems, phase measurement units (PMUs), remote terminal units (RTUs), and intelligent electronic devices (IEDs). The common characteristic of all these devices is that they generate SG information that needs to be collected via the backhaul networks, that is, the SN, in real-time synchronized mode of operation. Since synchronization is a critical issue, global positioning system (GPS) enabling is usually proposed in the literature [24]. As to PLC transmission, both BB-PLC and NB-PLC seem appropriate to serve SN communications, taking into account that the former option supports low whereas the latter supports high transmission rates. Other options may include wireless and optical media [43]. 6.4 AGGREGATION AND POWER CONTROL IN SGC NETWORKS EMPLOYING PLC The extended ICT infrastructure that enhances the traditional power grid with artiicial intelligence gives rise to the generation of SG information. That is, as power generators generate power, SG entities, either AMR or other SG devices at the customer premises, act as generators of SG information, which, in the framework of PLC-enabled SGNs, is transferred via the power lines. The huge amount of data produced by the SG information generators necessitates proper management to optimize their exploitation. Moreover, as the number of SG information generators scales up, so does the need for proper handling of the resulting data trafic; the pursued facilitation of power grid monitoring comes at the cost of the growing complexity of the underlying communications architecture. The extreme heterogeneity of SG information generators and the necessity to support a variety of SG information services increase the need to optimize the SGN operation. 6.4.1 aggregatIon oF Plc traFFIc Depending on the type of trafic produced by the smart devices and on the associated services and business-to-business (B2B) or business-to-customer (B2C) service level agreements (SLAs), the SG data trafic may be categorized into several classes. Certain applications, such as substation automation, require extremely low transmission delay and are classiied as time-critical applications. Other applications, such as the transmission of AMR data generated in CPNs, which do not require an 144 Smart Grids immediate response fall into the class of delay-tolerant services. From the data accuracy point of view, applications such as AMR or voice over IP (VoIP) may tolerate a certain degree of data loss, allowing the application of lossy aggregation techniques, whereas applications such as substation automation require highly accurate data transmission from the SG data source to the utility’s back ofice, and vice versa. The various QoS constraints explicitly imposed by the requirements of the services and applications supported may be properly traded off in the attempt to increase the eficiency of an SGN and, in certain cases, to increase its service capacity to accommodate more services or applications. In any case, imposing QoS constraints for satisfactory data services provisioning critically affects the operational characteristics of NANs, FANs, and SNs. Data aggregation has been considered in the recent literature as a mechanism to manage the trade-off between communications eficiency and data accuracy [44,45]. Moreover, several recent studies have highlighted data aggregation as a way to preserve anonymity, without reducing the contextual information of the data reported to the utilities [46]. Since security and privacy are of primary importance for SG applications, data aggregation seems appealing for application in SGNs. Although originally employed in the framework of wireless sensor networks (WSNs), the principles of data aggregation are also valid in the case of SGNs [47]. As the topology of the distribution grid may be modeled as a tree structure, and WSNs usually form tree-like structures for the routing of information, many data aggregation protocols destined for WSNs can be adapted to serve SGNs as well. The design targets determining the application of data aggregation to QoSconstrained SGNs are primarily related to determining (i) which data lows can or should be aggregated, (ii) where the aggregation points should be located, and (iii) which aggregation procedure guarantees conformity with the delay-related QoS constraints of certain SG service classes. Notably, as delay reduction comes at the cost of bit error rate (BER) deterioration, the above design targets are not easy to achieve. Indicatively, determining the optimal aggregation tree in WSNs and SGNs may be reduced to the solution of the well-known minimum spanning tree (MST) problem, which is known to be NP-complete [48]. Also, proper scheduling of data transmission reporting by the SG devices, which is related to serving time-critical applications, has been widely studied and has been proven to be NP-hard [49]. Taking into account the above considerations, determining the appropriate parameters for SGN operation is a dificult task, relecting the computational complexity that comes as a side effect of the enhanced monitoring and automation functionality offered by the SG. 6.4.2 PoWer control at Plc InFormatIon generators In line with the need to optimize the transmission schedules of the information generators, the recent IEEE P1901 standard proposes two approaches for proper MAC, one based on the CSMA family of protocols and one implementing TDMA [6]. The latter establishes the necessity for collision-free transmission periods to (i) alleviate the negative effects of restoring information from collided packets received from the SG devices side and (ii) guarantee QoS provisioning for time-critical applications. Role of PLC Technology in Smart Grid Communication Networks 145 Although general directions toward exploiting collision-free periods are given, exact scheduling algorithms have not yet been proposed. To cope with the lack of a standard MAC scheme, several approaches have been recently proposed [50,51]. These approaches apply techniques originally destined for wireless transmission. However, in all the schemes proposed, any similarities of PLC transmission with wireless transmission were not fully exploited; multipath propagation, severe signal attenuation, and background noise drastically aggravate the eficiency of PLC transmission and should be taken into account. To this end, proper control of the injected power spectral density (IPSD) by the SG devices may offer advantages in enabling frequency reuse through transmission-slot reuse. By controlling the IPSD of the SG information generators, the distribution grid may be segmented into independent transmission areas, where simultaneous transmissions are allowed, increasing SGN throughput while reducing the transmission delay. The merits expected from the employment of power control combined with data aggregation in the context of tree-like PLC SG structures are presented in [49–52]. Interestingly, apart from the documented communications-related beneits, power control may also yield signiicant energy savings, which, considering the large numbers of SG information generators employed in SGNs, may prove nontrivial. In conclusion, the introduction of ICT technologies enables full monitoring and control of the distribution grid at the cost of increased complexity and reduced eficiency of the necessary communication schemes. However, data aggregation alone or combined with proper IPSD control of the SG information generators may alleviate the negative effects of increased data trafic, simultaneously guaranteeing proper QoS provisioning and preserving end-user information anonymity, a feature of great importance to DSOs. 6.5 DESIGN GUIDELINES FOR SMART GRID IMPLEMENTATIONS This section presents several basic guidelines that should be followed when designing SG networks. The design steps that should be followed are: • Proper deinition of the SG requirements: The irst step when implementing SGs is to deine the relevant requirements according to the DSO speciications. Many DSOs have deployed backbone iber communication networks to monitor their crucial operations, especially in the high-voltage grid [53]; hence, in many cases, SNs are already deployed. On the other hand, DSOs have not yet deployed last-mile communication networks to exchange information with the various nodes of the transmission and distribution grids as well as with the end users. In these cases, the deployment of NANs and FANs is imperative. Therefore, the current stage of SG penetration in each case will determine the priorities in SGN deployment. Despite the speciic needs of DSOs, all SGNs should enable end users’ connectivity with the utilities’ premises through a reliable, secure, highly available, cost-effective, and resilient infrastructure. The need for reliability and security may lead to the selection of wired options for SGN implementation, whereas the necessity for optimal management of the cost–resilience trade-off 146 Smart Grids may lead to the installation of redundant infrastructures or sophisticated PLC transmission schemes and optimal network planning, as indicated in [54]. • SG services: After the requirements imposed by the DSOs have been speciied and the SGN type has been selected, the SG services must be deined. There are many reviews on QoS requirements of various SG services deployed over NANs, FANs, or SNs. As previously discussed in this chapter, SNs incorporate delay-sensitive applications. However, most of these networks have already been deployed and are properly managed. As far as DA is considered, the service requirements might vary signiicantly, since protection services impose strict delay constraints, whereas, on the other hand, metering services might be more tolerant to delays. In the case of NANs, the need for security, resilience, and anonymity calls for a robust and scalable SGN rather than a high-capacity one. • Communications network support: Directly related to the SG services is the selection of the SGN transmission medium. To support mission-critical applications, the implementation of one-hop communication technologies is imperative; hence, wireless cognitive approaches are preferable. On the other hand, PLC seems preferable for delay-tolerant applications. The services QoS speciied in the previous step usually leads to the deployment of hybrid networks. Techniques that offer interoperability or handle the performance trade-offs of the various transmission options are necessary. The SG interoperability standard [18] deines the interfaces between various SG entities, but does not handle issues related to data aggregation techniques or to the interoperability of various communications schemes. • Design of IT applications: Last but not least is the design of IT applications that constitute the interface of the SG nodes with the SG. As far as utilities are concerned, these applications should address the issue of big data; that is, they should effectively handle the data generated by the SG devices. In this context, many application-based platforms employing semantic repositories [55] or cloud-based approaches [56] are considered in the literature. Regardless of the application platform employed, the introduction of novel aggregation schemes leads to a signiicant reduction of unnecessary network trafic, enhancing the SGN performance. 6.6 CONCLUSIONS The transition toward smart power grids that support two-way lows of energy and information is imperative. In this framework, new services are introduced related to energy systems, communications platforms, and IT applications. The PLC transmission option is a strong candidate to support the various types of SGNs that should interoperate to offer reliable, end-to-end SG services: 1. BB-PLC gives rise to a high-capacity infrastructure capable of eficiently handling the trafic generated by CPNs. 2. Both NB-PLC and BB-PLC options may collect the trafic generated by CPNs and route this to the utilities’ back ofices, thus supporting NANs’ Role of PLC Technology in Smart Grid Communication Networks 147 communication. When combined with data aggregation techniques, PLC may satisfy an essential prerequisite of the SG, namely, preserving information anonymity. Also, when combined with injected power control, slot reuse in TDMA-based PLC transmission enhances network performance to meet the QoS constraints of time-critical applications. 3. NB-PLC seems appropriate to handle the low-rate trafic generated in heterogeneous FANs, especially when transmission of SG information from remote SGC nodes on the MV grid is required. When combined with the use of eficient network layer protocols, FANs may support the seamless integration of wireless and wired segments, increasing network availability. In conclusion, there is no speciic approach to designing SGNs. SG applications of variable QoS responding to different DSOs’ requirements may lead to different SG implementations. The need for resilience, security, reliability, and scalability, combined with the need for a cost-effective design, will determine how the various communications networks supporting the SG will be implemented. REFERENCES 1. European Commission, The EU climate and energy package 2020. Available at: http:// ec.europa.eu/clima/policies/package/index_en.htm. 2. Hochgraf, C., Tripathi, R., Herzberg, S., Smart grid charger for electric vehicles using existing cellular networks and SMS text messages, in Proceedings of the First IEEE International Conference on Smart Grid Communications (SmartGridComm), pp. 167–172, 4–6 October 2010, Gaithersburg, MD. 3. Bumiller, G., Lampe, L., Hrasnica, H., Power line communication networks for large-scale control and automation systems, IEEE Communications Magazine, 48(4), 106–113, 2010. 4. Raiei, M., Eftekhari, S.M., A practical smart metering using combination of power line communication (PLC) and WiFi protocols, in Proceedings of 17th Conference on Electrical Power Distribution Networks (EPDC), pp. 1–5, 2–3 May 2012, Tehran. 5. Aggarwal, A., Kunta, S., Verma, P.K., A proposed communications infrastructure for the smart grid, in Proceedings of Innovative Smart Grid Technologies (ISGT), pp. 1–5, 19–21 January 2010, Gaithersburg, MD. 6. IEEE P1901, IEEE standard for broadband over power line networks: Medium access control and physical layer speciications, IEEE Std 1901-2010, pp. 1–1586, December 2010. 7. ITU-T Recommendation G.9960, Uniied high-speed wire-line based home networking transceivers—Foundation, January 2010. 8. Patel, A., Aparicio, J., Tas, N., Loiacono, M., Rosca, J., Assessing communications technology options for smart grid applications, in Proceedings of IEEE Conference on Smart Grid Communications, pp. 126–131, 17–20 October 2011, Brussels. 9. Sarai, A., Tsiropoulos, G., Cottis, P., Hybrid wireless-broadband over power lines: A promising broadband solution in rural areas, IEEE Communications Magazine, 47(11), 140–147, 2009. 10. Fang, X., Misra, S., Xue, G., Yang, D., Smart grid—The new and improved power grid: A survey, IEEE Communications Surveys and Tutorials, 14(4), 944–980, 2012. 11. Lopes, A.J., Lezama, R., Pineda, R., Model-based systems engineering for smart grids as systems of systems, Procedia Computer Science, 6, 441–450, 2011. 148 Smart Grids 12. Metke, A.R., Ekl, R.L., Security technology for smart grid networks, IEEE Transactions on Smart Grid, 1(1), 99–107, 2010. 13. Sadeghi, S., Yaghmaee Moghddam, M.H., Bahekmat, M., Heydari Yazdi, A.S., Modeling of smart grid trafics using non-preemptive priority queues, in Proceedings of 2012 2nd Iranian Conference on Smart Grids (ICSG), pp. 1–4, 24–25 May 2012, Tehran, Iran. 14. National Institute of Standards and Technology, NIST framework and roadmap for smart grid interoperability standards, release 1.0, January 2010. Available at: http://www. nist.gov/public_affairs/releases/upload/smartgrid_interoperability_inal.pdf (accessed March 2013). 15. Hwang, T., Choi, M., Kang, S., Lee, I., Design of application-level reference models for micro energy grid in IT perspective, in Proceedings of 8th International Conference on Computing and Networking Technology (ICCNT), pp. 180–183, 27–29 August 2012, Gyeongju. 16. Lo, C., Ansari, N., The progressive smart grid system from both power and communications aspects, IEEE Communications Surveys and Tutorials, 14(3), 799–821, 2012. 17. Leeds, D.J., The smart grid in 2010: Market segments, applications and industry players, GTM Research, 2009. Available at: http://www.greentechmedia.com/research/report/ smart-grid-in-2010 (accessed March 2013). 18. IEEE Std 2030-2011. IEEE guide for smart grid interoperability of energy technology and information technology operation with the electric power system (EPS), end-use applications, and loads, IEEE Std 2030-2011, pp. 1–126, 10 September 2011. 19. Molderink, A., Bakker, V., Bosman, M.G.C., Hurink, J.L., Smit, G.J.M., Management and control of domestic smart grid technology, IEEE Transactions on Smart Grid, 1(2), 109–119, 2010. 20. IEC, Road vehicles—Vehicle to grid communication interface—Part 1: General information and use-case deinition, ISO/IEC 15118-1, 2010. 21. Deshpande, J.G., Kim, E., Thottan, M., Differentiated services QoS in smart grid communication networks, Bell Labs Technical Journal Wiley Subscription Services, 16(3), 61–81, 2011. 22. Khan, R.H., Khan, J.Y., A comprehensive review of the application characteristics and trafic requirements of a smart grid communications network, Computer Networks, 57(3), 825–845, 2013. 23. Kansal, P., Bose, A., Bandwidth and latency requirements for smart transmission grid applications, IEEE Transactions on Smart Grid, 3(3), 1344–1352, 2012. 24. IEEE Std 1646-2004. IEEE standard communication delivery time performance requirements for electric power substation automation, IEEE Std 1646-2004, pp. 0_1–24, 2005. 25. Usman, A., Shami, S.D., Evolution of communication technologies for smart grid applications, Renewable and Sustainable Energy Reviews, 19, 191–199, 2013. 26. Galli, S., Logvinov, O., Recent developments in the standardization of power line communications within the IEEE, IEEE Communications Magazine, 46(7), 64–71, 2008. 27. Rahman, M.M., Hong, C.S., Lee, S., Lee, J., Razzaque, M.A., Kim, J.H., Medium access control for power line communications: An overview of the IEEE 1901 and ITU-T G.hn standards, IEEE Communications Magazine, 49(6), 183–191, 2011. 28. Galli, S., Scaglione, A., Wang, Z., For the grid and through the grid: The role of power line communications in the smart grid, Proceedings of the IEEE, 99(6), 998–1027, 2011. 29. Galli, S., Kurobe, A., Ohura, M., The inter-PHY protocol (IPP): A simple coexistence protocol for shared media, in Proceedings of IEEE International Symposium on Power Line Communications and Its Applications, ISPLC, pp. 194–200, 29 March–1 April 2009, Dresden, Germany. 30. Oksman, V., Galli, S., G.hn: The new ITU-T home networking standard, IEEE Communications Magazine, 47(10), 138–145, 2009. Role of PLC Technology in Smart Grid Communication Networks 149 31. Galli, S., Scaglione, A., Wang, Z., Power line communications and the smart grid, in Proceedings of First IEEE International Conference on Smart Grid Communications (SmartGridComm), pp. 303–308, 4–6 October 2010, Gaithersburg, MD. 32. IEEE, Low frequency narrow-band power line communications. Available at: http:// grouper.ieee.org/groups/1901/2/ (accessed January 2013). 33. Wang, Z., Wang, Y., Wang, J., Overlapping clustering routing algorithm based on L-PLC meter reading system, in Proceedings of IEEE International Conference on Automation and Logistics, ICAL ‘09, pp. 1350–1355, 5–7 August 2009, Shenyang, China. 34. Liu, X., Wang, W., Zheng, J., Hai, T., Zhang, L., Liu, B., Tangential connection clustering routing algorithm for L-PLC based AMR systems, in Proceedings of 7th International Power Electronics and Motion Control Conference (IPEMC), vol. 4, pp. 2932–2936, 2–5 June 2012, Harbin, China. 35. Yu, R., Zhang, Y., Gjessing, S., Yuen, C., Xie, S., Guizani, M., Cognitive radio based hierarchical communications infrastructure for smart grid, IEEE Network, 25(5), 6–14, 2011. 36. Vineeta, Thathagar, J.K., Cognitive radio communication architecture in smart grid reconigurability, in Proceedings of 1st International Conference on Emerging Technology Trends in Electronics, Communication and Networking (ET2ECN), pp. 1–6, 19–21 December 2012, Surat, Gujarat, India. 37. Yu, R., Zhang, Y., Chen, Y., Hybrid spectrum access in cognitive neighborhood area networks in the smart grid, in IEEE Wireless Communications and Networking Conference (WCNC), pp. 1478–1483, 1–4 April 2012, Shanghai. 38. Lopez, G., Moura, P.S., Custodio, V., Moreno, J.I., Modeling the neighborhood area networks of the smart grid, in Proceedings of IEEE International Conference on Communications (ICC), pp. 3357–3361, 10–15 June 2012, Ottawa, ON. 39. Chauvenet, C., Tourancheau, B., Genon-Catalot, D., Goudet, P.-E., Pouillot, M., A communication stack over PLC for multi physical layer IPv6 networking, in Proceedings of First IEEE International Conference on Smart Grid Communications (SmartGridComm), pp. 250–255, 4–6 October 2010, Gaithersburg, MD. 40. Dawson-Haggerty, S., Tavakoli, A., Culler, D., Hydro: A hybrid routing protocol for low-power and lossy networks, in Proceedings of First IEEE International Conference on Smart Grid Communications (SmartGridComm), pp. 268–273, 4–6 October 2010, Gaithersburg, MD. 41. Ko, J., Terzis, A., Dawson-Haggerty, S., Culler, D.E., Hui, J.W., Levis, P., Connecting low-power and lossy networks to the internet, IEEE Communications Magazine, 49(4), 96–101, 2011. 42. Kapareliotis, E.S., Drakakis, K.E., Dimitriadis, H.P.K., Capsalis, C.N., Fault recognition on power networks via SNR analysis, IEEE Transactions on Power Delivery, 24(4), 2428–2433, 2009. 43. Qureshi, M., Raza, A., Kumar, D., Kim, S.S., Song, U.S., Park, M.W., Jang, H.S., Yang, H.S., Park, B.S., A survey of communication network paradigms for substation automation, in Proceedings of IEEE International Symposium on Power Line Communications and Its Applications, ISPLC, pp. 310–315, 2–4 April 2008, Jeju, Jeju Island. 44. Krishnamachari, B., Estrin, D., Wicker, S., The impact of data aggregation in wireless sensor networks, in Proceedings of the 22nd International Conference on Distributed Computing Systems Workshops, pp. 575–578, July 2002, Vienna, Austria. 45. Solis, I., Obraczka, K., In-network aggregation trade-offs for data collection in wireless sensor networks, International Journal of Sensor Networks, 1(3–4), 200–212, 2006. 46. Lu, C., Liang, X., Li, X., Lin, X., Shen, X., EPPA: An eficient and privacy-preserving aggregation scheme for secure smart grid communications, IEEE Transactions on Parallel and Distributed Systems, 23(9), 1621–1631, 2012. 150 Smart Grids 47. Gungor, V.C., Lu, B., Hancke, G.P., Opportunities and challenges of wireless sensor networks in smart grid, IEEE Transactions on Industrial Electronics, 57(10), 3557–3564, 2010. 48. Arein, A.S., Kashem Mia, M.A., NP-completeness of the minimum edge-ranking spanning tree problem on series-parallel graphs, in Proceedings of 10th International Conference on Computer and Information Technology, pp. 1–4, 27–29 December 2007, Dhaka, Bangladesh. 49. Incel, O.D., Ghosh, A., Krishnamachari, B., Chintalapudi, K., Fast data collection in tree-based wireless sensor networks, IEEE Transactions on Mobile Computing, 11(1), 86–99, 2012. 50. Kim, M.S., Kim, J., Kim, J., Yoo, Y., Design and implementation of MAC protocol for SmartGrid HAN environment, in Proceedings of IEEE 11th International Conference on Computer and Information Technology (CIT), pp. 212–217, 31 August–2 September 2011, Pafos. 51. Bu, S., Yu, F.R., Liu, P.X., Zhang, P., Distributed scheduling in smart grid communications with dynamic power demands and intermittent renewable energy resources, in Proceedings of IEEE International Conference on Communications Workshops (ICC), pp. 1–5, June 2011, Kyoto, Japan. 52. Moscibroda, T., The worst-case capacity of wireless sensor networks, in Proceedings of 6th International Symposium on Information Processing in Sensor Networks, IPSN, pp. 1–10, 25–27 April 2007, Cambridge, MA. 53. Livieratos, S., Vogiatzaki, E., Cottis, P., A generic framework for the evaluation of the beneits expected from the smart grid, Energies, 6, 988–1008, 2013. 54. Canale, S., Di Giorgio, A., Lanna, A., Mercurio, A., Panili, M., Pietrabissa, A., Optimal planning and routing in medium voltage power line communications networks, IEEE Transactions on Smart Grid, (99), 1–9, 2012. 55. Pena, A., Penya, Y.K., Distributed semantic repositories in smart grids, in Proceedings of 9th IEEE International Conference on Industrial Informatics (INDIN), pp. 721–726, 26–29 July 2011, Lisbon, Portugal. 56. Simmhan, Y., Prasanna, V., Aman, S., Kumbhare, A., Liu, R., Stevens, S., Zhao, Q., Cloud-based software platform for big data analytics in smart grids, Computing in Science and Engineering, 15(4), 38–47, 2013. 7 Power Grid Network Analysis for Smart Grid Applications Zhifang Wang, Anna Scaglione, and Robert J. Thomas CONTENTS 7.1 7.2 7.3 Introduction .................................................................................................. 152 System Model ............................................................................................... 153 Topology Measures of Power Grids.............................................................. 154 7.3.1 Topological Metrics .......................................................................... 156 7.3.2 Small-World Wiring ......................................................................... 156 7.3.3 The Nodal Degree Distribution ........................................................ 158 7.3.4 The Graph Spectrum and the Scaling of Connectivity .................... 162 7.4 Line Impedance Distribution ........................................................................ 165 7.5 A Model to Generate RT Test Networks ...................................................... 168 7.5.1 clusterSmallWorld Subnetwork ........................................................ 169 7.5.2 Lattice Connections .......................................................................... 170 7.5.3 Assignment of Line Impedances ...................................................... 170 7.6 Topological Analysis for the Distribution Networks .................................... 171 7.6.1 Structure of the Distribution Network .............................................. 171 7.6.2 Graph Theoretic Analysis of a Sample MV Distribution Network .....172 7.6.3 The PMF of the Node Degrees in the Sample 396-Node MV Distribution Network ........................................................................ 174 7.6.4 LV Distribution Network .................................................................. 176 7.7 Summary ...................................................................................................... 176 References .............................................................................................................. 176 Smart grid is an umbrella term for the modernization of electricity delivery systems with enhanced monitoring, analysis, control, and communication capabilities in order to improve the eficiency, reliability, economics, and sustainability of electricity services. The modernization happens at both the high-voltage (HV) transmission and the medium- or low-voltage (MV/LV) distribution grids and has a diverse set of goals, including facilitating greater competition between providers, encouraging greater use of alternative energy sources, implementing the automation 151 152 Smart Grids and monitoring capabilities needed for bulk transmission, and enabling the use of market forces to drive energy conservation. Many of the existing technologies already used by electric utilities will be taking full advantage of the smart grid, but additional communication and control capabilities will also be established to optimize the operation of the entire electrical grid. The smart grid is also positioned to adopt new technologies, such as smart metering, synchronous phasor measurement units (PMUs), plug-in hybrid electric vehicles (PHEVs), various forms of distributed generation, solar and wind energy, electric load management systems, distribution automation, and many more. In order to design an eficient communication scheme and examine the eficiency of any networked control architecture in smart grid applications, we need to statistically characterize its information source, namely, the power grid itself. Investigating the statistical properties of power grids has the immediate beneit of providing a natural simulation platform, producing a large number of power grid test cases with realistic topologies, with scalable network size, and with realistic electrical parameter settings. The second beneit is that one can start analyzing the performance of decentralized control algorithms over information networks whose topology matches that of the underlying power network, and use network scientiic approaches to determine analytically whether these architectures would scale well. 7.1 INTRODUCTION This chapter provides a comprehensive study on the topological and electrical characteristics of a power grid transmission network based on a number of synthetic and real-world power systems. Besides some basic graph theoretic metrics, it especially considers the following characteristics: • Small-World Properties: The transmission network of a power grid is sparsely connected and manifests salient small-world properties that are characterized by the larger clustering coeficient and much smaller average shortest path compared with a random graph network with the same network size [1]. However, it is found that the power grid assumes a different kind of small-world topology than that of the Watts–Strogatz small-world model. • Nodal Degree Distribution: This metric relects how the nodes in a power grid network connect with neighbors and closely relates to the network topology robustness under node removal [2,3]. Although most literature assumes that power grids have an exponential (or geometric) nodal degree distribution [4,5], it is found that the empirical distribution estimated from real-world power systems clearly deviates from an exponential distribution, especially in the range of small node degrees. The probability generation function is applied to analyzing the node degree distribution and estimating the parameters for it. • Graph Spectrum and Connectivity Scaling Property: These two metrics associate with the eigenvalues of the adjacency matrix and the Laplacian matrix, respectively, and they carry important topology characteristics of a network. Power Grid Network Analysis for Smart Grid Applications 153 • Distribution of Line Impedance: The line impedance dominates the electrical characteristics of a power grid network; hence, a study on its distribution is important and necessary. It is found that the distribution of line impedance is heavy-tailed and can be captured quite accurately by a clipped double Pareto lognormal (DPLN) distribution. This chapter introduces an algorithm that is able to generate random topology power grids featuring the same topology and electrical characteristics as those found in the real data. Finally, it also reports recent studies on the topological and electrical characteristics of a sample MV power distribution network. 7.2 SYSTEM MODEL The power network dynamics are coupled by the network equation YU = I , (7.1) where U and I represent the complex phasor vectors of bus voltages and injected currents; and Y is the network admittance matrix, which is determined by not only the connecting topology but also its electrical parameters. Given a network with n nodes and m links (which may also be referred to as buses and branches or lines in power grid analysis, or vertices and edges in graph theory and network analysis), each link l = (s, t) between nodes s and t has a line impedance zpr(l) = r(l) + jx(l), where r(l) is the resistance and x(l) the reactance. Usually for an HV transmission network the reactance dominates, that is, x(l) ≫ r(l). The line admittance is obtained from the inverse of its impedance, that is, ypr(l) = g(l) + jb(l) = 1/zpr(l). Assume that a unit current lows along the link l = (s, t) from node s to t; then the resulting voltage difference between the ends of the link equals Δu = U(s)−U(t) = zpr(l), or equivalently Δu = 1/ypr(l). Therefore, zpr(l) can be interpreted as the electrical distance between nodes s and t, and ypr(l) relects the coupling strength between the two end nodes. Deine the line-node incidence matrix A: = (Al,k)m × n as: Al,s = 1, Al,t = −1 if the lth link is between nodes s and t with arbitrary directions; and Al,k = 0 otherwise. Then the network admittance matrix Y: = (Ys,t)n × n can be obtained as Y = AT diag ( y pr ) A, (7.2) with entries as follows: Y ( s, t ) = − y pr ( s, t ) , y pr ( s, t ) , Y ( s, s ) = t ≠s Y ( s, t ) = 0, if link s − t exists, for t ≠ s ∑ (7.3) otherwise. 154 Smart Grids The network Laplacian L: = (Ls,t)n × n can then be obtained as L = ATA, with the entries L ( s, t ) = −1, L ( s, s ) = k, L ( s, t ) = 0, if link s − t exists, for t ≠ s with k = deg ( s ) = Σ t ≠ s − L ( s, t ) otherwise. (7.4) A close comparison of the matrix structures of L and Y uncovers some interesting analogies. The Laplacian matrix L fully describes the topology of a network, while the network admittance matrix Y contains information not only about the system topology, but also about its electrical coupling. The off-diagonal entries of Y, Ys,t equal the line admittance of the link between node s and t (with a minus sign), whose magnitude relects the coupling strength between the two nodes. The diagonal entries of the Laplacian L represent the total number of links connecting each node with the rest of the network, while a diagonal entry of Y represents the total coupling capability of one node with the rest of the network. Therefore, the network admittance matrix Y can be viewed as a complex-weighted Laplacian; and the Laplacian L can be equivalent to a “lat” network admittance matrix, which assumes all the links in the network have the same line impedance (with a common proportional factor). 7.3 TOPOLOGY MEASURES OF POWER GRIDS First, we look at some basic topological metrics for a power grid network: the network size n, the total number of links m, the average nodal degree 〈k〉, the average shortest path length counted in hops 〈l〉, and the Pearson coeficient. All these metrics can be derived from the Laplacian matrix: • • • • The total number of links is m = 1 2 Σi L (i, i ). The average nodal degree is 〈 k 〉 = 1 n Σi L (i, i ). The network adjacency matrix can be obtained as Madj = −L + diag{L}. The average shortest path 〈l〉 measured in hops can be calculated with Dijkstra’s algorithm [6] based on Madj. Deine the nodal degree vector as k = [k1, k2, …, kn] = diag(L), and k is the average degree of a node seen at the end of a randomly selected link (i,j), that is, k = ( 2m ) −1 ∑ ( k + k ) = ( 2m ) ∑ k −1 i 2 i j (i, j ) i = 〈k 2 〉 . 〈k〉 (7.5) Thus, we can compute the Pearson coeficient of node degrees of the network as ρ= ∑ ∑ (k − k )(k − k ) (k − k ) (k − k ) i j (i, j ) (i, j ) 2 i j 2 , (7.6) 155 Power Grid Network Analysis for Smart Grid Applications which is a measure of the correlation of node degrees in the network [7]. It has been observed that the Pearson coeficient for some kinds of networks is consistently positive, while for other kinds it is negative [8]. Therefore, some researchers proposed using the Pearson coeficient to differentiate technological networks from social networks [9]. However, it is found that the Pearson coeficient of power grids does not have restrictive characteristics, but ranges over a wide interval from negative to positive, which can be veriied by the metric evaluation results listed in Table 7.1. Power grids have been found to have the salient features of small-world graphs (see the work by Watts and Strogatz [1]). That is, while the vast majority of links are similar to those of a regular lattice, with limited near-neighbor connectivity, a few links connect across the network. These bridging links signiicantly shorten the path length that connects every two nodes and critically increase the connectivity of the network. At the same time, their scarcity puts the connectivity at risk in the case of link failure for one of these critical bridges. The characterizing measure to distinguish a small-world network is called the clustering coeficient, which assesses the degree to which nodes tend to cluster together. A small-world network usually has a clustering coeficient signiicantly higher than that of a random graph network, given the comparable network size and total number of edges. The random graph network mentioned here refers to the network model deined by Erdös and Rényi, with n labeled nodes connected by m edges that are chosen uniformly randomly from the n(n − 1)/2 possible edges [10]. The clustering coeficient is deined as the average of the clustering coeficients for all nodes [1]: C (G ) = 1 n n ∑C , (7.7) i i =1 with Ci being the node clustering coeficient as Ci = λG (i)/τG (i), where λG (i) is the number of edges between the neighbors of node i and τG (i) the total number of edges that could possibly exist among the neighbors of node i. For undirected graphs, obviously τG (i) = ki(ki − 1)/2, given ki is the node degree. As pointed out in [11], the TABLE 7.1 Topological Metrics of the IEEE and Real-World Power Grid Networks IEEE-30 IEEE-57 IEEE-118 IEEE-300 NYISO WSCC (n, m) 〈I〉 〈k〉 ρ C(G) C(R) (30, 41) (57, 78) (118, 179) (300, 409) (2935, 6567) (4941, 6594) 3.31 4.95 6.31 9.94 16.43 18.70 2.73 2.74 3.03 2.73 4.47 2.67 −0.0868 0.2432 −0.1526 −0.2206 0.4593 0.0035 0.2348 0.1222 0.1651 0.0856 0.2134 0.0801 0.094253 0.048872 0.025931 0.009119 0.001525 0.000540 156 Smart Grids clustering coeficient for a random graph network theoretically equals the probability of randomly selecting links from all possible links. That is C ( R) = 7.3.1 m 〈k〉 . = n ( n − 1) 2 n − 1 (7.8) toPologIcal metrIcs The topology metrics evaluated for the Institute of Electrical and Electronics Engineers (IEEE) model systems, the New York Independent System Operator (NYISO), and the Western System Coordinating Council (WSCC) systems are shown in Table 7.1. Especially, we show the clustering coeficients of the IEEE, NYISO, and WSCC systems compared with the random graph networks with the same network size and the same total number of links. The former is denoted as C(G), and the latter as C(R). For the reader’s reference, the IEEE 30, 57, and 118 bus systems represent different parts of the American Electric Power System in the midwestern United States; the IEEE 300 bus system is synthesized from the New England power system. More information can be obtained from [12]. The WSCC grid is the interconnected system of transmission lines spanning the western United States plus parts of Canada and Mexico, and the NYISO system represents the New York state bulk electricity grid. From Table 7.1, one can observe two interesting properties pertaining to the topology of the grid: (a) The connectivity is very sparse, since the average nodal degree does not scale up as the network size increases. Instead, the average nodal degree falls into a very restricted range, which is related to the particular geographical area to which the network belongs, rather than the network size. For example, power grids in the western United States have an average nodal degree between 2.5 and 3, while in the northeastern United States the power grids are a little denser, with 〈k〉 around 4.5. (b) It is not accurate to attribute “small-world” features to power grids of any size. It is true that the samples we studied have a clustering coeficient signiicantly larger than that of a random graph network of comparable size. However, as we will discuss next, while being similarly sparse, power grids have better connectivity scaling laws than the small-world graphs proposed in [1]. 7.3.2 small-World WIrIng The Watts–Strogatz small-world model is generated starting from a regular ring lattice, then, using a small probability, rewiring some local links to an arbitrary node chosen uniformly at random from the entire network (to make it a smallworld topology). A tool to visually highlight small-world topologies is the Kirk graph, which was proposed by Kirk (2007) [13]. It is a simple but effective way to show a network topology. First, node numbers are assigned according to physical nodal adjacency, that is, physically closely located nodes are given close numbers; then all the nodes are sequentially and evenly spread around a circle, and links between nodes are drawn as straight lines inside the circle. Figure 7.1 shows 157 Power Grid Network Analysis for Smart Grid Applications 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 (a) 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 40 41 42 43 44 45 46 57 56 55 54 53 52 51 50 49 47 48 18 2019 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 3940 41 (c) 17 16 19 18 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 (b) 15 14 13 1211 10 9 8 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 42 43 7 44 45 46 57 56 55 54 53 52 51 50 49 47 48 6 5 4 3 2 1 42 43 44 45 46 57 56 55 54 53 52 51 50 49 4748 FIGURE 7.1 (a–c) Topology comparison using Kirk plot. three representative network topologies, using Kirk graphs, of an Erdös–Rényi rand-graph network, of a Watts–Strogatz small-world network, and of a real power grid—the IEEE-57 network. The three networks have the same network size and almost the same total number of links. The topology difference between the Erdös–Rényi network and the IEEE-57 network can be easily noticed, in contrast to a good degree of similarity between the Watts–Strogatz small-world model and the test power network. However, the rewiring of the test power grid is not independent, as it is in the Watts–Strogatz small-world model; instead, long hauls appear over clusters of nodes. This property is visually noticeable in the IEEE 57 network and differentiates it from the small-world graph, in which these links are chosen independently. This fact has an intuitive explanation: long hauls require having a right of way to deploy a long connection, and it is highly likely that the long wires will reuse part of this space. These physical and economic constraints inevitably affect the structure of the topology. Besides, there is more than meets the eye. The Watts– Strogatz small-world model has a scaling property that cannot be validated by 158 Smart Grids power grid topologies, precisely because the average nodal degree of a power network is almost invariant with the size of the network. Given a network size with its speciied average nodal degree, the model fails to produce a connected grid topology for reasonably large network sizes with realistic power grid degree distributions. The main reason for the poor scaling lies in the fact that, in order to produce a sparse but connected topology, the Watts–Strogatz small-world model requires [1,11] 1 ≪ ln ( n ) ≪ 〈 k 〉 ≪ n , (7.9) or, equivalently, 〈k〉 = n = e〈k〉. On the other hand, real-world power grids have a very low average nodal degree, with 〈k〉 = 2 – 5, regardless of the network size. This will limit the network size to much smaller than 150 in order to produce a connected topology using the model. However, in the real world, large power grids are connected even if they are much sparser than is required by the Watts–Strogatz small-world model. 7.3.3 the nodal degree dIstrIButIon Now we examine the empirical distribution of nodal degrees in the available realworld power grids. Figure 7.3a shows the histogram probability mass function (PMF) in log scale for the nodal degrees of the NYISO system. If the PMF curve approximates a straight line in the semilogarithmic plot (i.e., shown as log(p(k)) vs. k), it implies an exponential tail analogous to that of the geometric distribution. However, it is also noticed that for the range of small node degrees, that is, when k ≤ 3, the empirical PMF curve clearly deviates from that of a geometric distribution. The probability generation function (PGF) can be utilized to analyze the node degree distribution in power grids. The PGF of a random variable X is deined as GX ( z) = Σ k Pr( x = k ) z k. Given a sample dataset of X with size N, its PGF can also be estimated from the mean of z X, because ( ) E zX = = = 1 n ∑z 1 n ∑n ≈ x ( x =k ) zk k ∑ k x n( x = k ) k z n ∑ Pr ( x =k ) zk , (7.10) k where n(x = k) denotes the total number of the data items equal to k. Due to lim n →∞ n( x = k ) n = Pr( x = k ), we can have E(z X) ≈ GX(z) with a large enough data size. 159 Power Grid Network Analysis for Smart Grid Applications If a random variable can be expressed as the sum of two independent random variables, its PMF is the convolution of the PMFs of the component variables, and its PGF is the product of those of the component variables. That is, X = X1 + X 2 Pr( X = k ) = Pr( X1 = k ) ⊗ Pr( X2 = k ) (7.11) GX ( z ) = GX1 ( z ) GX2 ( z ) . It is found that the node degree distribution in power grids can be very well approximated by the sum of two independent random variables, that is, K = G + D, (7.12) where G is a truncated geometric with the threshold of kmax: (1 − p ) p Pr(G = k ) = k i ∑ i=0 (1 − p ) p k max (1 − p ) p , k +1 1 − (1 − p ) k = k = 0,1, 2, … , kmax, max (7.13) with the PGF as kmax (1 − p ) pz k k +1 1 − (1 − p ) ∑ G (z) = G = k k =0 max ( ( p 1 − (1 − p ) z ( 1 − (1 − p ) kmax +1 ) kmax +1 ) (7.14) ) (1 − (1 − p ) z ) and D is an irregular discrete { p1, p2 , … , pkt }, with Pr( D= k ) = pk , k = 1, 2, … , kt whose PGF is GD ( z ) = p1z + p2 z 2 + p3 z 3 + ⋯ + pkt z kt . Therefore, the PMF of K is Pr(K=k) = Pr(G=k) ⊗ Pr(D=k), and the PGF of K can be written as GK ( z ) = ( ( p 1 − (1 − p ) z ( 1 − (1 − p ) ) kmax +1 kmax +1 )∑ kt i =1 pi z i ) (1 − (1 − p ) z ) . (7.15) Equation 7.15 indicates that the PGF G K(z) has kmax zeros evenly distributed around a circle of radius of 1/(1 − p) that are introduced by the truncation of the geometric G 160 Smart Grids 1.5 1.5 1.0 1.0 0.5 0.5 0 0 Y Y (because the zero at 1/(1 − p) has been neutralized by the denominator (1 − (1 − p)z)) and has kt zeros introduced by the irregular discrete D with { p1, p2 , … , pkt}. Figure 7.2 shows the contour plots of PGF of node degrees for different groups of buses in the NYISO system. Three interesting discoveries are worthy of note: (a) clearly each plot contains evenly distributed zeros around a circle, which indicates a truncated geometric component; (b) besides the zeros around the circle, most contour plots also have a small number of off-circle zeros, which come from an embedding irregular discrete component; (c) the contour plot for each group of nodes has zeros with a similar pattern but different positions. This implies that each group of node degrees has similar distribution functions, but with different coeficients. Therefore, it is necessary and reasonable to characterize the node degree distribution according to the node types. Otherwise, if the node degrees were to be aggregated into one single group, as in Figure 7.2a, some important characteristics of a subgroup of node degrees would be concealed (e.g., comparing a and b–d in Figure 7.2). −0.5 −0.5 −1.0 −1.0 −1.5 −1 (a) 0 X 1 −1.5 1.5 1.0 1.0 0.5 0.5 0 0 −0.5 −0.5 −1.0 −1.0 Y 1.5 −1.5 (c) −1 0 X −1 (b) 1 −1.5 −1.5 (d) −1.0 −0.5 0 X 0 X 1 0.5 1.0 1.5 FIGURE 7.2 Contour plot of E(z X) of node degrees for different groups of buses in the NYISO system: (a) all buses; (b) generation buses; (c) load buses; (d) connection buses. The zeros are marked by plus signs. 161 Power Grid Network Analysis for Smart Grid Applications Based on the contour plots, one can easily locate the zeros in PGF, and further determine the coeficients of the corresponding distribution functions. The estimated coeficients for each group of nodes in the NYISO and all the nodes in the WSCC system are listed in Table 7.2. Because the WSCC system data we have only contains unweighted raw data without distinguishing node types, its node degree distribution is only analyzed for one aggregate group. Figure 7.3 compares the PMF with estimated coeficients and the empirical PMF of the NYISO system, and shows that the former matches the latter with quite good approximation. The TABLE 7.2 Estimated Coeficients of the Node Degree Probability Density Functions of the NYISO and WSCC Systems Node Groups All Generation Load Connection All WSCC kmax Max (k) kt ρ 34 36 26 18 16 37 37 29 21 19 3 1 3 3 3 0.2269 0.1863 0.2423 0.4006 0.4084 , pkt} 0.4875, 0.2700, 0.2425 1.000 0.0455, 0.4675, 0.4870 0.0393, 0.4442, 0.5165 0.3545, 0.4499, 0.1956 100 100 Empirical PMF Fitting PMF Empirical PMF Fitting PMF 10−1 10 10−2 log (PMF) log (PMF) { p1, p2 , 10−3 10−4 −1 10−2 10−3 10−5 10−6 0 10 (a) 20 k-node degree 30 10−4 40 Empirical PMF Fitting PMF 20 k-node degree 30 40 Empirical PMF Fitting PMF 10−1 10−2 log (PMF) log (PMF) 10−1 10−3 10−2 10−3 10−4 10−4 (c) 10 100 100 10−5 0 (b) 0 5 10 15 20 k-node degree 25 30 10−5 (d) 0 5 10 15 k-node degree 20 25 FIGURE 7.3 Comparing the empirical and itted PMF of node degrees in the NYISO system: (a) all buses; (b) generation buses; (c) load buses; (d) connection buses. 162 Smart Grids results from both systems have validated our assumption of node degree distribution in power grids: it can be expressed as the sum of a truncated geometric random variable and an irregular discrete random variable. The results also demonstrated the effectiveness of the proposed method of analyzing node degree distribution by using the PGF. 7.3.4 the graPh sPectrum and the scalIng oF connectIvIty The eigenvalues of the adjacency matrix Madj and the Laplacian matrix L carry important topological features of a network. The two matrices are in fact exchangeable, and either of them fully describes a network topology. The set of eigenvalues of its adjacency matrix Madj is called the spectrum of a graph. Graph spectral density is deined as ρ (λ) = 1 n n ∑ δ (λ − λ ) , j (7.16) j =1 with {λj, j = 1, 2, …, n} forming the graph spectrum of the network. The kth moment of ρ(λ) represents the average number of k-hop paths returning to the same node in the graph [11]. However, please note that these paths can contain nodes which are already visited. The graph spectral density of network topology from different categories has distinctively different patterns [11]. An Erdös–Rényi random graph network G(n;p), given a large network size n and a nontrivial link selection probability p(n) = cn−z, with z < 1 and c as a constant regardless of network size, has its normalized spectral density converging to a semicircular distribution as n → ∞. This is known as Wigner’s semicircle law [14,15]. That is 2 4 − ( λ λ 0 ) ρ (λ) = 2πλ 0 0 if | λ |< 2λ 0 (7.17) otherwise with λ 0 = np(1 − p) and p = m n(n − 1) 2 = 〈 k 〉 n − 1, where m is the total number of links and 〈k〉 is the average node degree. And ρɶ = 2πλ 0ρ(λ) = 4 − λɶ 2 with λɶ = λ / λ 0 depicts a semicircle around the origin with a radius of 2. Next, we refer to spectrum as the set of eigenvalues and to spectral density as its density of values. Figure 7.4 shows the spectral density of a ring lattice (a), of two real-world power grids (b) and (c), and of a power grid generated from the proposed random topology (RT)-nested-Smallworld model (d), compared with the semicircle law corresponding to random graphs with the same network size and number of links. The plots demonstrate that power grid graphs have a very distinctive spectral density, far from that of regular or completely random networks; it also shows that the proposed model RT-nested-Smallworld (see Section 7.5) is in good agreement with the real data in terms of graph spectral density. 163 Power Grid Network Analysis for Smart Grid Applications 20 4 20 18 16 14 12 10 8 6 4 2 0 −2 (b) −4 4 20 18 16 14 12 10 8 6 4 2 0 −2 (d) −4 Lattice−1-D: 2935 Random graph 15 10 5 0 (a) −4 0 2 NYISO: 2935 Random graph network −2 0 Normalized graph spectral density 20 18 16 14 12 10 8 6 4 2 0 −2 (c) −4 −2 (e) 2 RT-nestedSW: 2988 Random graph network −2 0 2 4 WSCC: 4941 Random graph network −2 0 2 4 9 8 7 6 5 4 3 2 1 0 −4 −3 −2 −1 0 ~ λ 1 2 3 4 FIGURE 7.4 Normalized graph spectral density of different networks, ρɶ (λ ) versus λɶ : the dotted line of semicircle represents the graph spectral density of random graph networks; (a) a ring lattice; (b) RT-nested-Smallworld; (c) NYISO; (d) WSCC; (e) a 396-node MV distribution network. Another important measure is the second smallest eigenvalue of the Laplacian matrix, λ2(L), called algebraic connectivity. This measure is sometimes termed the Fiedler eigenvalue, because it was irst introduced by Fiedler (1989) [16]. How well a network is connected and how fast information data can be shared across the network is relected by λ2(L). The smallest eigenvalue of the Laplacian is always zero, that is, λ1(L) ≡ 0, and the number of times 0 appears as an eigenvalue in the 164 Smart Grids Laplacian is the total number of connected components in the network. The eigenvalue λ2(L) is greater than 0 if and only if the network is a connected graph. If the algebraic connectivity λ2(L) is close to zero, the network is close to being disconnected. Otherwise, if (1/n)λ2(L) tends to be 1, with n as the network size, the network tends to be fully connected. Table 7.3 shows the algebraic connectivity of IEEE model systems and the NYISO system. Figure 7.5 plots the connectivity scaling curve of power grid versus network size and compares it with that of one-dimensional (1-D) and two-dimensional (2-D) lattices. The 1-D lattice is a ring-structured topology, with nodes connected to most TABLE 7.3 Algebraic Connectivity of the IEEE and Real-World Power Networks IEEE-30 IEEE-57 IEEE-118 IEEE-300 NYISO WSCC n λ2(L) 30 57 118 300 2935 4941 0.21213 0.088223 0.027132 0.0093838 0.0014215 0.00075921 101 100 λ2 10−1 10−2 n−0.976 n−1.376 n−0.933 10−3 n−1.0604 n−2 10−4 n−2 −5 10 101 102 103 n (network size) 104 FIGURE 7.5 Connectivity scaling curve versus network size: 2-D lattice with k = 4 (x-dotted line); 2-D lattice with k = 2 (Δ-dotted line); 1-D lattice with k = 4 (▫-dotted line); 1-D lattice with k = 2 (∇-dotted line); nested-Smallworld RT with subnetwork size = 30 (small ▫); nestedSmallworld RT with subnetwork-size = 300 (⬨); power grids (star). Power Grid Network Analysis for Smart Grid Applications 165 adjacent neighbors on both sides. The 2-D lattice is a regular 2-D meshed grid with each boundary side merging with the other side and each node connected to the most adjacent neighbors around it. For a 1-D lattice, its connectivity scales as λ2(L) ∝ n−2; for a 2-D lattice, its connectivity grows as λ2(L) ∝ n−1; interestingly, for power grids, connectivity grows as λ2(L) ∝ n−1.376~−1.0604, lying between those of the 1-D lattice and the 2-D lattice. In trying to it the random wiring of a power network, we postulate a new possible model, RT-nested-Smallworld, in Section 7.5. The model results from nesting several “small-world” subnetworks, whose size is likely to produce a connected topology with a realistic degree distribution, into a regular lattice again. More details can be found in Section 7.5. The intuition guiding this modeling is that the network would have produced connected topologies with a connectivity that was intermediate between the 1-D and 2-D lattices. Figure 7.5 shows an excellent match between the algebraic connectivity of this type of model and the real data. 7.4 LINE IMPEDANCE DISTRIBUTION This chapter not only examines the network topology of a power grid, but also attempts to reproduce its electrical characteristics accurately, because the key element of a power grid network is its network admittance matrix Y, which associates both with the connecting topology and with its line impedances. The transmission line impedance in a power grid is, in fact, a complex number. When it comes to the statistical analysis, we study the distribution of the magnitude of Zpr. The line impedance can be represented as Zpr = R + jX, where R is the resistance and X is the reactance. Usually X is the dominant component, whereas R only takes a trivial value, which in many cases can even be neglected. Therefore, one can easily reconstruct the complex value of Zpr given its magnitude. The empirical data on line impedances of power grids are taken from IEEE model systems and the NYISO system. The irst clear observation from the empirical histogram probability density distribution (PDF) is that the distribution of the line impedances is heavy-tailed. The candidate distribution functions include gamma, generalized Pareto (GP), lognormal, DPLN, and two new distributions that we call lognormal-clip and DPLN-clip, which will be explained later. In searching among the heavy-tailed distributions for a it, the parameters of the candidate distribution can be estimated via the maximum likelihood (ML) criterion from the data, and the appropriately modiied Kolmogorov–Smirnov (K-S) test can be used to check whether the hypothesized distribution is a good it. The DPLN distribution was introduced by Reed and Jorgensen (2004) [17], and has proved to be very useful to model the size distributions of various phenomena, such as incomes and earnings, human settlement sizes, and so on. We choose the DPLN distribution as a candidate to it the line impedance data in a power grid because the latter implicitly relate to human settlement size in the network. Generally speaking, large-scale generation is usually located near water or the fuel source, but remotely from densely populated areas with large-scale human settlements. Therefore, longdistance transmission lines have been constructed in order to provide the interconnection and serve the large-scale settlements. And the long-distance lines usually exhibit high line impedance. 166 Smart Grids The lognormal-clip and DPLN-clip are especially suited for itting the NYISO data because the line impedances in this system appear to approximate a lognormal or DPLN distribution very well except for having an interrupted tail, which is captured by the clipping. Therefore, it can be assumed that the NYISO data resulted from some original impedance data being “clipped” by an exponential cutoff tailing coeficient. That is, given the original impedance data Y following a speciic distribution, −Y Z Y ∼ f Y (y), the clipped impedance data are X = Z max 1 − e ( max ) , with Z max being the cutoff threshold, so that Y = − Z max log(1 − X / Z max ). And the resulting clip distribution turns out to be ( ) X ~ f X ( x ) = Y(′x ) fY (Y( x ) ) = ( ) Z max fY − Z max log (1 − x Z max ) . Z max − x (7.18) It is reasonable to introduce the clipping mechanism into the line impedance distribution because in real-world power grids transmission lines are limited in length, and so, correspondingly, is the line impedance, which is proportional to the length. As shown next, the line impedances in real-world power grids have heavytailed distributions. Hence, the candidate distribution functions listed below are selected from the class of heavy-tailed distributions: Gamma Γ ( x | a, b ) = 1 x a −1e x / b ; baΓ ( a ) (7.19) GP ( x − θ) 1 gp ( x | k, σ, θ ) = 1 + k σ σ −1− (1 k ) ; (7.20) Lognormal logn ( x | µ, σ ) = 1 xσ 2 π 2 e −(log x −u ) 2 σ2 ; (7.21) DPLN dPlN ( x | α, β, µ, σ ) = 2 αβ − α −1 log x − µ − α σ A ( α, µ, σ ) x Φ α + β σ log x − µ + β σ2 + A ( −β, µ, σ ) x β −1Φ c , σ (7.22) Power Grid Network Analysis for Smart Grid Applications where A ( θ, µ, σ ) = e( θµ + θ σ Lognormal-clip 2 2 167 / 2) ; logn clip ( x | µ, σ, Z max ) = Z max X µ, σ ; logn − Z max log 1 − Z max − x Z max (7.23) DPLN-clip dPlN clip ( x | α, β, µ, σ, Z max ) = Z max X dPlN − Z max log 1 − α, β, µ, σ . Z max − x Z max (7.24) For each candidate distribution function, one can estimate its parameters using the ML method, and then run the K-S test in order to pick the distribution that gives the best K-S test result. Table 7.4 shows the best-itting distribution functions with the corresponding ML parameter estimates. Figure 7.6 compares the empirical PDFs and TABLE 7.4 PDF Fitting for the Line Impedances in the IEEE and NYISO Power Grids System Fitting Distribution IEEE-30 Γ(x|a, b) IEEE-118 Γ(x|a, b) IEEE-57 gp(x|k, σ, θ) IEEE-300 gp(x|k, σ, θ) NYISO-2935 lognclip(x|μ, σ, Zmax) dPlNclip(x|α, β, μ, σ, Zmax) ML Parameter Estimates (α = 0.05) a = 2.14687 b = 0.10191 a = 1.88734 b = 0.05856 k = 0.33941 σ = 0.16963 θ = 0.16963 k = 0.45019 σ = 0.07486 θ = 0.00046 μ = −2.37419 σ = 2.08285 Zmax = 1.9977 α = 44.25000 β = 44.30000 μ = −2.37420 σ = 2.082600 Zmax = 1.9977 168 Smart Grids NYISO: Empirical PDF of Zpr and DPLN-clip fitting PDF Density 104 (a) NYISO Zpr DPLN-clip 102 100 10−2 10−4 1 10−3 10−2 10−1 100 101 NYISO: Empirical CDF of Zpr and DPLN-clip fitting CDF F (Zpr) 0.8 0.6 0.4 NYISO Zpr DPLN-clip 0.2 0 (b) 0 0.2 0.4 0.6 0.8 1 1.2 Zpr (p.u.) 1.4 1.6 1.8 2 FIGURE 7.6 Comparison between empirical PDF/CDF distributions and itted distributions: (a) IEEE 300; (b) NYISO 2935. cumulative distribution functions (CDFs) (on logarithmic scales) with those from the best-itting distribution function for the NYISO system. The studies show that the gamma distribution is the best it for the line impedances in the IEEE 30 and 118 systems and the GP distribution for those of the IEEE 57 and 300 systems, while for the NYISO system lognormal-clip and DPLN-clip best it the data. This is also reasonable, because, even if it is not one to one, the line impedance usually grows with the geographic distance between the buses it connects. In addition, it is well known that the distribution of human settlements tends to be clustered; thus, it is reasonable that long tails appear between these clusters of smallersize buses, giving rise to these heavy tails. This observation seems to agree with the topological properties found in the power grid graphs, and it is reasonable to consider these heavy-weight impedances as good candidates for line impedances of the links that are rewired in the small-world islands, as well as the ones connecting difference islands in our nested-Smallworld model, which is introduced in the next section. 7.5 A MODEL TO GENERATE RT TEST NETWORKS There are a number of works in the power network literature that propose models to generate scalable-size power grid test cases, such as: the use of ring-structured grids to study the propagation pattern of disturbances [18]; a tree-structured power grid model for detecting critical transitions in the transmission network that cause cascading failure blackouts [19]; the Watts–Strogatz small-world network [1]; or the formation of a power grid topology starting from a uniform or a Poisson distribution for the nodal location (emulating what is often done in modeling wireless networks) [20]. While these proposed models provide some good perspectives on power grid Power Grid Network Analysis for Smart Grid Applications 169 characteristics, the generated topology fails to correctly or fully relect that of a realistic power system. This section presents a random topology power grid model, RT-nested-Smallworld, which constructs a large-scale power grid in a hierarchical way: irst forming connected subnetworks with size limited by the connectivity requirement; then connecting the subnetworks through lattice connections; and, inally, generating the line impedances from some speciic distribution and assigning them to the links in the topology network. The hierarchy in the model arises from observation of realworld power grids: usually a large-scale system consists of a number of smaller-size subsystems (e.g., control zones), which are interconnected by sparse and important tie-lines. The model contains three main components: (a) clusterSmallWorld subnetwork, (b) lattice connections, and (c) generation and assignment of line impedances. These will be described in detail in the following subsections. 7.5.1 clustersmallWorld suBnetWork Electrical power grid topology has “small-world” characteristics; it is sparsely connected with a low average nodal degree that does not scale with the network size. On the other hand, in order for a small-world model to generate a connected topology, the network size has to be limited. In this proposed model, different mechanisms from those of the Watts–Strogatz small-world model have been adopted to form a power grid subnetwork in order to improve its resulting connectivity, as shown in the following paragraphs. Consequently, the connectivity limitation on the network size can be expanded from what is indicated by Equation 7.9. The experiments have shown that for 〈k〉 = 2 – 3, the network size should be limited to no greater than 30, and for 〈k〉 = 4 – 5, no greater than 300. Therefore, the irst step of this new model is to select the size of subnetworks according to the connectivity limitation. Then a topology is built up through a modiied small-world model, called clusterSmallWorld. This model is different from the Watts–Strogatz small-world model in two aspects: the link selection and the link rewires. That is, instead of selecting links to connect most immediate 〈k〉/2 neighbors to form a regular lattice, it selects a number k of links at random from a local neighborhood N d0 with the distance threshold of d0, where k comes from a geometric distribution. The local neighborhood is deined as the group of close-by nodes with mutual node number difference less than the threshold d0, that is, N d(i0) = { j; | j − i |< d0} for node i. It is worth noting that the clusterSmallWorld model adopts a geometric distribution with the expectation of 〈k〉 for the initial node degree settings (i.e., for the link selection). Experiments have shown that the process of link selection and link rewiring that follows will transform the initial node degree distribution and inally result in a distribution with good matching approximation to the observed distribution. For the link rewiring, the Watts–Strogatz small-world model selects a small portion of the links to rewire to an arbitrary node chosen at random in the entire network to make it a small-world topology. The clusterSmallWorld model uses a Markov chain with transition probabilities of (α, β), as shown in Figure 7.7, to select clusters of nodes and therefore groups of links to be rewired. This mechanism is introduced in order to produce a correlation among the rewired links. After running the above Markov transition 170 Smart Grids α 0 1–α 1–β 1 β FIGURE 7.7 Markov chain for selecting cluster of rewiring nodes (0: not rewire; 1: rewire). n times (i.e., one for each node in the network), we get clusters of nodes with “1”s alternating with clusters of nodes with “0”s, where “1” means to rewire and “0” means not. Then, by a speciic rewiring probability qrw, some links are selected to rewire from all the links originating from each 1-cluster of nodes, and the corresponding local links are rewired to outside 1-clusters. The average cluster size for rewiring nodes is Kclst = 1/β, and the steady-state probabilities are p0 = β/(α + β), p1 = α/(α + β). Experiments are performed on the available real-world power grid data to estimate the parameters. The average rewiring cluster size Kclst, the rewiring probability of links qrw, and the ratio of nodes having rewire links p1 can be directly obtained from statistical estimates. Then the transition probability can be computed as β = 1/Kclst and α = βp1/(1 − p1). 7.5.2 lattIce connectIons The lattice connections are selected at random from neighboring subnetworks to form a whole large-scale power grid network. The number of lattice connections between neighboring subnetworks is randomly chosen to be an integer with the mean of 〈k〉. 7.5.3 assIgnment oF lIne ImPedances At this step m line impedances are generated from a speciied heavy-tailed distribution, and then sorted by magnitude and grouped into local links, rewire links, and lattice-connection links according to their corresponding proportions, as shown in Figure 7.8. Finally, line impedances in each group are assigned at random to the corresponding group of links in the topology. Impedances (p.u.) Lattice-conn Local 1 FIGURE 7.8 Grouping line impedances. Rewires m Link number Power Grid Network Analysis for Smart Grid Applications 7.6 171 TOPOLOGICAL ANALYSIS FOR THE DISTRIBUTION NETWORKS Usually the transmission network of a power grid is an HV network with voltage levels above 50 kV. The studies and analysis in this chapter have so far focused on the HV transmission network of a power grid. Now we extend the topological analysis to the distribution side of the power grid. A distribution network carries electricity from the transmission system and delivers it to end users. Typically, the network would include MV (<50 kV) lines, electrical substations and polemounted transformers, LV (<1 kV) distribution wiring, and, sometimes, electricity meters. The results presented are based on a sample 396-node MV distribution network, which comes from a real-world US distribution utility mainly located in a rural area. 7.6.1 structure oF the dIstrIButIon netWork In the LV and MV sections of the grid, the physical layout is often restricted by what land is available and its geology. The logical topology can vary depending on budget constraints, system reliability requirements, and load and generation characteristics. Generally speaking, there are a few typical kinds of topology in the distribution network: ring, radial, and interconnected. A radial network is the cheapest and simplest topology for a distribution grid, and the one most often encountered. This network has a tree shape, in which power from a large supply radiates out into progressively lower-voltage lines until the destination homes and businesses are reached. It is typical of long rural lines with isolated load areas. Today’s grid is radially operated with respect to the current transmission system, but this topology will no longer work when distributed energy resources (DER) are integrated into the grid. Unfortunately, this topology is the worst in terms of maximum communication delay, because the number of hops between its nodes tends to grow in the order of the size of the network. An interconnected or ring network is generally found in more urban areas and will have multiple connections to other points of supply. These points of connection are normally open, but allow the operating utility to set up various conigurations by closing and opening switches. Operation of these switches may be by remote control from a control center or by a lineman. The beneit of the interconnected model is that, in the event of a fault or required maintenance, a small area of the network can be isolated and the remainder kept on supply. Most areas provide three-phase industrial service. A ground is normally provided, connected to conductive cases and other safety equipment, to keep current away from equipment and people. Distribution voltages vary depending on customer need, equipment, and availability. Within these networks there may be a mix of overhead line construction utilizing traditional utility poles and wires and, increasingly, underground construction with cables and indoor or cabinet substations. However, an underground distribution network is signiicantly more expensive than lines deployed through an overhead construction. Distribution feeders emanating from a substation are generally controlled by a circuit breaker, which will open when a 172 Smart Grids fault is detected. Automatic circuit reclosers may be installed to further segregate the feeder, thus minimizing the impact of faults. It is important to remark that long feeders experience voltage drop, and thus require the installation of capacitors or voltage regulators. 7.6.2 graPh theoretIc analysIs oF a samPle mv dIstrIButIon netWork The logical topology of the sample 396-node MV distribution network analyzed here is shown in Figure 7.9. The power supply comes from the 115 to 34.5 kV stepdown substation. Most nodes or buses in the network are 12.47 kV, and only a small number of them are 34.5 or 4.8 kV. As shown in Figure 7.9, an MV network usually comprises different voltage levels, separated by transformers. In the following topology analysis of the sample MV network, it is assumed that wireless or wired couplers have been implemented at the locations of transformers and switches, so that the network connectivity will not be affected by transformer types or switch status. On the other hand, if couplers are missing, the network will be segmented into several sections either by the transformers or by the open switches. For the sample MV network analyzed here, most buses (>95%) in the network are at the same voltage level of 12.47 kV. Therefore, the topology analysis result of the separated 12.47 kV subnetwork is in fact very close to that of the whole connected graph. FIGURE 7.9 A 396-node MV distribution network in a rural area of the United States. Components: bus (circle), line branches (line ending with dots), switches (line ending with “+”s), transformers (lines ending with “x”s), open or out-of-service component (light gray); the node color represents its voltage level: 115 kV (gray X), 34.5 kV (gray O), 12.47 kV (black), 4.80 kV (gray). 173 Power Grid Network Analysis for Smart Grid Applications The topology metrics we evaluated include the following: • (n,m): the total number of nodes and branches, which represents the network size well. • 〈k〉: the average node degree, which represents the average number of branches a node connects to. • 〈l〉: the average shortest path length in hops between any pair of nodes. • ρ: the Pearson correlation coeficient, which evaluates the correlation of node degrees in the network. This measure relects whether a node adjacent to a highly connected node also has a large node degree. • λ2(L): the algebraic connectivity, which is the second smallest eigenvalue of the Laplacian matrix and is an index of how well a network is connected and how quickly information data can be shared across the network. • C(G): the clustering coeficient, which assesses the ratio of nodes tending to cluster together. The result of the analysis is shown in Table 7.5 in comparison with two other transmission networks: the IEEE-300 system represents a synthesized network from the New England power system and has a network size comparable to the 396-node MV distribution network we analyzed; the WSCC grid is the interconnected system of transmission lines spanning the western United States plus parts of Canada and Mexico, and contains 4941 nodes and 6594 transmission lines. It is well known that transmission and distribution topologies differ; nevertheless, to comment on these differences in a quantitative manner, as in this exercise, is useful for several reasons. For example, it allows us to better understand the characteristics of the transmission and distribution networks as information sources; it allows us to optimize the design of the distribution PMU-based wide-area measurement systems rather than attempting to duplicate the existing transmission, which is tailored to a network with very different topological characteristics; and it can tell us how the distribution topology can be “modiied” to achieve some advantageous characteristics of the transmission network, that is, shorter path lengths between nodes, better algebraic connectivity, and so on. From Table 7.5, we can see that the 396-node MV distribution network has an average node degree of 〈k〉 = 2.12, which is comparable to, although a little bit lower than, those of the other two transmission networks, the IEEE-300 system and the WSCC system. That means its average connecting sparsity is at about the same level TABLE 7.5 Topological Characteristics of the Transmission Networks and the MV Distribution Network IEEE-300 WSCC MV distribution n 〈k〉 〈l〉 ρ λ2(L) C(G) C(R) 300 4941 396 2.73 2.67 2.12 9.94 18.70 21.10 −0.2206 0.0035 −0.2257 0.0094 0.00076 0.00030 0.0856 0.0801 0.0000 0.009119 0.000540 0.005367 174 Smart Grids as the transmission networks it is compared with. However, the sample MV distribution network has a much longer average path length, 〈l〉 = 21.10 in hops, than the IEEE-300 system, and, interestingly, it is even longer than that of the much larger 4941-node WSCC system. More speciically, any node in this MV distribution network is about 16.50 hops away from node-1 or node-2, which are 115 kV buses at the HV side of the two step-down supply transformers and may likely serve as the trafic sinks in the power line communication (PLC)-based network. Looking at the algebraic connectivity λ2(L), the 396-node MV distribution network has a much weaker overall connectivity compared with the transmission networks, that is, λ2(L) = 0.00030 versus 0.0094 (IEEE-300) and 0.00076 (WSCC). This result shows that this topology is highly prone to becoming a disconnected graph under node failure (islanding). Finally, the most distinctive difference we found lies in the fact that the 396-node MV distribution network has a clustering coeficient equal to zero, compared with the clustering coeficient of 0.0856 for the IEEE-300 system and 0.0801 for the WSCC system. This means that no node in the sample MV distribution network is the vertex of a complete subgraph (triangle). MV distribution networks not located in rural areas are generally less prone to becoming disconnected graphs, as in urban areas it is not unusual for utilities to provide link redundancy by, for example, adding rings. If the distribution network becomes a disconnected graph, data connectivity obviously suffers if PLC is used. This vulnerability of the distribution network can be alleviated by judiciously adding wireless links to complement the PLC-based network with the goal of improving network connectivity as well as shortest path length characteristics. 7.6.3 the PmF oF the node degrees In the samPle 396-node mv dIstrIButIon netWork The average node degree of a power grid transmission network tends to be quite low and does not scale as the network size increases. The topology of a transmission network has salient small-world properties, since it features a much shorter average path length (in hops) and a much higher clustering coeficient than that of Erdös–Rényi random graphs with the same network size and sparsity. While small-world features have been recently conirmed for the HV transmission network, the sample MV network used here implies that a power grid distribution network has a very different kind of topology from that of an HV network, and obviously it is not a small-world topology. The node degree distribution of the 396-node MV distribution network, shown in Figure 7.10, shows that the maximum node degree in the network equals 4, which is much smaller than that found in the transmission side of the grid, where maximum nodal degrees of 20 or 30 can be found; and about 16% of the nodes connect to only one branch, 60% to two branches, 22% to three branches, and only 2% to four branches. Figure 7.4e depicts the network’s spectral density, which is a normalized spectral distribution of the eigenvalues of its adjacency matrix. The spectrum of an Erdös– Rényi random graph network, which has uncorrelated node degrees, converges to a semicircular distribution (see the semicircle dotted line on the background in Figure 7.4e). The spectra of real-world networks have speciic features depending on the details of the corresponding models. In particular, scale-free graphs develop 175 Power Grid Network Analysis for Smart Grid Applications 0.7 0.6 PMF (k) 0.5 0.4 0.3 0.2 0.1 0 1 2 3 k-node degree 4 5 FIGURE 7.10 PMF of the node degrees in the sample 396-node MV distribution network. 100 0.35 Probability mass function 0.30 0.25 10−1 0.20 0.15 10−2 0.10 0.05 0 10−3 0 1 2 Line length (103 m) 3 0 1 2 Line length (103 m) 3 FIGURE 7.11 PMF of the line length in the sample 396-node MV distribution network: (left) probability versus length; (right) log-probability versus length, where the existence of an exponential trend in the tail is clearly visible. a triangle-like spectral density with a power-law tail, whereas a small-world network has a complex spectral density consisting of several sharp peaks. The plot in Figure 7.4e indicates that the sample MV distribution network is neither a scale-free network nor a small-world network. The branch lengths in the MV distribution network are also analyzed. The corresponding PMF is shown in Figure 7.11. It indicates that most of the branches are shorter than 1067 m (3500 ft) and the branch length distribution has an exponential tail, with only a very small number of extremely long branches. 176 7.6.4 Smart Grids lv dIstrIButIon netWork It is dificult to obtain example data about LV distribution network topologies. Generally speaking, an LV distribution network is radial, and has a similar network topology to an MV distribution network, except that it may have more nodes with shorter branch length. 7.7 SUMMARY This chapter presents a comprehensive study on both the topological and electrical characteristics of electric power grids. It inds that the HV transmission network of the power grid is sparsely connected, with a low average node degree that does not scale with the network size; the power grid topology manifests obvious small-world properties, but its small-world rewires are not independent as in the Watts–Strogatz small-world model; instead, long hauls appear among clusters of nodes. This chapter provides evidence that the nodal degrees follow a mixed distribution that comes from the sum of a truncated geometric random variable and an irregular discrete random variable. It also proposes a method to estimate the distribution parameters by analyzing the poles and zeros of the average PGF. The chapter studies the graph spectral density and shows that the power grid network takes on a distinctive pattern for its graph spectral density. Study of the algebraic connectivity of power networks has shown that power networks are exceptionally well connected given their sparsity, featuring a better scaling law than the Watts–Strogatz small-world graphs. It is shown that the distribution of line impedances is heavy-tailed and can be captured quite accurately by a clipped DPLN distribution. The chapter introduces a novel random topology power grid model, RT-nested-Smallworld, intending to capture the above observed characteristics. Finally, the chapter extends the same topological analysis to the distribution side of the power grid and reports the results based on a sample 396-node MV distribution network that comes from a real-world US distribution utility mainly located in a rural area. REFERENCES 1. D. J. Watts and S. H. Strogatz, Collective dynamics of “small-world” networks, Nature, 393, 393–440, 1998. 2. R. Solé, M. Rosas-Casals, B. Corominas-Murtra, and S. Valverde, Robustness of the European power grids under international attack, Physical Review E, 026102, 2008. 3. Z. Wang, A. Scaglione, and R. J. Thomas, The node degree distribution in power grid and its topology robustness under random and selective node removals, in Proceedings of the 1st IEEE International Workshop on Smart Grid Communications (ICC’10SGComm), 23–27 May, Cape Town, South Africa, 2010. 4. R. Albert, I. Albert, and G. L. Nakarado, Structural vulnerability of the North American power grid, Physical Review E, 69, 1–4, 2004. 5. P. Crucittia, V. Latorab, and M. Marchioric, A topological analysis of the Italian electric power grid, Physica A, 338, 92–97, 2004. 6. E. W. Dijkstra, A note on two problems in connexion with graphs, Numerische Mathematik, 1, 269–271, 1959. Power Grid Network Analysis for Smart Grid Applications 177 7. J. L. Rodgers and W. A. Nicewander, Thirteen ways to look at the correlation coeficient, The American Statistician, 42(1), 59–66, 1988. 8. M. Newman, The structure and function of complex networks, SIAM Review, 45, 167–256, 2003. 9. D. E. Whitney and D. Alderson, Are technological and social networks really different, in Proceedings of the 6th International Conference on Complex Systems (ICCS06), Boston, MA, 2006. 10. P. Erdös and A. Rényi, On random graphs. I, Publicationes Mathematicae, 6, 290–297, 1959. 11. R. Albert and A. Barabási, Statistical mechanics of complex networks, Reviews of Modern Physics, 74(1), 47–97, 2002. 12. University of Washington, Electrical Engineering. Power systems test case archive, http://www.ee.washington.edu/research/pstca/. 13. E. J. Kirk, Count loops in a graph, Matlab Central, The Mathworks Inc., http://www. mathworks.com/matlabcentral/ileexchange/10722, 2007. 14. E. P. Wigner, Characteristic vectors of bordered matrices with ininite dimensions, The Annals of Math, 62, 548–564, 1955. 15. E. P. Wigner, On the distribution of the roots of certain symmetric matrices, The Annals of Math, 67, 325–328, 1958. 16. M. Fiedler, Laplacian of graphs and algebraic connectivity, Combinatorics and Graph Theory, 25, 57–70, 1989. 17. W. J. Reed and M. Jorgensen, The double pareto-lognormal distribution: A new parametric model for size distributions, Communications in Statistics, Theory and Methods, 33(8), 1733–1753, 2004. 18. M. Parashar and J. S. Thorp, Continuum modeling of electromechanical dynamics in large-scale power systems, IEEE Transactions on Circuits and Systems, 51(9), 1848–1858, 2004. 19. B. A. Carreras, V. E. Lynch, I. Dobson, and D. E. Newman, Critical points and transitions in an electric power transmission model for cascading failure blackouts, Chaos, 12(4), 985–994, 2002. 20. Z. Wang, R. J. Thomas, and A. Scaglione, Generating random topology power grids, in Proceedings of the 41st Annual Hawaii International Conference on System Sciences (HICSS-41), vol. 41, 7–10 January, Waikoloa, HI, 2008. 8 Open Source Software, an Enabling Technology for Smart Grid Evolution Russell Robertson, Fred Elmendorf, and Shawn Williams CONTENTS 8.1 8.2 Synopsis ........................................................................................................ 179 Introduction and Brief History ..................................................................... 180 8.2.1 Introduction ...................................................................................... 180 8.2.2 Brief History ..................................................................................... 180 8.3 Beneits and Challenges................................................................................ 181 8.3.1 Beneits ............................................................................................. 181 8.3.2 Challenges......................................................................................... 182 8.3.3 Licensing Considerations .................................................................. 182 8.4 Present State and Examples .......................................................................... 183 8.4.1 Present State ..................................................................................... 183 8.4.2 Examples........................................................................................... 183 8.4.3 Examples: Grid Protection Alliance (GPA) ...................................... 184 8.5 Future of the Electric Utility Industry .......................................................... 189 References .............................................................................................................. 190 8.1 SYNOPSIS The infrastructure that makes up the electric power grid has always been in a state of evolution. Design strategies, physical hardware, and materials continue to improve over time, and the connected nature of the grid makes it impossible for these changes to take place in isolation. Similarly, with the introduction of computers and intelligent electronic devices, sources of data continuously multiply and new tools and strategies are developed for better understanding and management of the power grid. A unique interdependency among the companies that make up the electric power grid creates a need for interactions on common ground. Open source software (OSS) addresses the need for common ground in application development. The purpose of this chapter is to introduce the concepts of OSS, provide a brief look at OSS history, cover some of its beneits and challenges, show how it meets 179 180 Smart Grids power system needs through examples of existing tools, and present a future vision of OSS in the electric utility industry. 8.2 8.2.1 INTRODUCTION AND BRIEF HISTORY IntroductIon OSS is a critical part of today’s computing fabric and provides the backbone for communications, the Internet, and many back-end services. OSS is not the same as “Freeware,” which is free to obtain but does not include the source code, so the user assumes the risk that it does what it is supposed to and nothing else. While OSS carries no licensing fees, by deinition it must include the source code. The exposure of the source code allows everyone to review it and determine that it does what it is supposed to do in the proper way. OSS is maintained in a trusted repository, and code changes are veriied before distribution. A diagram of the OSS development model as presented by David Wheeler1 is included in Figure 8.1. 8.2.2 BrIeF hIstory In the early days of computing, developers viewed writing software as a form of personal expression. Programming was a functional art form that developers were happy to share with others in the community. The open sharing of concepts and techniques facilitated collaboration and allowed rapid advances in software tools. Computer companies generated most of their income from the hardware, not the software it supported. However, by the mid- to late 1970s a shift in focus occurred, and computer manufacturers began to introduce proprietary software2 to increase differentiation, protect their personal interests, and develop a new income stream. In the electric utility community, as computing resources became available for power system engineers and operators, virtually all of their tools were developed collaboratively and user groups were formed to coordinate activities. Some of the user-supported activities and early applications have survived through the years, Source code Trusted repository Distributors Problem reports New feature requests Trusted developers Problem reports Developers Executables Users FIGURE 8.1 OSS development model. Open Source Software, an Enabling Technology for Smart Grid Evolution 181 but over time a large majority of applications have become proprietary. An early response to this transition was the emergence of the term “free software.” However, this term could be confusing, since to some it meant “free—as in freedom of speech and expression”3 while to others it meant “free—as in free beer.” In 1998, The Open Source Initiative (OSI)4 coined the term “open source” as part of a marketing initiative that it hoped would remove ambiguity and convey a more positive message. In spite of the fact that the majority of applications migrated to proprietary software, a huge portion of the overall computing infrastructure has remained open source. Some well-known examples are servers that support the Internet and communications, and, more recently, “big data” repositories. Driven by many converging forces, such as the need for agility, eficiency, and security, a new emphasis has been placed on developing high-quality open source tools. A 2003 study showed that OSS was already being widely used in the Department of Defense, and since that time a number of policy statements have been made requiring that OSS be considered equally or even preferentially to equivalent proprietary solutions. 8.3 BENEFITS AND CHALLENGES OSS provides many beneits compared with similar proprietary applications, but can be subject to many of the same issues. The quality and reliability of any software are dependent on the skill of the developer or development team, and are not dictated by the licensing under which it is distributed. From a user perspective, OSS has the potential for providing value and lexibility that are not available under a proprietary license. The following list of beneits is not comprehensive, but presents a variety of positive qualities that are inherently part of OSS. 8.3.1 BeneFIts Security: The code visibility provided by OSS provides the opportunity for wide review, and the lexible nature of OSS allows identiied changes to be deployed rapidly. Jim Stogdill describes this as maneuverability that can be used in cybersecurity combat.5 No Vendor Lock-in: The user is free to move from one application to another, or from one source of support to another, and has access to the source code, which can be maintained or enhanced in-house if desirable. Low Cost: Since there are no initial or annual licensing fees, there can be substantial savings in the acquisition and deployment of OSS applications. When enhancements are needed, development costs are limited to the task at hand, and do not carry the burden of supporting a larger organization. Bug ixes and some broadly appealing enhancements are often donated, which further reduces the total cost of ownership. High Quality: OSS applications must be of reasonably good quality to be successful in attracting community involvement and engaging users. Once established, the community provides continuous review by “many eyes,” and there is no way to “hide” a bug. Additionally, the development capabilities are not limited to one company’s resources. The code base is maintained by a steward, who veriies any changes before they are distributed. 182 Smart Grids State of the Art: OSS is not bound by previous commercial investment, and has the lexibility to use the latest techniques. It is often developed with intentional design for extension and enhancement, and the continuous community review avoids stagnation, which can occur within the developers of a single company. Perfect Platform for Collaboration: The absence of inancial concerns around licensing fees and the freedom to share code and ideas with a large community make OSS an ideal vehicle for bringing together many participants from different entities such as research, academia, government, utility, and vendors. Facilitates Innovation: The collaborative aspect of OSS described above eliminates the “not invented here” syndrome, and provides developers with more freedom to experiment. Community sharing of ideas and reviewing each other’s work foster creativity. Decreases Time to Deployment: The community environment of OSS allows direct application of the concept “many hands make light work,” and decreased restrictions combined with increased innovation and creativity result in decreased development time. There are no contract negotiations and no inancial incentive to stall a release to recoup the investment in the previous version, which also facilitates more timely delivery of initial deployment and enhancements. 8.3.2 challenges The biggest challenges for OSS are misconceptions and a lack of education. Nonproprietary can be misinterpreted to mean noncommercial, but OSS is typically very high-quality, well-managed, commercial-grade code. Another misconception is that OSS is developed by inexperienced college students; typically it is developed by skilled developers or teams. No licensing fee can be misunderstood to mean no cost, but training, support, transition, and enhancements are usually fee based. Uneducated assumptions regarding OSS could include the notions that it is less reliable and has more bugs than proprietary software, but major studies have shown that, in comparisons of OSS with proprietary operating systems and web server software, OSS scores much more favorably than proprietary counterparts. David Wheeler includes a good discussion of OSS myths and reality in his 2012 presentation titled “Open Source Software (OSS or FLOSS) and the US Department of Defense (DoD).” Another type of challenge for OSS is status quo and momentum. Electric utility companies are necessarily risk averse, and often ind comfort in continuing on the current course of action. However, as cost savings become ever more important, and the increased lexibility and agility of OSS are better understood, the resistance to change is decreasing. At the most basic level, OSS faces the same challenges and constraints as any proprietary development. It has to be the right solution for the problem, it has to be developed and maintained professionally, and it has to provide the desired user experience. 8.3.3 lIcensIng consIderatIons Users of OSS have less concern than developers about speciic licensing issues, as long as they are not involved in new development, and have identiied adequate Open Source Software, an Enabling Technology for Smart Grid Evolution 183 resources for maintenance and support. Developers of OSS need to carefully consider their goals regarding further development, marketing options, and integration with other software. OSI, a nonproit organization, was formed in 1998 to be the steward of the Open Source Deinition, build a community, educate, and promote the awareness and importance of nonproprietary software. It maintains a comprehensive library of open source licenses and is a valuable resource when exploring new licensing options. 8.4 PRESENT STATE AND EXAMPLES 8.4.1 Present state OSS is a vehicle to effectively share software platforms and technologies throughout the electric industry. This promotes a common ground for communication, data sharing, analysis, planning and management of the grid. It removes the common hindrances of the traditional proprietary software model. Methods and processes are developed by consensus of all parties with a reduced chance of miscommunications and misunderstandings. As noted on the OSI web page, OSS is continuing to gain acceptance. It is demonstrating the ability to deliver on the promise of better quality, higher reliability, more lexibility, lower cost, and no vendor lock-in. Fueled by the MITRE study for the US DoD in 2003,6 OSS has been continually gaining ground in US government agencies. In addition to the DoD, the US Department of Energy (DoE) is also encouraging OSS and using it as the vehicle of choice for technology transfer in applications developed through the department’s research projects. With the everincreasing recognition that OSS is a viable contender as a source for applications, electric utilities are taking advantage of this opportunity to reduce cost while maintaining quality and increasing lexibility. 8.4.2 examPles Two of the best-recognized and most successful OSS projects are the Linux7 operating system and the Apache web server. These are foundations of our modern computing world, but most electric utility users are not directly involved in those communities. Two examples of very successful OSS projects for power system applications are UWPFLOW8 and TEFTS9 from The University of Waterloo, Ontario, Canada. Listed below are additional examples of OSS in the power grid: IEEE Task Force on Open Source Software for Power Systems: http://ewh. ieee.org/cmte/psace/CAMS_taskforce/index.htm This Task Force explores the potential for open source software (OSS) in the Power Engineering Society (PES). The mission of the Task Force is twofold: (i) to diffuse the philosophy of OSS in the power systems community; and (ii) promote OSS for the beneit of the PES ranging all the way from simple pedagogical OSS to commercial-grade OSS. 184 Smart Grids OpenPMU: https://sites.google.com/site/openpmu/ The OpenPMU project is a suite of tools to help researchers create new power system technologies. The project was originally founded at Queen’s University Belfast, UK, and subsequently joined by the SmarTS Lab at KTH Stockholm, Sweden. For more information on the origins of the project, see here. GridLAB-D: http://www.gridlabd.org/ GridLAB-D™ is a new power distribution system simulation and analysis tool that provides valuable information to users who design and operate distribution systems, and to utilities that wish to take advantage of the latest energy technologies. It incorporates the most advanced modeling techniques, with high-performance algorithms to deliver the best in end-use modeling. GridLAB-D™ is coupled with distribution automation models and software integration tools for users of many power system analysis tools. GridLAB-D™ was developed by the U.S. Department of Energy (DOE) at Paciic Northwest National Laboratory (PNNL) under funding for Ofice of Electricity in collaboration with industry and academia. Total Grid Community: http://www.totalgrid.org/ An open source community dedicated to enabling a smarter energy grid. Green Energy Corp is pleased to sponsor this Community and looks forward to continued partnership with open source developers world-wide. OpenDSS: http://electricdss.sourceforge.net/ The OpenDSS is an electric power Distribution System Simulator (DSS) for supporting distributed resource integration and grid modernization efforts. OpenETran: http://sourceforge.net/projects/openetran/ OpenETran can presently simulate multi-conductor power lines, insulators, surge arresters, nonlinear grounds, and lightning strokes. It eficiently calculates energy and charge duty on surge arresters, and iterates to ind the critical lightning current causing lashover on one or more phases. It is also suitable for use in substation insulation coordination. Capacitor switching, TRV, and other applications may be added. 8.4.3 examPles: grId ProtectIon allIance (gPa) One example of a software development company dedicated to the development of OSS tools is GPA. GPA is a not-for-proit corporation formed to facilitate and support the development and deployment of comprehensive electric energy system-wide solutions and enhancements to electric sector security through integrated, adaptable, collaborative and rapidly evolving programs of threat identiication and analysis, and infrastructure research, development, demonstration, and technology transfer. GPA focuses on improvements to electric energy resiliency and maintenance of reliability and availability. The efforts will accommodate the evolving technology of electric energy systems, including grid modernization, advanced generation sources, and enhanced systems integration. Open Source Software, an Enabling Technology for Smart Grid Evolution 185 The Grid Solutions Division of GPA is located in Chattanooga, Tennessee, and develops OSS applications to meet the speciic demands of real-time data collection and storage to support electric grid operations and planning, and an extensible platform for the automation of disturbance analysis. GPA products are hosted in the CodePlex open source project community provided by Microsoft. GPA maintains product documentation and code on the CodePlex project web sites. For those who wish to contribute code, GPA has published coding guidelines. Many of GPA’s current products focus on synchrophasor applications. Synchrophasors are global positioning system (GPS) time-stamped measurements taken rapidly (at least 30 times per second) from different points on the power grid. These synchronized measurements are transmitted in real time to operations centers, where they provide information on the state and health of the power system. Synchrophasor information can be used to monitor power lows on the electric grid and to identify areas or conditions where problems exist. More information about synchrophasors can be found in the October 2010 North American Electric Reliability Corporation (NERC) report, “Real-Time Application of Synchrophasors for Improving Reliability.”10 GPA’s most recent development has been the creation of a framework to facilitate enterprise-class automated disturbance analysis systems. A research and demonstration project initially funded through the Electric Power Research Institute (EPRI) for the Tennessee Valley Authority has been extended to facilitate any type of analysis for disturbance recordings. This analysis can range from near real-time fault location to the assessment of historical records for steady-state system changes. A paper entitled “Open Source Fault Location in the Real World”11 was developed for the 2013 Georgia Tech Fault and Disturbance Analysis Conference. GPA software products are developed in the Microsoft .NET framework using the C# language, and all are managed through Codeplex12 projects and licensed under the Eclipse Public License.13 As with other OSS projects, there is no licensing fee for any of our products, but support is offered through either time and materials, or subscription-based agreements. GPA also hosts an annual users’ forum and tutorial where attendees can learn directly from the GPA development team and through presentations from other GPA product users. The GPA software environment is shown in Figure 8.2. A brief description of each GPA tool is included below. PMU Connection Tester: https://pmuconnectiontester.codeplex.com/ The PMU Connection Tester administered by the Grid Protection Alliance (GPA) veriies that a data stream from synchrophasor measurement device is being successfully received. The PMU connection tester supports the following phasor data protocols: IEEE C37.118 (Version 1/Draft 7, Draft 6), IEEE 1344, BPA PDCstream, SEL Fast Message, UTK FNET streaming data protocol, IEC 61850-90-5 and Macrodyne. openPDC: http://openpdc.codeplex.com/ The openPDC administered by the Grid Protection Alliance (GPA) is a complete Phasor Data Concentrator software system designed to process streaming time-series data in real-time. Measured data gathered with 186 Smart Grids GPA products OpenPDC SIEGate (openPG) PMU connection tester OpenXDA openFLE OpenHistorian ...and others Platform building blocks Time series processing core Low-level services Security services Grid solutions framework MS framework class library Common language runtime Platform Operating system Hardware FIGURE 8.2 GPA software environment. GPS-time from many hundreds of input sources is time-sorted and provided to user deined actions as well as to custom outputs for archival. Screen shots and an overview are shown in Figure 8.3. openHISTORIAN: https://openhistorian.codeplex.com/ The openHistorian administered by the Grid Protection Alliance (GPA) is a system for collecting, archiving, retrieving and displaying time-series event data in real-time from multiple data sources. A version of the openHistorian (Version 1.0) has been included with the openPDC since its initial release in 2009. This project is intended to extend and productize this initial version of the openHistorian to improve its performance and its lexibility to store and retrieve a broader range of time-series data types. This work on the openHistorian 2.0 began in early 2012. This site has been established so the community can follow and participate in this development activity. A binary alpha release of the openHistorian 2.0 was made available in 2013. The openHistorian 2.0 has been optimized for: • • • • • Assurance of archived data integrity/continuity Broad data source connectivity High-performance data capture and retrieval Eficient, high-volume data storage High availability. An overview of key features is shown in Figure 8.4. Open Source Software, an Enabling Technology for Smart Grid Evolution 187 Output adapter layer Measurements Input adapter layer Action adapter layer FIGURE 8.3 openPDC overview. SIEGate: https://siegate.codeplex.com/ The objectives of the SIEGate (the Secure Information Exchange Gateway pronounced Psy-gate) project are (1) to improve the security posture and minimize the external cyber-attack surface of electric utility control centers, and (2) to reduce the cost of maintaining current control-room-to-controlroom information exchange. .d 188 Smart Grids Collect/convert Configure Alarm/serve RT data Save SDK allows adapters to be developed for data consumption Recent data served with very low latency High-efficiency data storage with lossless data compression Extract Tools provided for bulk historical data extraction Publish Both publish/subscribe API and web services interfaces Visualize/analyze Companion openVisN tool allows users to view and export data FIGURE 8.4 openHistorian features. Funded by the US Department of Energy, the project team includes the Grid Protection Alliance (GPA) as the project prime and principal software developer, with the University of Illinois at Urbana-Champaign (Illinois) as a research, design, development, and testing partner; the Paciic Northwest National Laboratory (PNNL) as a design review, testing, and evaluation partner; PJM as a demonstration partner; and Alstom Grid as a demonstration and commercialization partner. Work on SIEGate began in 2010 and is expected on conclude in 2013. As shown below, SIEGate provides a security isolation layer between critical internal infrastructure and external systems to protect reliability and market sensitive data. SIEGate reduces the cost of data exchange through ease-of-coniguration. SIEGate implements a true publish-subscribe architecture where the sending gateway owner authorizes data as available for subscription by speciic consuming gateways. Once authorized, the consuming gateway automatically discovers the data that have been made available to it by other SIEGate nodes and allows the selective subscription to them. SIEGate data available for publication and subscription includes measurements, such as SCADA or synchrophasor data; iles, such as SDX iles; and higher-level notiications or alarms that are of signiicance for overall grid operation. These alarms may be conigured to promulgate to all interconnected SIEGate nodes so that global alarms can be raised. See Figure 8.5 for an overview. openPG: https://openpg.codeplex.com SIEGate supersedes all future builds of the openPG. All development work on the openPG has ended and efforts have been transferred to SIEGate as the future platform for secure synchrophasor data exchange openXDA: https://openxda.codeplex.com openXDA is an extensible platform for processing event iles from disturbance monitoring equipment such as digital fault recorders (DFRs) and power quality meters. In one implementation of openXDA analytics, based on the waveform data in these iles and on user-provided transmission line parameters, openXDA determines the fault type and calculates the line-distance to the fault. Open Source Software, an Enabling Technology for Smart Grid Evolution SIEGate manager application SCADA/ EMS SE/CA Gateway input Data channel Reliability plans Outage data Gateway output External Firewall Firewall Command channel SIEGate 189 Phasor data Control center infrastructure Internal FIGURE 8.5 SIEGate overview. Extensible disturbance analytics DFR (COMTRADE) PQmeter (PQDIF) SCADA data Phasor data XDA file watcher Data positioning Analytics openFLEϟ Structured data Time domain Analytics ... Frequency domain Analytics ... Presentation internal web; tablet TED.db Transmission event data Notification SCADA; email System integration reporting, detailed analysis FIGURE 8.6 openXDA overview. openXDA is an extension of work conducted in 2012 on openFLE for the Electric Power Research Institute (EPRI) which is posted under the project http://EPRIopenFLE.codeplex.com/. The openFLE project included an open source C# parser for power quality iles (IEEE 1159.3-2003). For more information, see http://www.gridprotectionalliance.org/pdf/openFLE_ Overview_Landscape.pdf. See Figure 8.6 for an overview. 8.5 FUTURE OF THE ELECTRIC UTILITY INDUSTRY The blackout of 2003 has had an effect on the power grid that has lasted well beyond those few hours. It was a clear wake-up call and provided strong motivation to improve grid monitoring and control strategies, and to seriously consider closed-loop control systems. Since the blackout, a multitude of projects to improve the resiliency of the grid have been initiated, ranging from the DoE and Smart Grid Investment Grants, to self-funded projects at forward-thinking electric companies. The recurring theme is the collection, processing, analyzing, archiving, and sharing of data, and new tools have been developed through these projects. At the October 16–18, 2012, North American SynchroPhasor Initiative14 conference, Western Electricity Coordinating Council announced a trial run of a closed-loop 190 Smart Grids control system at a Bonneville Power Administration site. This is the irst experiment of this magnitude, and GPA products will be used to collect and manage the data. The demonstrated success of OSS in the utility industry and increasing government support, combined with increasing pressure for eficient, agile, secure software applications, will result in even wider development and deployment of OSS platforms and tools. Grid Open Source Software Alliance (GOSSA) is a new nonproit organization being formed as a resource center for OSS in the electric utility industry. The GOSSA initiative is supported by the DoE, NERC, EPRI, large utilities, academia, and many other industry stakeholders. GOSSA will provide a single point of reference for electric power-focused OSS. An industry-wide GOSSA workshop was held at the National Rural Electric Cooperative Association headquarters near Washington, DC, on September 6, 2013. OSS is a vital and growing sector in today’s computing environment, and is the perfect platform for evolutionary changes needed to improve tools for monitoring and managing the electric power grid. Personal research and education in this area are strongly encouraged, and a list of references is provided at the end of this chapter. REFERENCES 1. D. A. Wheeler, 2009, Open Source Software (OSS or FLOSS) and the U.S. Department of Defense (DoD). Essay, November 4, http://www.dwheeler.com/essays/dod-oss.pdf. 2. BusinessDictionary.com, http://www.businessdictionary.com/deinition/proprietarysoftware.html. 3. GNU operating system, http://www.gnu.org/philosophy/free-sw.html. 4. Open Source Initiative, http://www.opensource.org/history. 5. J. Stogdill, Coding is maneuver. http://www.slideshare.net/jstogdill/coding-is-maneuver. 6. MITRE DoD 2003, http://dodcio.defense.gov/Portals/0/Documents/FOSS/dodfoss_pdf. pdf. 7. The Linux Foundation, http://www.linuxfoundation.org/. 8. University of Waterloo, UWPFlow, University of Waterloo. http://ece.uwaterloo. ca/~ccanizar/software/plow.htm. 9. University of Waterloo, Transient Stability and Energy Function Analysis, University of Waterloo. http://ece.uwaterloo.ca/~ccanizar/software/tefts.htm. 10. NERC (North American Electric Reliability Corporation), 2010, Real-time application of synchrophasors for improving reliability. NERC. http://www.nerc.com/docs/oc/ rapirtf/RAPIR%20inal%20101710.pdf. 11. Open Source Fault Location in the Real World, http://172.21.1.150/pdf/Open%20 Source%20Fault%20Location%20in%20the%20Real%20World%20(R3).pdf. 12. CodePlex, http://www.codeplex.com/. 13. EPL (Eclipse Public License), v. 1.0, http://www.eclipse.org/legal/epl-v10.html. 14. NASPI (North American SynchroPhasor Initiative), http://www.naspi.org/. 9 Contribution of Microgrids to the Development of the Smart Grid Tine L. Vandoorn and Lieven Vandevelde CONTENTS 9.1 9.2 Introduction .................................................................................................. 191 Microgrids .................................................................................................... 194 9.2.1 Beneits ............................................................................................. 194 9.2.2 Operating Modes .............................................................................. 195 9.2.2.1 Grid-Connected Mode ....................................................... 195 9.2.2.2 Islanded Mode.................................................................... 196 9.2.2.3 Remote Islanded Mode ......................................................200 9.2.3 Smart Microgrids.............................................................................. 201 9.3 Microgrids in the Smart Grid Paradigm.......................................................202 9.3.1 Similar Beneits and Aims................................................................202 9.3.2 Microgrid Control .............................................................................202 9.3.2.1 Smart Microgrid Control ...................................................202 9.3.2.2 Local Microgrid Control ....................................................202 9.4 Conclusion ....................................................................................................208 Acknowledgments..................................................................................................208 References ..............................................................................................................209 9.1 INTRODUCTION An approach to dealing with the large increase in decentralized unpredictable power sources, the aging grid infrastructure, and the increasing (peak) consumption of electric power, while mitigating the massive investments required for this, is to add more intelligence to the power system. The use of widespread information and communication technologies (ICT) to monitor and manage the distributed energy resources (DER) allows the grid to become more eficient and sustainable [1]. These ICT and new remote management abilities couple the grid elements via an interactive intelligent electricity network, the so-called smart grid. In brief, a smart grid involves the use of sensors, communication, computational ability, and control to 191 192 Smart Grids enhance the overall functionality of the electric power delivery system [2]. The smart grid offers many advantages, such as increasing the share of clean renewable energy, reducing costs, improving system reliability in the face of increasing intermittency, enabling more customer choices and involvement in their energy management, a more eficient usage of the current assets, and help with using electric vehicles [3]. One barrier to the development of the smart grid is that the individual beneits for consumers should be clear. One of the main challenges to increasing customer acceptance of smart grids is dealing with customers’ concerns regarding cost, privacy, and security, together with their fear of possible discomfort when they provide lexibility to the grid, for example, potential disruptions for peak shaving. It is thus important to overcome the lack of understanding about the smart grid’s functionality and increase customers’ knowledge of the beneits for them speciically. As well as lacking awareness about smart grids, consumers are not aware of which appliances are contributing most to their electricity bills. This leads to frustration because, despite the best of intentions, they are unable to signiicantly lower their bills [4]. Smart devices can help to address this issue by monitoring their consumption and comparing it with that of other consumers. End-user acceptance will ultimately determine the success of many aspects of smart grids [4]. With the massive penetration of distributed generation (DG) units, the current it-and-forget principle of integrating these units into the electric power system is no longer a sustainable option, and a coordinated approach is required. One method to capture the emerging potential of DG, and to cope with the problems caused by the unconventional behavior and increasing penetration of DG, is to take a system approach instead of considering each unit separately [5–8]. In the system approach, the generators and loads are regarded as a subsystem, or microgrid, as depicted in Figure 9.1. Microgrids are small-scale electricity networks, consisting of an aggregation of (converter-interfaced) DG units, (controllable) loads, and storage elements, which are connected to the utility network through a single point of connection, the point of common coupling (PCC), as shown in Figure 9.2. In comparison with a single DG unit, a microgrid has more control lexibility to ensure the system’s reliability and power quality requirements [9]. By cooperating, the microgrid elements can jointly achieve certain objectives, such as beneiting from market participation and Distributed generation units (DG) Power-electronic converter = Stand-alone mode PV panel Wind turbine CHP Fuel cell AC Grid-connected Energy storage mode (Power-electronically interfaced) Utility grid PCC distributed loads FIGURE 9.1 Microgrid with (power-electronically interfaced) loads, storage, and DG units in stand-alone (islanded) or grid-connected mode. 193 Contribution of Microgrids to the Development of the Smart Grid G G Generators MV distribution network MV/HV Transmission HV/MV DG transformer network (HV) transformer MV/LV transformer DG DG PCC PCC LV network DG DG DG LV microgrid LV microgrid FIGURE 9.2 Microgrid connected to utility network through a PCC. reducing electricity costs through lattening of the aggregated load proile [10,11]. Hence, microgrids are regarded as an important element in the successful integration of massive amounts of DG units and renewable energy sources [12,13], and help with power quality issues [11,14,15]. Microgrids also offer the potential to integrate local electrical and thermal networks to achieve coherent generation, storage, and load control, and to beneit from combined heat and power (CHP) generation, which is an important means of eficiency improvement [9]. Smart grids focus on system-wide improvements, with microgrids as the fundamental building blocks of the smart grid [16]. The grid will not become smart at once, but through a gradual evolution; hence, the smart grid will probably emerge as a system of integrated smart microgrids [3,17]. Microgrids are better suited than the overall smart grid to clarify the beneits for consumers, because they are inherently geographically conined entities in which the microgrid elements are aggregated to focus on the speciic beneits within the local entity, while smart grids focus on the more system-wide beneits. Microgrids can provide power to a small community, which can range from a residential district or an isolated rural community, to academic or public communities such as universities or schools, to industrial sites. Industrial parks can be managed as microgrids, for example, to decrease energy dependency, operate as low-carbon business parks,* and increase economic competitiveness (increasing reliability, reducing the purchase of energy, reducing peak consumption). Microgrids can operate either connected to the utility network, in the grid-connected mode, or independently of a main grid, in islanded mode. The key objectives of microgrids are to provide a coordinated integration of DG in the electric power system, improve reliability, allow more eficient use of energy, and become a controllable entity in the network. Hence, the objectives of smart grids and microgrids are closely related. In this sense, microgrids are fundamental building blocks of the smart grid that speciically focus on local issues and beneits, making end users not only consumers but also active participants in the electricity network by enabling them to control and manage (part of) the energy that they will consume. * www.ace-low-carbon-economy.eu. 194 9.2 Smart Grids MICROGRIDS 9.2.1 BeneFIts Microgrids offer signiicant beneits from both the grid perspective and the consumer perspective. A key advantage from the grid’s point of view is that the microgrid elements are collectively regarded as a single controllable unit, enabling the microgrid to deliver the cost beneits of large units [18]. In this way, utilities do not have to consider each unit separately for their system management. While this is not always obvious to utilities, microgrids can improve the economics and reliability of their service, help to meet renewable energy obligations, reduce congestion, improve power quality under high penetration levels of DG, and support demand response. For example, microgrids can optimize a “troublesome” business complex or a university campus, making it more economic and reliable to serve [1]. From the customers’ point of view, irstly, the impact of the microgrid on the reliability of the distribution network is relevant, certainly in the future, with more unpredictable generation and higher consumption (peaks) [19–21]. Reliability is the irst requirement of the electric power system. The cost of unreliability can have a considerable impact on the economy. The outage cost in the United States is estimated to amount to some $79 billion per year, of which momentary outages (<5 min) account for $52 billion [22,23]. The consequences of grid failure are gigantic, because industry as a whole, and the transportation, inancial, communication, and other critical sectors in particular, are largely dependent on a guaranteed and reliable supply of energy [24]. Most of the reliability improvements will need to be introduced in the distribution system, where most outage and power quality issues arise. To increase local reliability, backup gensets and other emergency supplies are often installed near critical loads to provide uninterruptible power supply (UPS) functionality [25]. The costs of these designated devices, which only run for a limited amount of time, are signiicant. In this context, microgrids can provide UPS functionality by exploiting locally available DG, storage, and loads. Secondly, aggregation of mixed assets in a microgrid can also provide considerable economic beneits to the microgrid subsystems by allowing them to participate in the electric power and ancillary services markets. Other beneits that can result from the introduction of microgrids are reduced feeder losses, removal of transmission and distribution bottlenecks, reactive power and local voltage support, various other ancillary services such as heterogeneous power quality and reliability (i.e., differentiation based on the speciic needs of the customers), increased eficiency through CHP, and easier large-scale integration of DG units [19,26–29]. The main reasons for aggregating units in a microgrid are: • To deal with power quality challenges. For example, some customers require a higher power quality than others. In this sense, it is more economically attractive to increase the power quality locally than to upgrade the power quality of the entire power system. With respect to voltage quality in the distribution networks, microgrids are especially beneicial in providing voltage control, as voltage is a local issue in contrast to grid frequency. Contribution of Microgrids to the Development of the Smart Grid 195 • To increase reliability. The power system is exposed to incidents that may affect the security of supply, such as lighting strikes, arc lashes on rainy days, falling trees, man-made faults, and other accidents. Although these incidents are mostly covered by the protection devices, they can affect the reliability of the system. The common defense mechanism is to invest in grid assets. However, grid extensions, for example, in order to avoid the failure of one line leading to a blackout of a part of the system, are increasingly countered by social resistance, following the not-in-my-backyard (NIMBY) principle. Even if these investments are made, natural disasters can completely annihilate parts of the transmission and distribution infrastructure. Hence, even if a certain area is not directly affected by the disaster, the power supply may be interrupted for a long time. Microgrids can restore the local system more quickly. • Optimal utilization of DG. Microgrids can coordinate their local generation, consumption, and storage to use the available (renewable) energy optimally. Also, the connection of DG units in low-voltage (LV) networks can give rise to overvoltage problems. Rather than using the current full on-off control of the DG units, the microgrid can coordinate the loads and generation units to restrict the voltage in the microgrid while capturing as much renewable energy as possible. • Peak load limitations. A microgrid can manage its own generation and load, and hence reduce peak consumption, which can have a signiicant impact on the electricity bill. • Reduction of transmission and distribution losses (with the subsequent energy costs). By locally coordinating its assets, the microgrid can become more independent of the rest of the power system. 9.2.2 oPeratIng modes Thanks to the single connection point between the microgrid and the utility network, microgrids can run in two operating conditions: grid-connected and islanded modes. Depending on the operating mode, the microgrid can tackle one or more of the aforementioned issues. 9.2.2.1 Grid-Connected Mode In the grid-connected mode, the microgrid supports the utility grid while exchanging power with it. In this mode, the microgrid can take economic decisions, such as controlling the power exchange with the utility network dependent on on-site generation, cost, consumption, and prices on the energy markets. For energy management, grid-connected microgrids have the following three assets at their disposal [8]: 1. DER controllers (DG and optional storage) 2. Load management 3. Control of the power exchange with the main grid Generally, the normal condition of the microgrid is grid-connected operation. A irst advantage is that, for the rest of the network, the microgrid can be seen as a controllable 196 Smart Grids entity. This provides signiicant beneits both for the microgrid participants, through scaling and aggregating, and for the distribution network operator, who does not need to consider all units separately [19,27–29]. A second feature is that, because of the single connection point between the microgrid and the utility network, the power exchange can be determined unambiguously and controlled to a predeined value. Hence, in the electricity markets, there is no need to measure and trade the output power of each unit, but solely the aggregated exchange. From the customer’s perspective, the main beneit of a grid-connected microgrid is that it can deliver the required scaling beneits to enable proitable participation in the electricity markets. First, aggregation of several DG units can enable units that would be too small individually (a minimum capacity is required in the markets) to participate in the electricity markets and get better prices for their energy. Secondly, by aggregating different kinds of units, the risks of deviating from predictions of production and consumption can be reduced. Thirdly, microgrids can provide ancillary services to the networks, such as reserve provision or reactive power support, become more self-reliant, and decrease transmission and distribution losses. 9.2.2.2 Islanded Mode Although the normal operation mode of a microgrid is grid connected, it offers the unique characteristic of islanding. By locally generating the consumed energy, and controlling the consumption to cancel the interchange of energy with the main grid, the microgrid can operate in a stable islanded mode. In islanded mode, the microgrid operates independently of the main grid, and, consequently, the DG units, loads, and storage devices are collectively responsible for maintaining the integrity of the microgrid without the assistance of a main grid. There are various reasons to go into islanded mode. Reliability is one of the major concerns of the electric power system, mainly due to the high outage costs. These costs are, however, different for different types of customers, for instance: • Residential consumers mainly face discomfort in the case of a power outage. • Industrial facilities can be faced with large costs due to loss of production. • Hospitals are affected signiicantly in regard to personal safety. Critical systems such as data centers and hospitals have emergency backup power, mostly from diesel-powered generators, but these are ineficient, environmentally unfriendly, and costly. Diesel generators are prone to failure, and diesel delivery can be problematic in the case of large grid outages when the transportation routes are cut off. Also, unlike microgrids, irstly, there is usually a delay between the grid outage and the full operation of diesel generation, and, secondly, diesel backup systems are not built to run for a long time. The transition from a purely diesel-powered backup generator to a hybrid microgrid system, which integrates different on-site sources with energy storage and load control, signiicantly increases the reliability of the backup system [30]. Hence, in order to ensure the continuity of supply for the local load, a (short-term) islanded mode can occur in the case of special situations such as grid faults, outage of the bulk supply, or power quality problems [12,31–33]. It is important to note that the grid fails more often than most people realize. In an average year, according to a 2008 Lawrence Berkeley National Laboratory Contribution of Microgrids to the Development of the Smart Grid 197 (LBNL) study, the power is out for an average of 92 min in the Midwest and 214 min in the Northeast of the United States. Many of these power outages are weather related, such as caused by tornadoes, hurricanes or tropical storms, ice storms, lightning, wind, and rain. According to the Center for Research on the Epidemiology of Disasters, 100–200 million people were affected by weather-related disasters between 1980 and 2009, with economic losses ranging from $50 billion to $100 billion annually. Another possible reason for grid outages is that much of the infrastructure that serves the US power grid is aging. The average age of power plants is over 30 years, with most of these facilities having a life expectancy of 40 years [34]. Electric transmission and distribution system components are similarly aging, with power transformers averaging over 40 years of age and 70% of the transmission lines being 25 years old or older. LBNL statistics show that 80%–90% of all grid failures in North America originate at the distribution level of electricity service [3,23]. For instance, momentary sags and surges on local distribution feeders are a common cause of breakdowns and work stoppages at high-tech facilities such as semiconductor fabs, research labs, and data centers. The remaining 10% of the grid failures stem from generation and transmission problems, which can cause wider-scale outages affecting larger numbers of customers. Microgrids are able to deal with these power outages by operating in islanded mode. In this way, microgrids offer a potential improvement of the reliability, quality, and costs of the system. Hence, microgrids are a good option in applications in which electrical energy must be guaranteed at all times, for example, hospitals, server centers, and industrial facilities. However, they are a step beyond either emergency backup systems or stand-alone solar panel arrays. They use special software and power electronics to integrate multiple sources of power and energy storage to provide electricity around the clock, even when the sun is not shining or regulations limit the use of diesel generators. Other than dealing with rare (weather) incidents, islanding is also useful when the main grid is not robust enough due to factors such as long distances from the main grid [35]. In this way, in large, highly dispersed countries such as Canada, the United States, and Japan, major efforts into microgrid research are being made. Microgrids can also switch into islanded mode due to planned maintenance operations. In conclusion, true microgrids are much more than backup systems. They include real-time control to match the microgrid generation, storage, and consumption. In the normal, grid-connected mode, they interact with the utility network in order to deliver power when the cost is higher and consume power from the grid when the cost is lower. In islanded mode, they can achieve independence from the main grid. Examples 1. Reliability: Hurricane Sandy (United States) In October 2012, superstorm Sandy reached the eastern shores of the United States. This led to an electricity outage in many areas. Over two million connections were de-electriied, of which 750,000 were in New York City. Two days after the superstorm, six million American people still had no electricity, of whom two-thirds lived in the states of New York and 198 Smart Grids New Jersey, according to the U.S. Department of Energy. President Obama stressed that restoration of the electricity supply was a priority. However, 1 week after Sandy, 1.4 million American people still had no electricity, according to the Department of Energy in Washington. This shows the vulnerability of the electricity grid when faced with extreme weather circumstances and the dificulty of restoring this massive system. Microgrids offer a unique opportunity to provide a more reliable energy supply. Although many places had no electricity, a number of colleges, including Princeton and New York Universities, relied on their microgrids to keep power lowing to vital services during the Sandy storm. Princeton University typically gets its electric power from both the local grid and on-site generation. During Hurricane Sandy, Princeton was able to switch off the grid and power part of the campus with about 11 MW of local generation. 2. Reliability: 2011 Earthquake and Tsunami (Tohoku, Japan) After the Great East Japan Earthquake and Tsunami in 2011, around 4.4 million households in northeastern Japan were left without electricity. In some areas, the outages lasted for weeks. Likewise, the port city of Sendai was heavily hit, and faced 2 days of outage. But in one small section of the city the lights stayed on. The Tohoku Fukushi University had an experimental microgrid project fed by three types of energy generators: fuel cells, solar panels, and natural gas microturbines. The microgrid system was able to island and operate as an independent energy source, making it less vulnerable to the problems in the larger system. For 2 days, the microgrid continued to provide reliable power to a number of loads, including a hospital, while the larger macrogrid had halted power supply. Even if a certain area is not directly affected by a natural disaster, the power supply may be interrupted for weeks or even months if the connection to the utility network is interrupted by the event. As the microgrid is not dependent on the utility grid, the immediate construction of microgrids appears feasible in some areas [36], for example, in Japan after the tsunami. Microgrids can be planned and assembled in a short time. Furthermore, a power system disturbance can trigger a cascading outage, in which microgrids are virtually unaffected. 3. Reliability: Jail (United States) At the Santa Rita Jail in California, a local microgrid project was oficially launched in 2012. Since mid-1990, the jail has made a major effort in eficiency upgrades. This has been pushed by the local government to protect the jail against rolling blackouts and spikes in the electricity prices, due to the California electricity crises in 2000–2001. Therefore, the jail has a signiicant number of photovoltaic (PV) panels, a fuel cell power plant, and some small wind turbines as on-site generators, as well as some diesel generators and batteries for backup. Next to eficiency improvements (lower costs), power reliability is a primary concern for the jail to ensure that large amounts of electricity are constantly kept available. Now, the facility forms a microgrid that can operate in islanded mode. The system keeps the power on, for example, when storms take down the grid, which is essential for safety at the maximum security facility. 4. Integration of DG: Jail (United States) In Alcatraz, near San Francisco, there is a similar story. A feasibility study conducted by researchers from the University of Washington indicated that installing an undersea transmission cable from Alcatraz Island to Contribution of Microgrids to the Development of the Smart Grid 5. 6. 7. 8. 199 San Francisco would be too costly. Hence, at irst, some diesel generators were responsible for feeding the island. However, as the price of diesel began to climb, and the cost of solar PV fell, developing a state-of-the-art microgrid appeared attractive. Thanks to the combination of a PV installation and battery packs, it is expected that the solar-powered microgrid will reduce the running time of the diesel generators from 100% to 40%. Power quality: Industrial areas In highly sensitive industrial areas, for example, chemical or semiconductor manufacturing facilities, reliable power of high quality is required. High power quality is also very important in modern health care [37]. Electronic equipment is commonly designed for power supplies with speciic voltage and current characteristics. When voltage or current levels deviate from the speciied quality standards, electronic equipment can be damaged or fail, which may cause high-tech medical equipment to malfunction, resulting in extended patient discomfort, misdiagnoses, increased equipment downtime and service costs, and even life-threatening situations [37]. Health-care facilities with extremely sensitive electronic equipment and high reliability needs are increasingly recognizing the limitations of traditional solutions to costly power quality problems. Microgrids can be tailored to the speciic needs of these end users and provide so-called premium power customers with inely calibrated power lows [37]. They can provide an integrated power supply with different levels of power quality. Hence, instead of alleviating the power quality levels of the entire grid, microgrids can ensure the required power quality locally at a more acceptable cost. Maintenance: Hydro-Québec (Canada) The islanded microgrid of the Canadian electric utility Hydro-Québec was the result of an urgent need for the maintenance of the transmission line feeding a substation named Senneterre, to which a privately owned thermal power plant (Boralex) was connected. The substation feeds three distribution lines, serving 3000 customers in the municipality of Senneterre and its surrounding area. Standing on wooden structures, this line, more than 55 years old, required urgent replacement of its angle and stop portals [38]. This type of maintenance can only be performed on a de-energized line [38]. To avoid service interruption to customers during the restoration of the transmission line, the Boralex power plant was used for islanding the microgrid. The irst maintenance work of the transmission line began in fall 2006 [38]. Mobile: Military bases The U.S. Department of Defense gives high priority to mobile tactical microgrids. These microgrids need to be highly modular. They need to be deployed very quickly (in a matter of days), and afterward they need to be deconstructed and moved to another location. The ability to rapidly construct and deconstruct is not the only reason for using the microgrid concept in military systems. Microgrids can signiicantly attenuate the cost of the transport of oil and gas to military bases, which are often stationed in remote locations. Reliability and energy cost: BC Hydro Boston Bar substation (Canada) In areas with frequent loss of electrical power, for example, in countries with a weak grid infrastructure, very long lines, or harsh weather circumstances, islanded systems can reduce the loss of electrical power. An example is deployed by the Canadian electric utility BC Hydro. In the Boston Bar area, 200 Smart Grids BC Hydro can switch to the islanded operating mode of the interconnected feeder during outages in its radial transmission line. This line had proved to be insuficiently reliable because it is a single radial supply with a long feeder, which is, moreover, subjected to frequent outages because of falling rocks and trees, mud, storms, and motor vehicles. By including the ability to island, the reliability of the system has signiicantly increased. 9. Dealing with the growing demand for electricity: Independence The increasing use of electricity* demands the construction of new power plants and power lines. However, due to the common NIMBY attitude, these investments are being signiicantly delayed. Microgrid structures can ease this attitude, as they limit the construction of new lines, and, when they are installed, the beneit to the local consumer is obvious, as opposed to the case of building long transmission lines. Hence, they can relieve distribution and transmission expansion issues. 9.2.2.3 Remote Islanded Mode Although not strictly according to the deinition of a microgrid, an islanded microgrid can also exist in the case of remote electriication, where no main grid is available due to, for example, geographical issues. According to the World Bank’s 2010 development report, 1.6 billion people in developing countries do not have access to electricity [39]. The most important reason is the extensive investment needed to install power lines across large distances, across rough terrain, or both, in order to extend the electricity supply to a few people. It is recognized that electricity is a key driver of economic growth and combating poverty. Hence, islanded microgrid projects provide great opportunities to realize sustainable human development. According to the International Energy Agency (IEA 2009), 83% of people who do not have access to electricity live in rural areas [40], for example, over one-third of India’s rural population. The Indian government has set the target of providing all Indian households with electricity in the near future [41]. If a conventional grid connection is not cost effective to accomplish this, consideration is given to decentralized electric power with local distribution, that is, islanded microgrids, with an example in the Sundarbans Islands region in India [41]. Two examples of remote islanded microgrids are given below. 1. Serving loads in remote locations that lack infrastructure (Uganda, India) Uganda has a limited supply of electricity, with less than 10% of the Ugandan population having access to electric power. Rural areas in particular are not yet electriied, because of the high costs of the grid extensions this would require. Without electricity, economic growth is constrained, limiting the chances of combating poverty in these regions. Also, hospitals are only able to deliver a limited service. A solution can be provided by small, decentralized (hydro)power stations. Therefore, in Uganda, several projects are running, such as the Suam project, which enables the use of small microgrids, often fed by hydropower, for electriication. * For energy eficiency reasons, both transportation and heating are becoming more and more electriied, for example, through the use of electrical vehicles and heating pumps. Contribution of Microgrids to the Development of the Smart Grid 201 Remote microgrids are also suitable for electrifying Indian villages that do not yet have power. Nearly 400 million Indian people, most living in rural communities, lack access to the grid. To access electricity, these people need to go to the city. They also generally use kerosene-illed lamps, which are unhealthy and unsafe, instead of electric lighting. Microgrids offer a great opportunity to deliver electricity to these communities. 2. Serving islands (Greece) Many of the Greek Aegean islands are not connected to the mainland grid, and thus rely on polluting, costly, and ineficient diesel generation to fulill their electricity demands. Fortunately, most islands have optimal wind potential and solar resources. The exploitation of these resources would provide a cheaper alternative to the current diesel generation and result in signiicant savings and lower electricity prices. 9.2.3 smart mIcrogrIds Realistically, the growth of the smart grid will not be a revolution. Instead, a gradual evolution is expected [3,17]. Because of their lexibility and scalability, microgrids are likely to play a key role in the evolution of the smart grid [42], with smart microgrids as small pilot versions (building blocks) of the future power system [16]. Hence, the smart grid will probably emerge as a system of integrated smart microgrids [3]. Smart microgrids are microgrids combined with an overlying intelligence scheme. The intelligent software implemented in the microgrid can: • Contain energy management systems • Monitor the system (energy supply, storage, and demand) and actively intervene in consumption/generation • Identify and maximize energy eficiency opportunities • Use an extra communication and sensor layer to maximize cost savings and reduce carbon emissions • Generate more active participation by consumers A balance between the cost of incorporating intelligence into the grid and the subsequent beneits needs to be achieved. Furthermore, the infrastructure and control centers for smart grids can be hard to implement on a large scale, and can take many years. Rather than investing in incorporating intelligence on a large scale, investing in small smart microgrids (such as business areas) can be done at a lower cost and more quickly. Also, a high level of intelligence built into all parts of the system is not necessary: different areas require different levels of smartness. The areas that allow more and beneit most from high levels of smartness of the system, such as areas with high penetration of DG, can it directly into the microgrid concept. In these microgrids, smart features can be installed more quickly than in the rest of the network, with local levels of intelligence higher than average. In this way, the smart microgrids can enable a new energy strategy while avoiding the cost of making the whole system smart. 202 9.3 9.3.1 Smart Grids MICROGRIDS IN THE SMART GRID PARADIGM sImIlar BeneFIts and aIms Essentially, the goals of microgrids and smart grids are the same: to minimize costs, meet the growing demand, integrate more sustainable generation resources by maximizing generation assets, and increase the eficiency of the power system. The main difference between microgrids and smart grids is that the microgrid is a building block of the overlying smart grid. Microgrids aim at local control and optimization of their local assets, with a contribution to the higher-level utility network. Because of this local scope, microgrids directly focus on the value proposition of their customers. In contrast, as smart grids focus on more system-wide issues, an important challenge for smart grids is to make the value of the achieved beneits clear to customers. 9.3.2 mIcrogrId control As in the conventional power system, microgrid control can be conceived as a hierarchical control structure. Microgrid control can be divided into: • Local primary control • Supporting (smart grid) services in an overlying (secondary/tertiary) control; smart microgrid control 9.3.2.1 Smart Microgrid Control Smart microgrid control is conceived as a high-level (secondary) control structure, which deines new set points for the local controllers. Smart microgrid control can beneit fully from smart grid features, such as communication and metering data. With respect to this control level, the microgrid is a building block of the smart grid. Like the primary controller, the secondary controller can tackle technical issues, but it can deliver more functionalities, such as achieving economic, societal, and environmental objectives. The secondary controller can achieve further optimization of the microgrid (e.g., fuel savings), coordinate the primary control actions, restore preagreed consumption patterns, restore the voltage to its nominal value (amplitude, frequency, or both), and control the import and export of microgrid energy. For instance, the secondary or tertiary controller can address the social issues of the producers if the physical position in the grid impacts the amount of power that can be injected (and hence also impacts the inancial reimbursement for the energy delivered to the grid). 9.3.2.2 Local Microgrid Control The local microgrid control is a primary control level operating without communication. In this sense, it goes beyond the smart grid control, as it is faster and depends on local measurements only. In islanded microgrids, the local microgrid controller is responsible for the reliability of the system. In grid-connected microgrids, the local controller can contribute to achieving a proper voltage quality with minimal communication effort. The contribution of the local microgrid controller to easing smart grid communication is clear, as the local controller can achieve fast and reliable control actions, not depending on communication and remote measurements. Contribution of Microgrids to the Development of the Smart Grid 203 An overview of various primary control strategies is given in [43]. For (extended) islanded microgrids, the droop concept is most promising to ensure reliable system operation. The active power/grid voltage (P/V) droop concept in general and the voltage-based droop (VBD) controller [44] speciically are highlighted in this chapter. Droop controllers in microgrids are derived from the well-known active power/grid frequency (P/f) droop control principle incorporated into the large central synchronous generators in the conventional electric power system. While the P/f droop control method works well in a microgrid with mainly inductive line impedances, it leads to concern when implemented on an LV microgrid, without signiicant inertia, and where the line resistance should not be neglected [9]. In the case of mainly resistive lines, the active power (P) is linked with the voltage amplitude difference (V), while the reactive power (Q) is linked with the phase angle difference and hence frequency (f). This leads to P/V and Q/f droops as opposed to the conventional P/f and Q/V droops [45–47]. The active and reactive powers are measured and drooped to obtain the root-mean-square (RMS) voltage and its frequency (ω = 2πf), respectively: Vg,i = Vg,ref − K P ( Pi − Pi,ref ) (9.1) ωi = ωref + KQ ( Qi − Qi,ref ) (9.2) where KP and KQ are the droop coeficients. The VBD control of [44] is a variant of the P/V droop control that focuses on the optimal integration of renewables in the network and also incorporates a direct current (dc)-bus controller. In the VBD control, the P/V droop controller is divided into two droop controllers, and constant-power bands are included, as depicted in Figure 9.3. Further details of these control loops are given in [44]. The parameter 2b determines the width of the constant-power band. When the terminal voltage of the DG unit is in the constant-power band, that is, (1 − b)Vg,ref < Vg < (1 + b)Vg,ref, the DG unit delivers its reference power. Otherwise, the power of the DG unit is changed. A distinction is made between dispatchable and less dispatchable DG units. For dispatchable DG units, analogously to central large generators, Pref represents the scheduled power that can be determined in the electricity markets. In this case, the value of b is small, such that the unit reacts to small variations of the grid voltage. For less dispatchable DG units, such as many renewables, a wider constant-power band is used. In this way, these units do not react to all load variations, but change their output power only in the case of more extreme voltage variations compared with dispatchable DG units. These extreme voltages only occur when the power limits of the dispatchable DG units are nearly reached. Consequently, an optimized integration of renewables is possible, because the VBD control prioritizes the power changes of the units, by setting different values of b, without the need for communication. In these units, Pref can, for example, represent the instantaneous maximum power point of a wind turbine. Also, as voltage is a local parameter, the voltage limits are sometimes already reached in areas where the ratio of dispatchable to less dispatchable DG units is low. The current it-and-forget strategy of integrating DG solves these voltage problems by turning the DG units off. In contrast, the VBD controller uses soft power 204 Smart Grids Vg/Vdc droop controller Vg P/Vg droop controller Pdc Vg Vdc Pdc Pdc,nom Vg,nom bVg,nom bVg,nom Vdc Vg Vg,nom Vdc,nom Q/f droop controller Q vg,droop f f ∫ sin vg,ref + Duty ratio VSI Voltage controller α fnom – Q Qnom ig zv FIGURE 9.3 VBD controller and virtual impedance loop. curtailment to capture more of the renewable energy. In this way, voltage problems can be overcome, as the renewables also take part in the voltage control. As shown in Figure 9.3, the Vg/Vdc and Q/f droop controllers, which are digital controllers, determine the droop voltage vg*,droop,k = Vg,k sin ( α k ) (k is the discrete time instance). Discrete values are used because pulse width modulation with sampling period Ts is used in the converter. The amplitude Vg of the droop voltage is obtained from the Vg/Vdc droop controller and the phase angle is obtained from the frequency f in the Q/f droop controller. Together, they determine v*g,droop,k = Vg,k sin ( α k −1 + 2πfkTs ) (9.3) Figure 9.3 also shows the virtual output impedance loop. A resistive output impedance zv = Rv is chosen, as this provides more damping in the system [48] and complies with the power control strategies of the loads and generators, where the active power is changed based on the grid voltage: vg* = vg*,droop − Rvig where: v*g is the reference voltage for the voltage controller * vg,droop is the voltage obtained by the VBD controller ig is the current injected into the microgrid, as illustrated in Figure 9.4 (9.4) 205 Contribution of Microgrids to the Development of the Smart Grid Rest of microgrid LC filter VSI P Line impedance Primary energy source Pdc Pc T1 ig T4 Vdc Pdc L Load C + Cdc vg T2 T3 FIGURE 9.4 DG unit connected to the microgrid. Parameters (vg, ig, Vdc, P, Pdc). This chapter illustrates the contribution of the VBD control in a grid-connected microgrid to operating the grid in a smarter manner. 1. Market operation: The amount of power export/import is preagreed in the markets, for example, on a day-ahead basis. For example, when the electricity price is high, the microgrid will import as little power as possible from the grid, or will even export power. One requirement for market participation of a microgrid is thus that the power exchange between the microgrid and the utility networks can be controlled to an agreed value. This is conventionally achieved by using a microgrid central controller that communicates new set points to all units. However, this may lead to possible delays and a signiicant communication burden. Therefore, in [49], the usage of a so-called smart transformer is discussed. The smart transformer is a tapchanging transformer, connected at the PCC of the microgrid, as depicted in Figure 9.5. The transformer is smart in the sense that the control strategy for changing the microgrid-side voltage is able to control the power exchange to the set value by changing its taps. In medium-voltage networks, on-load tap-changing transformers are sometimes already available; hence, controlling these as smart transformers requires few modiications. In the lower-voltage networks, most PCC transformers are manual tap-changing transformers, from which the voltage can only be controlled off-line and PCC PCC transformer Tap changer Microgrid PPCC ··· PCC Utility network VPCC DG units Loads ··· FIGURE 9.5 Smart transformer located at the PCC of a microgrid. 206 Smart Grids not automatically. This puts a signiicant stress on electrical grids that face increased penetration of DG units. Historically, the planning of LV grids is based on a worst-case scenario, ensuring that, in the case of maximum consumption, the voltage does not drop below the lower voltage limit. Therefore, many (manual) tap-changing transformers at the beginning of the LV lines are set somewhat above the nominal voltage. However, with the increasing presence of DG units, the risk of overvoltages becomes higher. Also, planning becomes more dificult because of the larger voltage variations (e.g., sunny daytime vs. nighttime). Hence, the ability to change the tap settings automatically becomes more useful, as it is more effective, faster, and cheaper to implement than the conventional approach of investing in more lines and larger transformers. Therefore, in future, it may be beneicial to install an on-load tap changer (OLTC). Moreover, it is well known that many of the assets in distribution networks are reaching the end of their life and will have to be replaced in the coming years. Hence, the manual transformers can gradually be upgraded to OLTCs. The OLTC transformer with smart control strategy, that is, the smart transformer, controls the PCC power (PPCC) by changing the microgrid-side voltage (VPCC). If the microgrid elements are equipped with the VBD control strategy, these elements automatically react by changing their input/output power depending on the grid voltage. For example, when the PCC power exported to the utility network is lower than its preagreed value (PPCC,ref ), the smart transformer will lower VPCC. The microgrid DG units react to this voltage drop by increasing their output power P, hence increasing PPCC. In this way, the PCC power can be controlled by the smart transformer, without the need to communicate new set points to all grid elements as they automatically react. A second advantage is that, in this way, a virtual islanded mode is achieved. The utility grid is not seen as a slack bus, but is conceived as a constant-power load or generator, that is, generating the power agreed upon in the markets. The smart transformer can facilitate distribution network management, as the distribution network operator only needs to communicate to the microgrid central controller, which, instead of facing all of the microgrid elements separately, can rely on an automatic response of the microgrid elements to a changing grid voltage induced by the smart transformer. 2. Voltage control in LV networks: Although developed for islanded microgrids, the VBD controller can be used in grid-connected microgrids as well [50]. Moreover, it offers signiicant advantages, irstly in the voltage control (avoiding overvoltage) and secondly in avoiding on-off oscillation of DG units (increasing the capture of renewable energy). Firstly, a well-known issue when connecting DG units in LV systems is possible overvoltage due to the injection of active power by the DG unit. Generally, in LV networks, other than the system planning, there is little or no means of voltage control. On the other hand, the extensive penetration of small DG units into these networks offers a unique opportunity to use their controllability for voltage control in the network. 207 Contribution of Microgrids to the Development of the Smart Grid The conventional approach to voltage control is by injecting or consuming reactive power, mirroring the voltage control in the transmission network. In [51], the required amount of reactive power Q that a DG unit should consume to erase its impact on the voltage is calculated. A minimal impact is achieved when the DG unit consumes (minus sign) reactive power according to: Q=− R P, X (9.5) where R and X are the line parameters and P is the generated active power of the DG unit. Here, a uniform distribution of load along the feeder and a constant resistance and reactance per unit length are assumed. Still, a small impact remains due to the power losses associated with the transport of power over the network, which are not included in Equation 9.5. In this chapter, LV networks are considered, which are mainly resistive; hence, they have a high R/X, generally greater than 3. According to Equation 9.5, in these networks with R ≫ X (LV microgrids), a DG unit needs to consume a lot of reactive power to minimize the voltage rise due to its active power injection into the grid. Conversely, reactive power consumption has only a small effect on the voltage of the system. Hence, it will predominantly increase the network losses. Furthermore, such large amounts of reactive power generally cannot be consumed by the generators without signiicant overrating. The VBD control achieves an automatic soft curtailment of the DG units, but because of the usage of constant-power bands, this is done in a predeined priority order. The renewables are softly curtailed only when absolutely necessary. In this way, the DG units can control the system voltage by using the VBD controller, without the need for designated devices for voltage control. Currently, renewables fully turn off when their terminal voltage exceeds a certain voltage level. As a single unit can have a large impact on the grid voltage, this may lead to a signiicant voltage drop, such that the unit may 1,000 245 240 0 200 400 600 800 1,000 1,200 Output power (W) Voltage (V) 250 0 1,400 FIGURE 9.6 On/off control of DG leading to grid oscillations: measurements of PV panels in Oostende, Belgium (— = DG unit 1, - - - = DG unit 2, ––––– = VPCC). 208 Smart Grids TABLE 9.1 Renewable Energy Capture: Comparison of On/Off Control versus VBD Control DG 1 (E [J]) DG 2 (E [J]) DG 3 (E [J]) On-Off Control VBD Control Gain (%) 9794 9122 4002 11,539 9402 4741 18 3 18 turn on again. This may lead to on-off oscillations of renewables in LV networks. An example of a measurement of such an oscillation is depicted in Figure 9.6. The on-off oscillations can also occur between multiple units that turn on and off in turn. In [50], it is illustrated that the grid-connected VBD control can, irstly, avoid these on-off oscillations, and, secondly, subsequently achieve a higher renewable energy capture. A result from this paper is depicted in Table 9.1. The output energy for three DG units, which are connected to a feeder with multiple RL loads, is shown. The details of the coniguration are given in [50]. The on-off control of the units leads to on-off oscillations in the system, which, in turn, lead to a lower renewable energy capture than with VBD control. Hence, although soft curtailment instantly results in a loss of the available renewable energy, it may increase the total energy capture by avoiding on-off oscillations, when the priority is set, for example, by using the VBD concept. 9.4 CONCLUSION This chapter illustrates that microgrids can signiicantly contribute to developing the smart grid. Some examples of microgrids in grid-connected and islanded modes are given, including the beneits they bring to the smart grid. The microgrid control consists of two layers: a fast and automatic primary controller and an overlying slower controller. The latter can take full advantage of the communication and metering of the smart grid, and hence its fully into the scope of the smart grid. The primary controller contributes signiicantly to the smart grid in the sense that it facilitates the smart grid control and eases the communication burden. This is illustrated through a discussion of the contribution of the VBD control, irstly, in controlling the power exchange between microgrid and utility networks with a minimum of communication, and, secondly, in voltage control. ACKNOWLEDGMENTS The work of Dr. Ir. T. Vandoorn is inancially supported by a Fellowship of the FWO-Vlaanderen (Research Foundation Flanders, Belgium). This research has been carried out in the frame of the Interuniversity Attraction Poles Programme initiated by the Belgian Science Policy Ofice (IAP-VII-02). Contribution of Microgrids to the Development of the Smart Grid 209 REFERENCES 1. K. Iniewski (ed.), Smart Grid Infrastructure & Networking. McGraw-Hill Professional, New York, 2012. 2. C. W. Gellings, The Smart Grid, Enabling Energy Eficiency and Demand Response. The Fairmont Press, Lilburn, 2009. 3. H. Farhangi, The path of the smart grid, IEEE Power and Energy Magazine, 8(1), 18–28, 2010. 4. DNV KEMA, Global inventory and analysis of smart grid demonstration projects, Technical Report, October 22, Netbeheer Nederland, 2012. 5. R. H. Lasseter, J. H. Eto, B. Schenkman, J. Stevens, H. Vollkommer, D. Klapp, E. Linton, H. Hurtado, and J. Roy, CERTS microgrid laboratory test bed, IEEE Transactions on Power Delivery, 26(1), 325–332, 2011. 6. M. Popov, H. Karimi, H. Nikkhajoei, and V. Terzija, Modeling, control and islanding detection of microgrids with passive loads, in Proceedings of 14th International Power Electronics and Motion Control Conference (EPE/PEMC), September 6–8, Ohrid, Republic of Macedonia, 2010. 7. A. Engler, O. Osika, M. Barnes, N. Jenkins, and A. Arulampalam, DB1 local micro source controller strategies and algorithms, European Commission, February, 2004. www.microgrids.eu/micro2000. 8. J. Driesen and F. Katiraei, Design for distributed energy resources: Microgrid planning and architectures for improved reliability and integration, IEEE Power and Energy Magazine, 6(3), 30–40, 2008. 9. Y. W. Li and C. N. Kao, An accurate power control strategy for power electronics interfaced distributed generation units operating in a low voltage multibus microgrid, IEEE Transactions on Power Electronics, 24(12), 2977–2988, 2009. 10. R. H. Lasseter, A. Akhil, C. Marnay, J. Stephens, J. Dagle, R. Guttromson, A. Meliopoulous, R. Yinger, and J. Eto, The CERTS microgrid concept, white paper on integration of distributed energy resources, in California Energy Commission, Ofice of Power Technologies— U.S. Department of Energy, LBNL-50829, www.certs.lbl.gov, April, 2002. 11. R. H. Lasseter and P. Paigi, Microgrid: A conceptual solution, in Proceedings of IEEE Power Electronics Specialists Conference (PESC 2004), June, Aachen, Germany, 2004. 12. M. Bollen, J. Zhong, O. Samuelsson, and J. Björnstedt, Performance indicators for microgrids during grid-connected and island operation, in Proceedings of IEEE PowerTech Conference, June 28–July 2, Bucharest, Romania, 2009. 13. A. Vaccaro, M. Popov, D. Villacci, and V. Terzija, An integrated framework for smart microgrids modeling, monitoring, control, communication, and veriication, Proceedings of IEEE, 99(1), 119–132, 2010. 14. N. Pogaku, M. Prodanović, and T. C. Green, Modeling, analysis and testing of autonomous operation of an inverter-based microgrid, IEEE Transactions of Power Electronics, 22(2), 613–625, 2007. 15. M. A. Pedrasa and T. Spooner, A survey of techniques used to control microgrid generation and storage during island operation, in Proceedings of the 2006 Australian Universities Power Engineering Conference (AUPEC2006), December 10–13, 2006. 16. M. Savaghebi, A. Jalilian, J. C. Vasquez, and J. M. Guerrero, Secondary control scheme for voltage unbalance compensation in an islanded droop-controlled microgrid, IEEE Transactions on Smart Grid, 3(2), 797–807, 2012. 17. M. Kezunovic, V. Vittal, S. Meliopoulos, and T. Mount, The big picture: Smart research for large-scale integrated smart grid solutions, IEEE Power and Energy Magazine, 10(4), 22–34, 2012. 18. A. G. Tsikalakis and N. D. Hatziargyriou, Centralized control for optimizing microgrids operation, IEEE Transactions on Energy Conversion, 23(1), 241–248, 2008. 210 Smart Grids 19. P. M. Costa and M. A. Matos, Assessing the contribution of microgrids to the reliability of distribution networks, Electric Power Systems Research, 79(2), 382–389, 2009. 20. K. De Brabandere, Voltage and frequency droop control in low voltage grids by distributed generators with inverter front-end, PhD Dissertation, KU Leuven, Leuven, Belgium, October, 2006. 21. R. H. Lasseter and P. Piagi, Control and design of microgrid components, Power Systems Engineering Research Center (PSERC), Technical Report, January, 2006. 22. X. Tan, Q. Li, and H. Wang, Advances and trends of energy storage technology in microgrid, International Journal on Electrical Power and Energy Systems, 44(1), 179–191, 2013. 23. K. H. LaCommare and J. H. Eto, Understanding the cost of power interruptions to U.S. electricity consumers, Ernest Orlando Lawrence Berkeley National Laboratory, Technical Report, 2004. 24. S. M. Amin and B. F. Wollenberg, Toward a smart grid, IEEE Power and Energy Magazine, 3(5), 34–41, 2005. 25. C. Marnay and R. Firestone, Microgrids: An emerging paradigm for meeting building electricity and heat requirements eficiently and with appropriate energy quality, Environmental Energy Technologies Division, US Department of Energy, presented at the European Council for an Energy Eficient Economy 2007 Summer Study, Technical Report, June 4–9, La Colle sur Loup, France, 2007. 26. T. C. Green and M. Prodanovíc, Control of inverter-based micro-grids, Electric Power Systems Research, 77(9), 1204–1213, 2007. 27. H. Jiayi, J. Chuanwen, and X. Rong, A review on distributed energy resources and microgrid, Renewable and Sustainable Energy Reviews, 12(9), 2472–2483, 2008. 28. C. Marnay and G. Venkataramanan, Microgrids in the evolving electricity generation and delivery infrastructure, in Proceedings of IEEE PES General Meeting, June 18–22, Montréal, QC, 2006. 29. P. Basak, A. K. Saha, S. Chowdhury, and S. P. Chowdhury, Microgrid: Control techniques and modelling, in Proceedings of the 44th International Universities’ Power Engineering Conference (UPEC 2009), September 1–4, Glasgow, Scotland, 2009. 30. T. Casey, The spiders microgrid project, 2013 February 12. Available at: www. cleantechnica.com. 31. European Commission, New ERA for electricity in Europe. Distributed generation: Key issues, challenges and proposed solutions, 2003. Available at: www.smartgrids.eu. 32. F. Katiraei, M. R. Iravani, and P. W. Lehn, Micro-grid autonomous operation during and subsequent to islanding process, IEEE Transactions on Power Delivery, 20(1), 248–257, 2005. 33. H. Mohamad and P. Crossley, Islanded operation of UK radial distribution earthing strategy, in Proceedings of the 44th International Universities’ Power Engineering Conference (UPEC 2009), September 1–4, Glasgow, UK, 2009. 34. R. J. Campbell, Weather-related power outages and electric system resiliency, Congressional Research Service Report for Congress, Technical Report, August 28, 2012. 35. E. Planas, A. Gil-de-Muro, J. Andreu, I. Kortabarria, and I. Martínez de Alegría, General aspects, hierarchical controls and droop methods in microgrids: A review, Renewable and Sustainable Energy Reviews, 17(5), 147–159, 2013. 36. Siemens, Microgrid white paper, 2011. Available at: www.energy.siemens.com. 37. Center for Energy, Marine Transportation and Public Policy at Columbia University, Microgrids: An assessment of the value, opportunities and barriers to deployment in New York state, New York State Energy Research and Development Authority (NYSERDA), Technical Report, September, 2010. Contribution of Microgrids to the Development of the Smart Grid 211 38. M. Gauthier, C. Abbey, F. Katiraei, J.-L. Pepin, M. Plamondon, and G. Simard, Planned islanding as a distribution system operator tool for reliability enhancement, in Proceedings of 19th International Conference on Electricity Distribution, May 21–24, Vienna, Austria, 2007. 39. The World Bank, Washington DC, World development report 2010, development and climate change, 2010. Available at: http://web.worldbank.org. 40. International Energy Agency (IEA), Access to electricity (WEO-2011), 2011. Available at: www.worldenergyoutlook.org/. 41. I. Mitra, T. Degner, and M. Braun, Distributed generation and microgrids for small island electriication in developing countries: A review, SESI Journal, Solar Energy Society of India, 18(1), 6–20, 2008. 42. N. Lidula and A. D. Rajapakse, Microgrids research: A review of experimental microgrids and test systems, Renewable and Sustainable Energy Reviews, 15(1), 186–202, 2011. 43. T. L. Vandoorn, J. D. M. De Kooning, B. Meersman, and L. Vandevelde, Review of primary control strategies for islanded microgrids with power-electronic interfaces, Renewable and Sustainable Energy Reviews, 19, 613–628, 2013. 44. T. L. Vandoorn, B. Meersman, L. Degroote, B. Renders, and L. Vandevelde, A control strategy for islanded microgrids with dc-link voltage control, IEEE Transactions on Power Delivery, 26(2), 703–713, 2011. 45. A. Engler and N. Soultanis, Droop control in LV-grids, in Proceedings of International Conference on Future Power Systems, November 16–18, The Netherlands, 2005. 46. A. Engler, O. Osika, M. Barnes, and N. Hatziargyriou, DB2 evaluation of the local controller strategies, January 2005. www.microgrids.eu/micro2000. 47. C. Sao and P. Lehn, Control and power management of converter fed microgrids, IEEE Transactions on Power Systems, 23(3), 1088–1098, 2008. 48. W. Yao, M. Chen, J. M. Guerrero, and Z.-M. Qian, Design and analysis of the droop control method for parallel inverters considering the impact of the complex impedance on the power sharing, IEEE Transactions on Industrial Electronics, 58(2), 576–588, 2011. 49. T. L. Vandoorn, J. D. M. De Kooning, B. Meersman, J. Guerrero, and L. Vandevelde, Voltage-based control of a smart transformer in a microgrid, IEEE Transactions on Industrial Electronics, 60(4), 1291–1305, 2013. 50. T. L. Vandoorn, J. D. M. De Kooning, B. Meersman, and L. Vandevelde, Voltage-based droop control of renewables to avoid on-off oscillations caused by overvoltages, IEEE Transactions on Power Delivery, 28(2), 845–854, 2013. 51. M. H. J. Bollen and F. Hassan, Integration of Distributed Generation in the Power System, IEEE Press Series on Power Engineering, Wiley, Hoboken, NJ, 2011. 10 Microgrids Mietek Glinkowski, Adam Guglielmo, Alexandre Oudalov, Gary Rackliffe, Bill Rose, Ernst Scholtz, Lokesh Verma, and Fang Yang CONTENTS 10.1 Introduction to Microgrids ........................................................................... 214 10.1.1 What is a Microgrid? ........................................................................ 214 10.1.2 Why Microgrids? .............................................................................. 214 10.1.3 Microgrid Market ............................................................................. 216 10.2 Microgrid Design .......................................................................................... 217 10.2.1 Optimal Microgrid Topology ........................................................... 217 10.2.1.1 Traditional Distribution Network Topologies .................... 218 10.2.1.2 Size and Location of Distributed Energy Resources ......... 220 10.2.2 Microgrid Control ............................................................................. 220 10.2.2.1 Centralized Approach ........................................................ 221 10.2.2.2 Decentralized Approach .................................................... 222 10.2.3 Microgrid Protection ........................................................................ 223 10.2.3.1 Adaptive Protection ........................................................... 223 10.2.3.2 Fault-Current Source.......................................................... 226 10.2.3.3 Fast Static Switch ............................................................... 226 10.2.3.4 Communication Architectures and Protocols for Adaptive Protection ........................................................... 227 10.2.4 Selecting the Right Communications Technology ........................... 227 10.2.5 Direct Current Microgrids ................................................................ 228 10.2.5.1 Differences Between ac and dc for Microgrids ................. 230 10.2.5.2 Examples of dc Microgrids ................................................ 231 10.2.5.3 Control and Protection of dc Microgrids ........................... 233 10.3 Microgrid Operations ................................................................................... 235 10.3.1 Operation Objectives ........................................................................ 236 10.3.2 Technical Challenges ........................................................................ 236 10.3.2.1 Intermittent Renewable Generation ................................... 236 10.3.2.2 Low Grid Inertia ................................................................ 237 10.3.2.3 Coordination among Distributed Energy Resources ......... 237 10.3.3 Grid-Connected Operation ............................................................... 237 10.3.4 Island-Mode Operation ..................................................................... 238 10.3.4.1 Spinning Reserve ............................................................... 238 10.3.4.2 Minimum Loading of Conventional Generators ............... 238 10.3.4.3 Step-Load Requirements ................................................... 239 213 214 Smart Grids 10.3.4.4 10.3.4.5 10.3.4.6 10.3.4.7 Power System Stability and Oscillations ........................... 239 Reactive Power Requirement and Voltage Control ............ 239 Limitations to Renewable Energy Penetration................... 239 Short-Term Energy Storage Helps Renewable Energy Penetration .............................................................240 10.4 Microgrid Economic Value Streams ............................................................240 10.5 Reliability ..................................................................................................... 241 10.5.1 Reliability Due to Maintenance ........................................................ 242 10.5.2 Microgrid System Reliability ........................................................... 242 10.5.3 Can Reliability of a Microgrid be Better Than That of a Utility Grid? ................................................................................................. 243 10.6 Outlook: Microgrids—Energy for the Digital Society ................................. 243 References .............................................................................................................. 247 10.1 10.1.1 INTRODUCTION TO MICROGRIDS What Is a mIcrogrId? An integrated energy system consisting of distributed energy resources (DERs) and multiple electrical loads operating as a single, autonomous grid either in parallel to or “islanded” from the existing utility power grid. Pike Research The microgrid concept is a logical evolution of simple distribution networks and can accommodate a high density of various distributed generation sources. A typical microgrid power system consists of: generators; wind turbines; solar photovoltaic (PV) arrays; other renewable technologies such as geothermal generation; main grid connection/interconnector switch (optional); energy-storage devices such as lywheels and batteries for short- and long-term storage and stabilization; grid management software; controllable loads; and a robust communication network. The typical size of a microgrid is 500 kW–15 MW. 10.1.2 Why mIcrogrIds? Some of the various reasons for increased interest in microgrids include: • Rising Cost and Burdens of Transmission and Distribution (T&D) Infrastructure: Building new T&D infrastructure has become dificult in some areas due to permission issues, public resistance, and the dificulty, cost, or both of upgrading or building new infrastructure. • Integration of Renewables and Storage Technologies: The greatest obstacle to the integration of renewables has been the risk to grid stability. Some renewable energy sources, most notably wind and solar, are Microgrids • • • • • 215 intermittent by nature, which increases the voltage and frequency luctuation on the grids. The risk is greatest when there is a high penetration of renewables. However, technology advancements in grid stabilization and energy storage have addressed these concerns in microgrids. Nowadays, the challenge of renewable energy generation peaking not matching up with demand peaks can be resolved using energy-storage modules. An energy-storage module is a packaged solution that stores energy for use at a later time. The energy is usually stored in batteries for speciic energy demands or to effectively optimize cost. The modules can store electrical energy and supply it to designated loads as a primary or supplementary source. Moreover, it provides a stable and continuous power supply regardless of the supply source status. Voltage and frequency regulation can also be improved by using storage modules [1]. Power Quality and Reliability: The need for high reliability and good power quality has increased as more customers install microprocessorbased devices and sensitive end-use machines. Public Policy: In a full reversal from the past, public policy today is favoring distributed generation that offers improved eficiency, lower emissions, enhanced power system security, and other beneits of national interest. Policies supporting this include tax credits, renewable portfolio standards, emissions restrictions, grants, and so on. More Knowledgeable Energy Users: Energy users are becoming more aware of alternative power approaches and are more willing to consider on-site generation options than in the past. Many are interested in combined heat and power (CHP) as well as reliability enhancements. Storm Recovery and Restoration: Hurricane Sandy and other “superstorms” of the last decade have highlighted, in painful and expensive ways, the need to protect consumers from the devastation caused by such natural disasters. Today, both utilities and industry recognize the need to collaborate to make the microgrid a reality and ensure our community, place of business, or entity is insulated from an extended power outage in the case of a natural disaster. The technology needed to make microgrids feasible for everyday use, such as that needed for power storage and stabilization, has reached a point where it is commercially viable and available. The reliability beneits of the microgrid are clear. When the main grid loses power, the microgrid can switch to island mode. As long as there is an adequate source of power at the local level (diesel generator, wind turbine, fuel cell, PV, etc.), the power continues to low while the rest of the community is enveloped in darkness [1]. Increased Interest in Security: While cyberattacks often dominate the discussion of grid security, utilities and governments need to ind ways to protect the grid from natural disasters as well as physical sabotage. Experts agree that any prolonged disruption of power, regardless of the cause, represents a threat to security and economic stability [1]. 216 Smart Grids 10.1.3 mIcrogrId market There are currently four leading end-use applications deployed today: Institutional/Campus Microgrids: Currently, the largest market for microgrids is found in institutional campus settings. These installations often make use of CHP systems to maximize eficiency and reduce dependence on local infrastructure. Reliability can be of extremely high importance, particularly if the campus includes vital loads such as a hospital and is located in a rural area that is subject to more frequent disruptions in the power supply [2]. Military Base Microgrids: The most active market now is in military facilities, both stationary bases and forward-operating bases. The US Department of Defense (DoD) has mandates to obtain 25% of its electricity from renewable sources by 2025, and to make US bases net-zero energy users. The Defense Critical Infrastructure Protocol also directs the DoD to identify critical operations and demonstrate the ability to operate them over a 30-day period without service from outside electric infrastructure or resupply of fuel [3]. These are formidable objectives, and microgrids will be instrumental in achieving them. A combination of energy security, reliability, and cost is driving the military’s activities, which extend to renewable generation, energy storage, and electric vehicles, among others. According to Pike Research, there are more than two dozen microgrid implementations going on now at military facilities in the United States, and there is support in Congress for more. The military has a long history of pioneering technologies that later become widely accepted—the Internet being the most compelling example—and the Department of Energy (DOE) is reportedly looking to leverage the DoD’s work to apply microgrid technology to emergency response and relief efforts [2]. Remote “Off-Grid” Microgrids: Microgrids can provide an alternative to conventional power grid development to remote facilities that do not have access to grid power, especially developing countries and isolated islands. These microgrids never connect to a larger macrogrid, and therefore operate in an island mode on a 24/7 basis. Many of these microgrids are designed to reduce diesel fuel consumption by integration of solar PV, distributed wind, and other potential renewable sources such as hydropower. Microgrids are easier to build than large power plants serving loads via large T&D networks, and they are focused on serving the speciic needs of the area in which they are located. They are also easily scalable. If these attributes prove to be salient for even a modest number of countries, the resulting potential market could be quite large [2]. Critical Loads: There is a segment of the microgrid market that is sometimes overlooked, or more often rolled into the larger institutional/campus category, and that is what is referred to as critical load. Historically, this term has most often been associated with facilities such as hospitals or telecommunications—things with an essential link to the functioning 217 Microgrids of modern society. In recent years, however, a new facility has emerged as “critical” not only from a business perspective but from a societal one as well: data centers. Given their explosive growth—they already account for a signiicant portion of all electricity consumption—and their increasing importance to every sector of the economy, data centers may well represent their own segment of the microgrid market. They have tremendous requirements for power, using up to 100 times as much as a comparably sized ofice building, and most, if not all, of the load must be served at all times. Top-tier data centers already go to great lengths to ensure reliability, with redundant grid connections and extensive on-site backup power capabilities. Going forward, advancing microgrid technologies could offer improvements in both reliability and eficiency/cost for data centers [2]. 10.2 MICROGRID DESIGN This section addresses several important aspects related to the design of microgrids, namely: • • • • Network topology Control approaches Protection approaches Direct current (dc) microgrid 10.2.1 oPtImal mIcrogrId toPology Microgrids can operate in parallel with or isolated from the utility grid during emergency conditions or planned events. This type of distribution grid structure offers potential for improvements in power supply eficiency and reliability of power supply in comparison with traditional and passive distribution grids. But what is the optimal topology for this kind of power distribution network? The well-known approaches include a radial, a normally open-loop, or a meshed structure for the distribution systems, which constitute the possible solutions for microgrid optimal topology (Figure 10.1). Radial Open loop Normally open FIGURE 10.1 Traditional distribution network topology. Meshed 218 Smart Grids Factors determining the selection of optimal microgrid network topology include [4]: • Size, type, and location of DERs and loads • Power quality and reliability targets (the minimum power quality that a microgrid has to provide to its customer affects the selection of an optimal network structure) • Economic constraints/available budget • Investment cost (primary equipment, protection and control, communication, etc.) • Operating and maintenance costs, including cost of power losses and energy not supplied due to interruptions • Technical constraints (e.g., protection system, voltage proile, and physical equipment dimensions) • Voltage level (usually medium-voltage networks are open-loop networks and low-voltage networks are radial, with normally open-loop topologies in some exceptional cases) The optimal microgrid topology selection is a complex task with multiple conlicting objectives and uncertainties. Value-based reliability planning provides a rational and consistent framework for addressing the fundamental economic trade-off question of how much reliability is adequate from the customer perspective, gives rise to eficient topologies that can meet the consumer requirements with the minimum system upgrade, makes the system more robust, and complies with consumer requirements [5]. It is well known that in a conventional distribution system the failure frequency of the medium-voltage network is decisive for the frequency of low-voltage network interruptions. For that reason, it is also obvious that the adoption of an islanded lowvoltage microgrid leads to a remarkable reduction of “local” interruptions caused by the medium-voltage network. However, a low-voltage microgrid has a small impact on the “global” reliability indices of a distribution system. 10.2.1.1 Traditional Distribution Network Topologies Traditional low-voltage distribution networks have been radial, and the focus of network planning has been primarily on preserving the radiality of the distribution system. For reliability reasons, it is common that medium-voltage distribution networks use normally open-loop structures (Figure 10.1). In both cases the system operation is performed radially. The meshed structure is the alternative scheme, and it offers a large degree of freedom for the optimization procedure, providing a large number of different topologies according to technical and economic constraints. Is it still convenient to keep the radial operation of microgrids, considering the large penetration level of the DER, and verify whether the adoption of advanced topologies could help overcome the obstacles that limit the penetration level of the DER? Radial is usually the most cost-effective topology in a traditional, passive distribution grid due to a reduced number of switching devices, shorter cable length, and simple nondirectional overcurrent (OC) protection systems. However, there are no options to change the network coniguration if faults occur in the radial topology, which usually leads to unnecessarily long customer interruptions. Microgrids 219 The utilities’ increased concern about a reduction of customer interruption time has led to the addition of connections that provide alternate power supply routes in the case of outages or scheduled power supply interruptions. These emergency connections end with an open switch to preserve the radial network structure during normal operation. In this way, a distribution network has normally open loops, which can be connected into the meshed structures and increase the reliability of the power supply and the availability of local generating and energystorage devices. Coniguration management is possible with open-loop topology for service restoration, power loss reduction, and load balancing to relieve feeder overloads. Also, it is expected that DERs will play a more important role, and their penetration will reach a high level in the future. In some cases, local generation may exceed local feeder load and lead to a reverse power low, feeding power into an upstream grid. Thus, distribution utilities will probably have to dismiss the traditional radial network operation, adopting more lexible meshed structures used today at higher voltage levels. Meshed networks have well-known advantages versus radial/normally open-loop schemes [4]: • • • • • • • Reduction of power losses under the same load demand and DER penetration Improved voltage regulation Greater lexibility and ability to cope with the future load growth Better asset exploitation Alternative power low roots and increased reliability of power supply Greater eficiency due to increased availability of local DERs Improvement of power quality due to fault level increase However, compared with a radial arrangement, a meshed network also presents some drawbacks, usually leading to a higher investment cost: • Increased short-circuit current could necessitate a substitution of the existing circuit breakers (CBs), due to the overcoming of their nominal faultinterrupting capacity. • Need for more complex protection schemes due to alteration of a nondirectional OC protection logic, which can prejudice its selectivity. • More complex planning and system engineering. Therefore, major factors affecting the decision-making process regarding the selection of optimal microgrid topology are: • • • • Penetration and dispersion levels of the DER Load characteristic (dispersion, reliability requirements, etc.) Anticipated events and envisaged operation of the microgrid Availability and suitability of equipment enabling meshed operation (network monitoring, power low and fault level management, protection system, etc.) 220 Smart Grids 10.2.1.2 Size and Location of Distributed Energy Resources There are two main options for integrating DERs in a microgrid. The irst option can be called a “single object power supply” [4], consisting of small-scale DERs, each of which resides with one consumer at a time. The second option is characterized by a residential area power supply in the form of one larger generator for each residential area, or a complete microgrid. In the case of the single object power supply, the DER units can be disconnected automatically from the main power supply system in the event of any fault and create a consumer microgrid. The load can be supplied without interruption resulting from internal microgrid faults; however, it is necessary that the nominal capacity of each local generator meets the peak of local load. In the second option, a microgrid has the advantage of using the compensating effect of the consumer’s collective, but a reduced redundancy in generation is a disadvantage. The presence of intermittent generation resources requires the use of energy storage to provide a power balance during isolated operation. Energy storage can be realized as a set of small-scale and distributed devices (located at every household/ building or next to the local DER) or as a community-scale single energy-storage device for an entire microgrid. 10.2.2 mIcrogrId control Microgrids are comprised of different components, such as: • Distributed generators, such as microturbines, fuel cells, PV, diesel generators • Energy-storage devices, such as batteries, lywheels, supercapacitors • Flexible loads, such as heating, ventilation, air-conditioning, and lighting • Reconigurable feeders, tap-changing transformers, and reactive power compensation These components can be controlled in a continuous or discrete way in order to keep a microgrid running in utility-connected or islanded operating modes, as well as to guarantee a seamless transition between two modes. Typical control tasks include: • Spinning reserve management to cope with emergency conditions • Generation, energy storage, and demand management for both steady-state and dynamic microgrid control • Network coniguration management • Coordination with distributed generation and feeder overload protection relays Most of these tasks are executed at different timescales, from the subsecond range up to hours and days. The short-term control actions are focused on system stability by reaching an instantaneous power balancing and acceptable voltage level. The mid- and long-term control actions focus on economic objectives and system security. 221 Microgrids A major control challenge related to microgrids is coordinating a variety of components in the microgrid to facilitate their parallel operations without loss of voltage and frequency stability, and doing so in the most economic way. Techniques and approaches vary, resulting in a continued wide divergence as to what type of microgrid control system will work best. There are several control approaches applicable to microgrids that have been evaluated by a scientiic community and tested in the ield at several pilots. They vary between fully decentralized and fully centralized approaches, with various hierarchical/hybrid schemes in between (Figure 10.2). 10.2.2.1 Centralized Approach This approach is based on the presence of a centralized controller, which collects measurements and other necessary information at a single point, where control decisions are taken and communicated back to the actuators. This approach is much more focused on optimal microgrid operation, including services outside the microgrid [6], while beneits of reliability may come as a second objective. Although the centralized approach provides better system observability and optimality of decisions, it has been recognized that it also has several downsides in a microgrid environment: • Failure of the central controller can be catastrophic, and redundancy increases the cost. • Need for a performing hardware for a central master controller (memory and central processing unit) and high-speed and low-latency communication network increases the cost. • System maintenance requires complete shutdowns. Microgrid control complexity Comm. requirements and costs Comm. dependence and latency Sensors and costs Responding speed Microgrid reliability 100% decentralized Device control complexity Performance and effectiveness Optimization of power flow and Energy utilization Efficiency Standardization Scalability/modularity Compatibility Where is the balance? 100% centralized FIGURE 10.2 Microgrid system control: decentralized versus centralized. (From Xu, Y., ORNL Power Systems and Power Electronics for Utility Applications, ORNL, Workshop on Power Electronics and Distributed Power systems, UTRC, December 2, 2011. Available at: http://web. ornl.gov/sci/ees/etsd/pes/pubs/UTRCWorkshop_ORNLPowerElectronics_Microgrid.pdf. [7]) 222 Smart Grids From an economic point of view, the layered/hierarchical system is the more practical solution. In this case, a local control focused on system stability is delegated to local controllers, while the central controller executes slow control actions of economic dispatch, calculating schedules for local controllers as well as an amount of reserve for emergency conditions. A layered centralized microgrid control system is described in [8]. For example, control architecture can consist of the following levels: • Local-level controllers • Microsource controllers use local information to control the voltage and the frequency of the microgrid in transient conditions. • Load controllers shed loads in emergency conditions. • Microgrid central controllers, which are used for • Forecasting of electrical and heat load and production capabilities • Economic dispatch of different production and consumption units • Online security assessment • Demand-side management • Black start and grid resynchronization A practical realization of the centralized microgrid approach is described in [9]. 10.2.2.2 Decentralized Approach The advantages of having the control system distributed and split up into individual peers are as follows [10]: • Each electrical device can be itted with a separate, less complex controller. • The failure of one controller does not have catastrophic impact, since replacement capacity can be brought online immediately. • The system is more scalable and extendable, not limited to onboard input/ output of the central master controller. • It is easier to develop upgrades, thanks to independence and modularity. • Communication redundancy can be easily achieved. • Each node is autonomous and yet closely integrated with its peers. In a fully decentralized approach, the local controllers of different controllable microgrid components are responsible for satisfying the demand (and possibly exporting power to the utility grid) autonomously (communication in such a distributed environment works through exchanging messages between the individual controllers) in the most economical way. In addition to economic goals, local controllers may have local control tasks such as heat generation, voltage control, backup for critical loads, and so on. These tasks suggest the importance of distributed control and autonomous operation [8]. The use of multiagent systems in controlling a microgrid has been investigated by leading academics in [8]. It is especially eficient in the case when different microgrid components have different owners; in this case, decisions should be taken locally, so centralized control is dificult. Practical examples of decentralized microgrid control Microgrids 223 include the Consortium for Electric Reliability Technology Solutions (CERTS) [11] and Powercorp [10]. Such a system will work for the majority of small microgrids, whose highest priority is reliability and sustainability during emergency conditions. 10.2.3 mIcrogrId ProtectIon One of the major technical challenges associated with a wide deployment of microgrids is the design of their protection system. Protection must respond to both utility grid and microgrid faults. If the fault is on the utility grid, the desired response is to isolate the microgrid from the main utility as rapidly as necessary to protect the microgrid loads. The speed of isolation is dependent on the speciic customer’s loads on the microgrid, and in some cases may require the use of electronic static switches. If the fault is within the microgrid, the protection system isolates the smallest possible section of the distribution feeder to eliminate the fault [12]. A further segmentation of the microgrid during the isolated operation, that is, a creation of multiple islands or submicrogrids, must be supported by the availability of the DER and load lexibility. Time OC protection is the most common method employed in medium- and lowvoltage distribution grids, as it is quite simple to conigure and is a cost-eficient solution. The presence of the DER changes the magnitude and direction of fault currents and may lead to protection system maloperation [13–16]: • Reduced fault sensitivity and speed in tapped DER connections. • Unnecessary tripping of utility breaker for faults in adjacent lines due to DER contribution. • Reduced fault contribution of inverter-based DER in isolated mode. • Increased fault levels may exceed the capacity of the existing switchgear. • Auto-reclosing and fuse-saving utility policies may fail. In such circumstances, coordination between relays may fail and generic OC protection with a single setting group in each relay may become inadequate; that is, it will not guarantee a selective operation for all possible faults. Therefore, it is essential to ensure that protection settings chosen for OC protection relays are dynamically modiied, adapting to changes in grid topology and location, type, and number of connected DERs. 10.2.3.1 Adaptive Protection Adaptive protection is deined as “an online activity that modiies the preferred protective response to a change in system conditions or requirements in a timely manner by means of externally generated signals or control action” [16]. Technical requirements for the practical implementation of an adaptive microgrid protection system are: • Use of digital directional OC relays, since fuses or electromechanical and standard solid-state relays do not provide the lexibility for changing the settings of tripping characteristics, and they have no current direction sensitivity feature. 224 Smart Grids • Use of new/existing communication infrastructure and standard communication protocols such that individual relays can communicate and exchange information with a central computer or between different individual relays rapidly and reliably to guarantee a required application performance [17]. The parameter adaptation time delay is not critical, because the communication infrastructure is used to collect information about the microgrid coniguration and to change relay settings accordingly only prior to the fault. The changes of the microgrid coniguration will be identiied within a few seconds, depending on the dimension of the network, and the protection system reconiguration has to be completed in times of the same order, provided that basic backup protection functions are present during the transitory phase. Next, two approaches to modifying protection settings are addressed. Both approaches are similar in using a central protection coordination unit that collects information about microgrid topology and decides whether adaptation of settings is needed. The difference between the approaches is in how the setting modiication physically happens after the decision to apply a new setting is taken. The irst approach is based on online switching between precalculated and “trusted” setting groups that are kept in the memory of the protection device (today’s intelligent electronic devices [IEDs] can hold four or more different setting groups). In this case the central coordination unit sends a command of which setting group number must be set as active, and the IED implements it locally, reporting back the new value of the active setting group. There are two main steps in realizing this approach: • Off-line calculation of settings: A set of meaningful microgrid conigurations as well as feeding-in states of DERs (on/off) is created for off-line fault analysis. Next, fault currents passing through all monitored CBs are calculated by simulating different short-circuit faults (three-phase, phaseto-ground, etc.) at different locations of the protected microgrid at a time. During repetitive short-circuit calculations, a topology or a status of a single DER or load is modiied between iterations. Based on these results, optimal settings (to achieve targeted sensitivity, selectivity, and speed) for each directional OC relay and for each particular system state are calculated. The settings are then uploaded into the memory of protection devices. • Online operation: During online operation, the central protection coordination unit monitors the microgrid state by receiving information from remote IEDs or sensors. This process runs periodically or is triggered by an event (tripping of CB, protection alarm, etc.). The adaptation algorithm then compares the actual microgrid coniguration with previously analyzed cases and retrieves the setting group number most suitable for the current coniguration. If the actual setting group number (also monitored) is different, the output of the adaptation algorithm will be a setting switching command, as shown in Figure 10.3. Alternatively, if the number of settings exceeds the memory size of the local protection device, a complete new setting group can be uploaded (replacing one of the settings not used in the memory) into the memory of the local device and then activated [17]. 225 Microgrids t Legend Ik2max = 2.4 kA seen by CB3.1 100 s CB3.2, CB6.2 Base trip curve CB2.2, CB3.1 CB5.2, CB6.1 10 s Dynamically modified trip curve Delayed trip 1s Selective trip 0.1 s 1 kA 10 kA FIGURE 10.3 Adaptation of the OC trip curves in response to microgrid topological change. (From Oudalov, A. and Fidigatti, A., International Journal of Distributed Energy, 5, 201–225, 2009.) The second approach is based on real-time calculation of settings as soon as a change in microgrid topology is detected by the central protection coordination unit. The calculated settings are then uploaded into the memory of the local device and activated. In some speciic implementations, the central unit may also perform primary protection functionality; that is, in the case of a fault, the central algorithm implemented in the central unit generates selective tripping signals, which are sent to the respective CBs. This application is limited to a single substation case, since communication may become a bottleneck for remote CBs, which will require the presence of high bandwidths and reliable communication channels [16]. Integration of state estimation routines into a coniguration of the protection system, which monitors the DER operating state and sets up a corresponding data exchange with the protection system, is also possible [18]. In this case, the prediction data of the availability of the DER are used in order to review the selectivity of the tripping characteristics for each new operating condition, based on the short-circuit power of the DER and the actual short-circuit power available from the main grid, and to adapt them accordingly if selectivity is no longer suitably provided. If the adaptation is successful and the boundary conditions are not violated, the tripping characteristics of respective relays will be matched. If there is no possible solution without violating the boundary conditions, a signal will be generated that forbids the acceptance of the operation predicted by the energy management system (EMS). 226 Smart Grids 10.2.3.2 Fault-Current Source As mentioned above, one of the major challenges in microgrids is fault detection in the presence of a DER with power electronics (PE) interfaces, especially in the isolated mode in the absence of the utility power fault source. One possible solution is to use an independent system that supplies a fault current comparable to the level provided by the utility grid. In this case, protection devices do not need to adjust protection settings. This concept is also suitable in the case of the active use of fuses as protection devices. The proposed system, designed as fault-current source (FCS), is based on PE, but using a simpler and thus less costly technology than is normally used in inverters in this power range. Its power circuit remains idle during normal operation of the network, but when a fault is detected (by sensing a local network voltage level) the FCS is activated. It attempts to restore the original network voltage by injecting a fault current into the network for a speciied duration, enough to cause a fuse or CB to clear the fault. The source of the fault current is an ultracapacitor [19]. 10.2.3.3 Fast Static Switch Depending on the switch technology, momentary interruptions may occur during transfer from grid-connected to islanded mode. If high-speed isolation is required by the load, the static switch is a suitable technology. Static switch is a PE-based solution used for fast and automatic islanding and for grid resynchronization. It is capable of isolating sensitive loads and microsources from the distribution system in less than one-quarter of a cycle during faults and other disturbances. There are two major technologies used in static switches: • Silicon-controlled rectiiers (SCRs) • Insulated gate bipolar transistors (IGBTs) Some lessons learned from testing the static switch prototype are shared in [19]. The differences among the varieties of switching technologies are based on the speed of response. A wide range of interconnection speeds allows lexibility to match application requirements. The SCR-based switch can respond in one-half to one cycle. Some SCR switches in transfer switch coniguration can transfer within one-quarter of a cycle. This is a relatively expensive solution due to the use of semiconductor materials and the need for a CB to handle fault conditions. The IGBT-based switch can respond in the 100 μs range. Also, IGBTs can clamp the instantaneous currents and turn off in very short time frames. Key technical questions which must be addressed when using static switch in the microgrid include: • Coordination of microsource control with a static switch (especially in the case of active synchronization) • Coordination with other protection devices (adjustment of tripping curves and load-shedding sequences) Microgrids 227 10.2.3.4 Communication Architectures and Protocols for Adaptive Protection Successful implementation of the adaptive protection system requires that protection IEDs can communicate among themselves or with a central controller. By means of the communications, IEDs can provide and receive information about the actual network topology of the microgrid, the status of local generators, energy-storage devices, and loads, and commands to execute an action (e.g., change of the active setting group). Within a range of communication protocols that enable a protection system to become adaptive, International Electrotechnical Commission (IEC) standard 61850 is of great interest [20]. The most important IEC 61850 advantages are: • • • • Guaranteed interoperability between devices from different vendors Standardized data models Simpliied engineering process and IED coniguration Scalable solutions Its main disadvantage is that it requires an Ethernet network, which is a common technology for automation at high- and medium-voltage levels, but today still means extra cost and technical complexity in low-voltage installations. Moreover, it poses additional complexity for the maintenance personnel, for a system that is used to working more comfortably with hardwired systems. For adaptive protection application, this standard has communication services for usual operations in electrical installations, such as change of settings and setting groups, report of information, commands, and so on. One remarkable service is known as the generic object-oriented substation event (GOOSE), which makes direct information exchange between IEDs possible, accepting any type of data (single signals, double signals, measurements, etc.), guaranteeing that they are transmitted in less than 3 ms (faster than a cable between digital outputs and inputs), and are always available, because they are retransmitted indeinitely. Obviously, the application of GOOSE messages for communication between IEDs within the microgrid provides a valuable tool, allowing their conigurations to be adapted to any operation scenario, such as the transition from the grid-connected mode to the islanded mode. Each IED can receive data from all others very rapidly, so the reconiguration process is very fast [16]. 10.2.4 selectIng the rIght communIcatIons technology Perhaps the most important aspect of microgrid communication design involves deploying suitable and future-proof wired/wireless or hybrid technologies. In general, while wired technologies, particularly optical iber, provide more capacity, reliability, and security, they can be costly to deploy, and inlexible once they are deployed. As distribution networks grow more sophisticated, and as more intelligent points are added with a need for communications, this unfavorable cost per point correlation is only exacerbated. So, while iber might be considered in situations where it is already found close to assets we wish to connect to, or perhaps 228 Smart Grids in circumstances where a new building or community is going up, it is becoming apparent that wireless communication technologies may in many cases be a better it for the needs of microgrids and other distribution systems. While, in the past, available wireless systems may not have met the requirements of the most performance-sensitive applications, this is no longer the case. In particular, for ield or distribution-area networks, it is now possible to leverage wireless mesh technologies to build a network capable of supporting multimegabit throughput, low latency (<20 ms round trip), enterprise-grade security, and resilient routing architectures supporting up to 99.999% of reliability. These “mesh” wireless systems typically function in the unlicensed spectrum found in the 2.4 and 5.0 GHz ranges and can provide coverage over a utility microgrid, along with connectivity into the various IEDs a utility might wish to connect for the purposes of information and control (Table 10.1). Complemented by terrestrial communications such as iber found at local substations or other interconnection points within the microgrid, and possibly by other wireless technologies where low density or lack of iber might dictate, it is possible to design a network that meets the communication requirements of the microgrid, both now and in the future. 10.2.5 dIrect current mIcrogrIds Much has been written recently about dc as applied to power grids. Historically, the power system started as dc during Thomas Edison’s efforts to electrify New York City at the end of the nineteenth century. This, however, was short lived, since dc could not be transformed to higher voltage levels, which prevented the long-range transmission of dc power. Under Nicolai Tesla and George Westinghouse, alternating current (ac) prevailed, and three-phase ac generation, transmission, and distribution became the choice for power systems. In the 1930s, Swedish company ASEA TABLE 10.1 Network Requirements Will Increase with Time Application Reclosers Capacitor bank Remote terminal unit Motor-operated disconnect Line regulator Advanced metering Interval data recorder Demand management Mobile workforce management voice Mobile workforce management data Aggregate Latency (ms) 10 s 100 s 1000 1000 100 s 100 s 100 s 100 s 100 s 100 s 10 s Bandwidth (kbps) <56 <56 56 <56 <56 56 <56 <56 <56 1000s 1000s Source: IBM Presentation for the Utilities Telecom Council on the Utility of the Future. Microgrids 229 developed a high-voltage dc transmission, demonstrating that under some conditions dc could overcome some of the hurdles that ac faced, such as inductive and capacitive effects, which reduce the throughput of the transmission line. In the 1950s, several commercial installations commenced, and the trend of using high-voltage direct current (HVDC) for applications such as long cable transmission (underground and marine) and tying otherwise asynchronous systems continues to this day. HVDC systems are primarily point to point; that is, an HVDC line has an ac/dc converter station at each end to interconnect with the ac system. The primary reason why HVDC transmission technology did not develop widely into an HVDC system with many interconnected dc lines is the lack of dc CBs, devices that could protect the system against faults and interrupt the resulting short-circuit currents. Today, dc is experiencing a major comeback. There are several reasons for this. Let us name just a few: • More and more electric loads are inherently dc powered. Consumer electronics (from laptops and smart phones to plasma TVs), information technology (IT) equipment (servers, network switches, storage, etc.), variable-frequency drives (VFD), and so on, all have dc as the internal supply voltage. It is estimated that the share of the inherently dc electronic loads will continue to grow. • The popularity and applications of dc machines (motors and generators) have grown. With the help of advanced rare-earth magnets, new designs, and permanent magnet machine technologies, dc machines achieve performance comparable to, if not better than, their ac equivalents. Coupled with their VFDs, they can provide a new range of functionalities and operational parameters such as torque characteristics at different speeds. • Renewable generation resources such as wind energy turbines and solar PV arrays also use dc before the power is converted to the main frequency ac and synchronized with the grid. • Batteries have become the primary means of electric energy storage, and these technologies are dc based. Other energy-storage technologies, such as the lywheel and the fuel cell, also use dc within them as an intermediate stage. • Advances in PE and power semiconductor technology over the last several years have made the transformation of current from ac to dc and vice versa so eficient and inexpensive that it has begun to compete in cost and eficiency with conventional ac transformers. • New PE equipment opens new functionalities that were never possible before: controlling the power low, current, and voltage levels, power factors as seen from the ac side, and sensing and interrupting short-circuit currents with unprecedented speed, all with no moving parts. Many believe that, for microgrids, dc will play an even more important role. Because of its small scale, implementation of dc power in a microgrid could be much simpler, and therefore one could beneit from the dc technology with much less effort and cost. 230 Smart Grids 10.2.5.1 Differences Between ac and dc for Microgrids • The irst difference involves the synchronization and integration of several energy sources. With dc, one can connect multiple rotating machines without a need for synchronization. Either dc machines or ac machines with rectiiers can be connected to a dc microgrid without expensive and complex paralleling schemes. Recall that microgrids do not have a so-called ininite bus, so each connection of a machine to an ac microgrid can cause a voltage and frequency swing that is dificult to control. • Operation of dc microgrids, and all dc grids for that matter, involves delivering (generating, distributing, and consuming) real power (P) only; no reactive power (Q) is present. Since balancing reactive power can often be a challenge in ac systems, dc is inherently easier to control and operate. There is no need for power factor correction equipment and control, such as power factor shunt capacitors or voltage-control shunt reactors. • System protection for ac is based on mechanical CBs, which cover the entire range of voltages and currents of today’s systems. CBs for ac are common and are based on mature technology; coupled with their economy of scale (tens of millions of devices produced each year for different applications and at various voltage and current levels), they provide an excellent and low-cost solution for circuit protection. CBs for dc are complex and limited primarily to low voltage. They are based on the principle of current limiting by developing arc voltage greater than the system voltage or by injection of transient high-frequency current to create artiicial current zero crossings. This currently gives ac systems the advantage over dc. • Operating procedures for disconnecting (islanding) and connecting to the main grid, as well as operating in parallel with the main grid for bidirectional power low, are sometimes simpler for dc microgrids. • Single-phase or two-phase dc systems are possible, and, with either two conductors (+ and −) or three conductors (+, −, and ground), they can offer signiicant cost reduction compared with ac. In most cases the three-phase three-wire or four-wire ac conductor systems (overhead lines, underground cables, or others) could be converted (retroitted) to dc, although the opposite would not be possible. A dc conductor system can deliver sqrt(2) higher power than the equivalent ac conductor system, primarily because the dc insulation system can be designed for the operating voltage whereas ac insulation has to be designed for sqrt(2) times operating ac root-meansquare voltage. Additionally, the conductor resistance for dc is lower than for ac, since the current density within the conductor is uniformly distributed through its cross section; for ac, the skin effect pushes the current to the surface. The magnetic skin depth for copper at 60 Hz is only ~8.5 mm. • Plugs, receptacles, switches, power supplies, sensors, motors, and other dc equipment are not yet as popular and commercially available as ac equipment; ac equipment has a large historical presence and its installed base is huge, with multiple vendors and a large commercial selection of brands. The evolution of the dc industry will take some time. 231 Microgrids 10.2.5.2 Examples of dc Microgrids In principle, dc microgrid architectures could be similar to their ac counterparts. They are making inroads into building systems and data centers, among other things [21,22]. Concepts such as radial, single bus, dual bus (with tie), mesh, or ring main are equally applicable to dc, depending on the size of the microgrids, that is, the number of sources and loads to be connected as well as the physical dimensions (grid length and area to be covered). In most cases so far the microgrids have been of either single-bus or ring-main (ring-bus) design. 10.2.5.2.1 Single-Bus dc Design This is the most popular concept, which applies to the smallest installations, such as ofices, ofice buildings, or small commercial enterprises (Figure 10.4). It can also be extended to other, noncritical applications such as on board a train or an electric bus. Some degree of redundancy can be built into such a simple system, but the main single point of failure, the dc bus itself, will remain. 10.2.5.2.2 Dual-Bus with a dc Tie Breaker or dc Switch In this particular topology, a dual-bus system on board a ship [23] is used for a degree of redundancy. The concept is that if one bus is not available, for example, due to faults or maintenance, the other bus can still power enough loads to keep mission-critical operation running. In traditional marine electrical propulsion systems, ac VFDs typically account for more than 80% of the installed power. In the new approach, the onboard dc grid concept is based on taking a multidrive system and distributing it so that the distributed rectiiers can be eliminated. In this way, generation sources and major loads, in this example diesel generators and electric propulsion motors, can be connected to two common buses. Such a system can be optimized from the energy production and consumption point of view, resulting in substantial energy savings, up to 20%. The new microgrid can be seen as one large multidrive system merging the various dc links that would otherwise reside in individual VFDs around the vessel. Power is distributed through a dual-bus single 1000 V dc circuit, thereby eliminating the need for main ac switchboards, distributed rectiiers, and converter transformers. ac grid (macrogrid) ac source(s) ac dc dc source(s) ac/dc ac/dc dc bus dc loads dc/ac dc/ac ac loads FIGURE 10.4 Simple diagram of a single-bus dc microgrid. Bidirectional ac/dc conversion 232 Smart Grids All generated electric power is fed either directly or via a rectiier, depending on the nature of the source, into a common dc bus. Each main load is then fed by a separate inverter unit (Figure 10.5). If an ac distribution network is still needed, for example, for auxiliary loads, it is fed using dc/ac converters. Additional converters for energy storage in the form of batteries, PV arrays, or supercapacitors for leveling out power variations can be added to the dc grid. In some applications, one can envision connecting dc energy sources or storage devices directly to the dc bus(es); however, adequate protection would have to be included against short-circuit currents. The onboard dc grid is lexible and can be conigured in several different ways depending on the physical constraints (space), length of feeders for sources and loads, dc currents, losses, and reliability of different components, among other factors. For example, with a centralized approach all converter modules are located in one or multiple lineups within the same space in which the main ac switchboard would normally be installed. In another implementation, the individual converter modules could be located close to the devices they serve, and the dc bus(es) extended throughout the ship. 10.2.5.2.3 Ring Main For even better reliability and to provide some redundancy of power delivery, a dc ring bus (or ring main or loop) provides an alternative path for supplying power while any section of the bus is out, either faulted or for routine maintenance (Figure 10.6). Such a system can be implemented in several different variants. For example, for more critical loads a more dedicated source can be placed on the same section of the ring; one can also use fewer separation points (fewer breaks in the ring) for lowercost implementation. In yet another variant, the ac grid can be connected in more than one point around the loop. The operational principle is rather simple and almost identical to the ac ring mains popular in many distribution systems in large European cities. Whenever a bus fault occurs, this section of the bus is isolated by the respective CBs (or other protective devices) and the remaining portions of the ring main operate as usual. There are no G G ∼ ∼ = G G ∼ = ∼ = = dc bus dc bus = = ∼ = ∼ = ∼ = ∼ = ∼ ac bus ac bus M FIGURE 10.5 ∼ M Simpliied diagram of a dual-bus dc system with a tie-switch. 233 Microgrids ac grid (macrogrid) ac source(s) ac dc dc source(s) ac/dc Bidirectional ac/dc conversion ac/dc dc bus dc loads dc/ac dc/ac dc ring (loop) ac loads FIGURE 10.6 Simpliied diagram of a ring-main dc microgrid. downstream or upstream devices. Each load and source can be operated from two directions of the ring main. 10.2.5.3 Control and Protection of dc Microgrids As with ac microgrids, the most critical aspect of designing, operating, and maintaining the dc microgrid is control and protection [24,25]. Even when the individual devices (energy sources, loads, conductors) are available, the design of control and protection logic has to be carefully thought out. Controlling dc microgrids involves primarily three aspects: 1. Switching microgrids from islanded to utility network operations, due to the availability of power, economics, utility outages, or other factors. The switching can be either hard, that is, instantaneous, as in the case of sudden, unplanned utility outage, or soft, as in the case of planned outages, economic considerations, or approaching a natural disaster such as a storm, hurricane, or looding. 2. Managing (balancing) the generation sources based on their availability (e.g., solar PV and wind), generating cost, and—in some instances—their location; this managing could involve different strategies depending on whether the microgrid is islanded or on grid. 3. Managing (controlling) load demand. Loads should be irst categorized as to their priority (mission critical, critical, essential, nonessential); their level of criticality can be dynamic, for example, a function of time, weather, or major external events. What makes the dc microgrid different is the fact that switching dc equipment (generation or load) on or off is potentially simpler, since no point-on-wave synchronization is necessary—no out-of-phase conditions, no reactive power low or phase angle differentials—and no frequency luctuations occur. At the same time, the dc challenge today involves the lack of established, documented, and wellproven dc control algorithms and processes, so that they have to be developed on a case-by-case basis. This will likely change as microgrids and the use of dc continue to grow. 234 Smart Grids Protecting dc microgrids against system disturbances such as faults, arcs, and open circuits can be quite different from the case of ac. All the aspects of fault detection, isolation, and restoration require different treatment. First of all, the waveforms of dc short circuit are fundamentally different. Depending on which devices are connected to the microgrid, it may have one of several types, as shown in Table 10.2. In most cases the short-circuit current of a dc microgrid will be a combination of the three generic waveforms shown in Figure 10.7, with different inductance to resistance (L/R) ratios, oscillating frequencies, and damping factors. Detecting a dc fault is, on the one hand, a little simpler, and can be a faster process, since any excursion of the current above a certain threshold for even a short period of time can indicate a fault. Therefore, the detection algorithm does not need to wait for a quarter of a cycle (5 or 4.166 ms for 50 and 60 Hz, respectively) as in ac systems. On the other hand, with a potential overshoot of short-circuit current when various dc link capacitors discharge (waveforms (b) and (c)), the dc system and its components have to withstand these instantaneous stresses comparably to ac. TABLE 10.2 Types of Short Circuits in dc Microgrids Source 1. Battery bank Load Nonbackfeeding loads 2. Rotating generator (dc) Nonbackfeeding loads 3. Rectiied ac from the utility grid Nonbackfeeding loads 4. Combination of sources Nonbackfeeding or backfeeding loads, or both 5. Ungrounded (loating) or high impedance grounded Ungrounded or high impedance grounded Short-Circuit Current Waveform Expected Exponential waveform from load current level to short-circuit current (Isc) determined by resistance of the faulty path, including internal battery resistance, conductors, and electrical arc, if any Depending on the type of machine Exponential waveform as in case 1, with superimposed contribution from the discharging capacitors Waveform will be complex, with several superimposed contributions from the discharging capacitors of dc links of various PE devices. Depending on the resistances involved, these contributions can be either underdamped or overdamped and with different L/R ratios or frequency and damping factors, or both. Single line to ground fault is limited to either small transient current (due to parasitic capacitance/inductance) or small dc current via grounding resistances 235 Current Microgrids L/R Steady-state fault current Prefault current Time (a) Time (b) Time (c) FIGURE 10.7 Three generic waveforms of short-circuit current in dc microgrids: (a) simple L-R connection with constant dc voltage; L/R ratio determines the rate of rise of the fault current; (b) an overdamped L-R-C oscillatory circuit; overshoot occurs due to the presence of a discharging capacitor(s); (c) underdamped L-R-C oscillatory circuit; frequency of oscillations is determined by a combination of L, C, and R, and decay of the oscillations is determined by R/sqrt(L/C). Isolating the fault means fault-current interruption. As mentioned before, interrupting the dc current by conventional means is not possible due to the lack of (natural) current zeros. The two possible methods of dc fault isolation are by means of dc breakers, PE converters, or both. Depending on the design and ratings, dc breakers can take up to several milliseconds to force the short-circuit current to zero. IGBT or similar PE-based devices can bring the current low to zero much faster, in a fraction of a millisecond, and without any internal arcing. The other advantage of handling the faults with PE is that the current does not have to be interrupted completely; for example, a small dc current could be maintained for further fault location or for the purpose of protection coordination with dc fuses, downstream and upstream dc breakers, and ac devices outside the microgrid. When the system is ready to be restored, PE devices can also “reclose” or re-energize the system almost instantaneously without recharging, and, if the reclose is at fault again, the current does not have to rise to its previous, original fault level. In fact, the system can be re-energized in a soft manner. This can result in virtually arcless, arcproof, and fault-tolerant systems, increasing safety for personnel and equipment alike. The availability of such fault-tolerant systems will also increase. In summary, dc microgrids offer unique advantages over their ac counterparts. With the growth of inherently dc loads, the advent of battery technologies, PE, dc machines, renewable energy, and a growing interest in plug-in electric vehicles, dc not only simpliies the design, integration, and operation of the microgrids, but also provides new functionalities and increased safety and availability. 10.3 MICROGRID OPERATIONS A microgrid can operate in either the grid-connected or islanded mode, and may experience mode transition. In the grid-connected operation mode, the microgrid is connected to the main grid through the point of common coupling and operates in parallel with the main grid to deliver power to the load. From the distribution 236 Smart Grids system perspective, a grid-connected microgrid can be treated as an individual controllable entity, acting like a generator to supply power or a load to consume power. On the other hand, the microgrid operating in the islanded mode is independent of or isolated from the main grid, due either to disturbance in the main grid or to its geographical isolation, such as a remote island. In the islanded mode, the microgrid operation, such as the voltage and frequency control and regulation, is implemented in a stand-alone system. The transition between the two modes can occur during the operation. For example, when a fault happens to the main grid, the microgrid is disconnected from the main grid to continue operation in the islanded mode, and, after the fault has been cleared from the main grid, the microgrid is resynchronized and connected back to the main grid via the common coupling point. Some types of microgrids, such as institutional/campus, industrial/commercial, community/utility, and military base microgrids, operate in the grid-connected mode most of the time, and only switch to the islanded mode when disturbances happen to the main grid. Some other types of microgrids, that is, remote off-grid and weakly grid-connected microgrids (such as military forward installations and remote mining industrial sites), operate in the islanded mode all the time, or very often. 10.3.1 oPeratIon oBJectIves The deployment of microgrids generally aims to increase the penetration level of DERs (especially the green energy resources), improve energy eficiency, enhance service reliability and power quality, and reduce greenhouse gas emission. In order to achieve these objectives, it is critical that the microgrid can maintain stable, economic, and secure operation in different modes. A stable operation relies on control of the voltage and frequency within certain limits when disturbances occur. Such disturbances may be load demand variation, renewable generation output luctuation, islanding from the main grid, and so on. Economic operation refers to the supply of the local load with minimum generation cost, which can be achieved by coordinated control among different devices, such as renewable generation, conventional distributed generators, energy storage, and the power import from the utility grid. Secure operation aims to guarantee the power supply for critical loads all the time and is very important for hospital, military base, and data center microgrids, in which the disruption of power supply for critical loads can cause a huge impact. 10.3.2 technIcal challenges Some of the technical challenges that must be overcome to achieve stable, economic, and secure microgrid operational status must deal with several aspects. 10.3.2.1 Intermittent Renewable Generation One of the major incentives to deploy microgrids is to facilitate the integration of renewable generation in the distribution system. The power output of renewable generation (such as solar and wind) is signiicantly inluenced by the season and weather, and these are characterized as intermittent power resources. That is, the Microgrids 237 power output of these resources can vary abruptly and frequently and imposes challenges on maintaining microgrid stability, especially in the islanded mode. 10.3.2.2 Low Grid Inertia A microgrid may include both conventional and modern DERs. Conventional distributed generators, such as diesel generators, usually utilize synchronous generators that directly connect to the grid. Modern distribution grids use the most renewable resources and energy-storage devices that are connected to the grid indirectly through the PE interface. These DERs either have a low inertia or are inertialess, and bring dynamic problems to microgrid operation. 10.3.2.3 Coordination among Distributed Energy Resources A microgrid may have various types of DERs, such as diesel generator, microturbine, fuel cell, CHP, energy-storage device, and so on. These DERs usually have different operation characteristics in their generation capacity, startup/shutdown time, ramping rate, operation cost/eficiency, energy storage charging/discharging rate, and intertemporary control limitations. The microgrid operation should consider the characteristics of different components and provide appropriate control strategies accordingly to improve economic and secure operation. 10.3.3 grId-connected oPeratIon In the grid-connected mode, since the microgrid is connected to the main grid, stable operation is not a major concern. Therefore, the operation objective is mainly to reduce operation cost and improve energy eficiency. The microgrid management can be implemented in either a centralized or a decentralized manner. This section illustrates the centralized microgrid management applications for grid-connected mode operation. Centralized control enables optimal operation via coordinated control among different types of controllable DERs. The centralized microgrid EMS performs the coordinated management of the microgrid in a central controller, which monitors overall system operating conditions, makes optimal control decisions in terms of minimizing operation cost, reducing fossil fuel consumption, providing services for the utility grid, and so on, and then communicates the setting points to DERs and control commands to controllable loads. Most of the current centralized EMS solution technologies fall into two categories: (1) day-ahead unit commitment combined with online economic dispatch (ED) and (2) online ED across multiple time intervals. These solutions have attempted to provide optimized operation strategy over a predeined period of time while taking into account the renewable generation and load forecast. The method of day-ahead unit commitment with online ED approach [26–31] generates an optimal operation plan for next day based on the day-ahead renewable generation and load forecast, and the ED is executed in real time using the day-ahead unit commitment results. The method of online ED over multiple intervals [32–43] can take the latest forecast on the renewable generation and load into account for the operation decision. However, it provides the control decision not only for the 238 Smart Grids current interval but also for the future intervals at each execution interval (e.g., every 5–15 min) in real time. Due to the heavy computational burden, simpliied optimization is usually deployed, considering only the power balance instead of more detailed operating constraints. 10.3.4 Island-mode oPeratIon In island-mode operation, the microgrid is operating independently of or isolated from the main grid. This mode of operation may be the result of a disturbance that has occurred in the main grid, or the microgrid may be geographically isolated from the main grid. Examples of this would include a remote island without a mainland link or a remote location without access to the grid, such as a forward military installation or remote mining industrial site. One of the drivers for isolated microgrids is the ability of these systems to integrate renewable energy sources. These renewable energy sources include generation sources such as wind turbines, PV solar systems, solar-thermal power, biomass power plants, fuel cells, gas microturbines, hydropower turbines, CHP microturbines, and hybrid power systems. The renewable energy sources are normally paired with fossil generation as part of the power generation systems for isolated microgrids. The exploitation of renewable energy sources, even when there is a good potential resource, can be problematic due to their variable and intermittent nature, especially when these resources are connected to isolated grids. When using renewable generation in isolated power systems, some of the major dificulties involve power quality and reliability, due to the stochastic nature of the sources. However, the inclusion of renewable energy resources in an islanded microgrid is attractive, since they can reduce the use of fossil fuel. Unit power costs for fossil generation are usually very high for islanded microgrids because power is typically supplied by diesel-generating sets with inherently high costs due to transportation of the fuel to remote islands and locations. There are a number of limitations to the attainable level of renewable energy penetration in isolated grids. These limitations are described below and include spinning reserve requirements, conventional generator minimum loading, conventional generator step-load response, system stability, and reactive power and voltage control requirements [44]. 10.3.4.1 Spinning Reserve Spinning reserves are needed to account for unscheduled or scheduled loss of DER [45]. The intermittent nature of wind and solar requires slightly higher spinning reserves. However, 1:1 reserves are typically not needed. Depending on the degree of wind/solar penetration, these reserves could reach 18%. 10.3.4.2 Minimum Loading of Conventional Generators Some rotating machines require minimum loading depending on the type of prime mover used. This particularly applies to diesel engine–driven generators, which might require up to 40% minimum load to function properly. 239 Microgrids 10.3.4.3 Step-Load Requirements Different generating sources can tolerate and respond to step-load changes (sudden increases or drops in active power demand). The mechanical inertia of prime movers and the speed at which prime movers can be adjusted to the changing electrical load play an important role. Hydrogenerators respond rather slowly (orders of seconds or tens of seconds), whereas PE-connected sources (wind or solar) are extremely fast (orders of milliseconds). In case of a sudden increase in active power demand, one has to take into account the maximum available power from the prime movers, which in the case of wind and solar is changing and could be below the maximum-rated power. 10.3.4.4 Power System Stability and Oscillations The stability of power systems depends on the inertia of the prime movers (as described above) as well as the level of attenuation (damping) that the system exhibits during a disturbance. Large power-generating stations utilize power system stabilizers (PSS), which help in damping out possible system swings. Smaller machines (diesel gensets, for example) typically do not employ these devices. On the other hand, the PE-connected sources (wind and solar) can support the system stability if their controls can perform the PSS equivalent functions. 10.3.4.5 Reactive Power Requirement and Voltage Control In large-scale power grids, the typical power factor is lagging (inductive) and of the order of 0.8–0.9 or better. In microgrids, the variety of generating sources, shorter or no transmission/distribution lines, and more complex PE loads make the reactive power requirements more complex. It does not necessarily mean that the power factor is lower than in the macrogrid. It only implies that the controls of microgrids (either distributed or centralized) have to be able to monitor, calculate, and respond to the more complex reactive power controls and voltage regulation. 10.3.4.6 Limitations to Renewable Energy Penetration It is believed that the larger the degree of renewable penetration the higher the requirements for spinning reserves (see also Section 10.3.4.1). In [44] the formula given is δmax = 1 – Pmin (1 − PSR ) where: δ represents the percentage of renewable energy penetration δmax is the maximum possible percentage of renewable energy penetration, given unlimited renewable power available Pmin is the minimum conventional generator loading PSR is the spinning reserve requirement of the power system Figure 10.8 [44] illustrates an example of the achievable penetration δ for a system where PSR = 20%, Pmin = 40%, and the conventional generator size is 50 kW. In this case, δmax is only 50%. 240 Smart Grids RE penetration (%) 60 50 40 30 20 10 0 0 50 100 150 Pload (kW) 200 250 300 FIGURE 10.8 δ versus Pload for a system where PSR = 20% and Pmin = 40%. 10.3.4.7 Short-Term Energy Storage Helps Renewable Energy Penetration Understandably, storing the energy generated by the intermittent sources (wind, solar) contributes to the smoothing effect and helps ride-through the power peaks and valleys [46]. An electric battery, or a compressed gas facility, a lywheel or a fuel-cell technology could be used for this purpose. Typically, the battery storage is capable of several minutes of active power generation, and the lywheel can typically cover 30 s or so [47]. For longer periods of energy storage, diesel gensets or fuel cells can be utilized. One has to remember here that these different sources have also different ramp-up times: batteries have the shortest time (milliseconds) where fuel cells require several hours to come online. 10.4 MICROGRID ECONOMIC VALUE STREAMS The cost element involved in the implementation of microgrids is a widely discussed topic among consumers and utilities. As with any capital project, implementing a microgrid involves an investment in infrastructure, including the power sources and the technologies needed to manage and connect the microgrid to the main grid. However, the capital outlay required for a microgrid is often much less and payback is signiicantly faster than for other initiatives intended to improve the availability and reliability of power. The payback for a microgrid is often measured in years, if not months, whereas the payback for a new utility-scale power generation might be measured in decades and the project might face a barrage of regulatory hurdles before it can even begin. A number of factors can help offset the cost of a microgrid: • Reduced Overall Energy Costs: One of the most important economic value streams for microgrids is certainly the reduced dependence on grid-sourced electricity and utility T&D services, although this ability to reduce dependence will result only if “average cost of internally generated power is less than grid sourced power” [26]. • Sales of Excess Power to the Macrogrid: In some regions, energy markets allow microgrid operators to sell the excess power generated, making the 241 Microgrids • • • • 10.5 microgrid a potential source of revenue. In addition, the heat generated from the source powering the microgrid can be used to create an additional stream of revenue for a business. For example, steam might be used to power up additional generators, or hot water could be used for absorption chilling [26]. Greater Fuel Choice: While wind turbines and diesel generators are probably the two most common sources of power for established microgrids, there are many choices available. New technologies and new discoveries are driving down the cost of natural gas in some areas. Solar is becoming more eficient and less costly. While fuel cells are still relatively expensive, they are low maintenance, and research continues to ind new ways to make them commercially viable. Leveraging a biomass fuel source may allow the company to claim carbon credits. These and other sources of energy can be combined as the consumer’s needs, technology, and market dynamics evolve [1]. Reduced T&D Losses: As much as 10% of energy is lost in transmission and distribution. By their nature, microgrids are local, and the power consumed has less distance to travel from generation to consumer, thus reducing a signiicant percentage of T&D losses [1]. Participation in Organized Demand Response Markets: Demand response can be understood as a change in net energy consumption (loads, distributed generation, storage) by a customer in response to a signal (price, load, emissions). By virtue of the microgrid’s ability to precisely control sources of supply and demand in response to market prices or other signals, interconnected microgrids may be able to participate in organized demand response programs. This helps in reducing the cost of energy by charging the system with low-priced energy and discharging later when the energy prices are high [26]. Enhanced Electricity Price Elasticity: By reducing consumption of energy from the macrogrid, microgrids may be able to reduce the output from high marginal cost or “peaking” plants, thereby reducing the clearing price for electricity in wholesale energy markets. All power consumers—including utility ratepayers—will beneit from this value stream. RELIABILITY Reliability is deined as the probability that a device will perform its intended function during a speciied period of time under stated conditions. Mathematically, this may be expressed as ∞ R ( t ) = Pr {T > t} = ∫ f ( x ) dx t where f(x) is the failure probability density function and t is the length of time (which is assumed here to start from time zero). Reliability is therefore a function of time—the longer the time, the more likely it is that it will fail—and, of course, a function of the initial (factory) quality and design robustness of the device. Availability is the opposite of reliability; it is the probability that a device will be available (will not fail) for a given period of time. 242 Smart Grids Since it is dificult to know the reliability of a device before it is actually put into service, another, more practical way of describing reliability is by two parameters: failure rate (FR), measured in (number per period of time), and mean time to repair (MTTR), measured in (minutes, hours, days, etc.). A commonly used parameter, mean time between failures (MTBF), is the reciprocal of FR. Availability related to failures is called inherent availability and is expressed as Ai = MTBF MTBF + MTTR Although a product of FR and MTTR (if MTBF≫MTTR) could yield a measure of unavailability (i.e., [1−R(t)]*t), and it is often used in this way to describe devices and systems, this calculation could be easily misinterpreted. For example, if a device fails once over a period of 1 year and its MTTR is 1 day, its impact on the operational availability of the system could be dramatically different from that of a device that fails 24 times in 1 year with an MTTR of 1 h, or a device that fails 1000 times per year with an MTTR of 86.4 s. Systems such as microgrids are made up of components (devices), which all contribute to reliability. However, the degree to which these components play a role in the overall reliability depends on a few factors, such as (a) topology, that is, the location of the device within the microgrid, and the degree of redundancy built into the system; and (b) the degree of criticality of the device for the mission (failures of some devices, such as CBs, could result in immediate loss of service, whereas failures of some other devices, such as power quality meters, could be tolerated until the next scheduled maintenance, for example). 10.5.1 relIaBIlIty due to maIntenance So far, reliability has been deined in terms of occurrence of failures, that is, unexpected events. To complete the analysis, one must also include maintenance of devices (expected, planned events), since this also contributes to the total system (un)availability. This can be easily done by quantifying maintenance frequency (MF) and maintenance duration (MD), parameters equivalent to FR and MTTR. MF is often expressed as the reciprocal of the mean time between maintenance (MTBM) and MD is also called mean down time (MDT). Operational availability is deined similarly as Ao = 10.5.2 MTBM MTBM + MDT mIcrogrId system relIaBIlIty There is no one single reliability number for a whole microgrid as a system. Instead, the reliability should be deined at speciic points, typically where loads are connected [27]. One has to realize that the same microgrid can have different reliability performance at different load locations. Topology, load requirements, and state of the microgrid all play a role. Microgrids 243 In order to deine the reliability at any given point of the microgrid, one has to combine the reliability data of all the devices and components as seen from the point of interest. There are basically two approaches. The irst approach is to use analytical formulae to combine all the component data based on their interconnectivity, in series or in parallel. The entire network can be then simpliied to a single set of two numbers, outage frequency (OF) and outage duration (OD). If alternative, nonconcurrent paths exist to the given point of interest (where the load is connected), one has to perform this process for both paths and weigh the OF and OD according to the relative share of each path in which the microgrid will operate, for example, 30% in path A and 70% in path B. 10.5.3 can relIaBIlIty oF a mIcrogrId Be Better than that oF a utIlIty grId? There is no simple answer to this question [28–32]. First of all, renewable resources such as wind and solar have mixed availability due to natural factors such as weather [30,31]. Although they might not experience many equipment failures, the natural factors can be interpreted as unavailability of a resource. On the other hand, the fact that the microgrid is rather small and consists of a relatively low number of components can have a positive inluence on reliability; fewer components mean fewer possibilities of failure. Yet, in the islanded mode, a microgrid does not have the support (or backup) of a utility grid behind it, and this can decrease the reliability of microgrid service. Yet another aspect has to do with the energy storage [29]. If such equipment is used, it can really improve the reliability of the service if it is properly designed and located within the microgrid, and if the appropriate control logic is implemented. 10.6 OUTLOOK: MICROGRIDS—ENERGY FOR THE DIGITAL SOCIETY The notion of a “microgrid” has been around for more than a decade [33,34] in the power engineering community. Yet these grids are far from being commercially prevalent (see DOE data, circa 2012, on installed microgrids), even though the improved supply reliability—by disconnecting from a stronger utility grid if this grid experiences operational problems—for end customers that own or are part of the microgrid is technically sound. T&D utilities and medium- to large-sized consumers have been recognizing the technical beneit for some time, but the economics, in combination with existing utility grid supply reliability, have not yet been right for microgrids to lourish. The idea of splitting up a grid into islands for increased reliability is not new, and it is employed, coupled with continuing research, to secure—or limit the negative impact on—an interconnected transmission grid in emergency situations such as a spreading blackout. By splitting up the system, parts of the network can survive, as was illustrated in the 2003 northeast blackout in North America, when New England disconnected from its neighbors and kept the lights on for its customers. The T&D space has been slow to embrace microgrids, and so far all that has materialized are generally either small examples where disconnection was not explicitly 244 Smart Grids 2010 Washington Maine Montana North Dakota Minnesota VT NH Oregon ldaho Wisconsin South Dakota Michigan Wyoming Nebraska Nevada Utah CT RI Pennsylvania New Jersey Iowa Ohio Delaware West DC Virginia Virginia Kentucky North Carolina Tennessee Illinois Indiana Colorado Kansas California Arizona MA New York Oklahoma New Mexico Missouri Arkansas Mississippi Alabama South Carolina Georgia Texas Louisiana Alaska Florida DOE R&D DOE/DoD Peak reduction ARRA SGDP Industry led Hawaii 2012 Maine Washington Montana North Dakota Minnesota VT NH MA New York CT RI Oregon ldaho South Dakota Wisconsin Michigan Wyoming Nebraska Nevada Utah New Mexico Ohio Delaware West Virginia DC Virginia Kentucky North Carolina Tennessee Illinois Indiana Colorado Kansas California Pennsylvania New Jersey Iowa Missouri Oklahoma Arkansas Arizona South Carolina Alabama Georgia Mississippi Texas Louisiana Alaska Hawaii Florida R&D test beds (DOE) R&D test beds (Non-DOE) DOE/DoD (Assessments and SPIDERS) DoD ESTCP Peak load reduction ARRA SGDP Industry/utility/ university/other fed led FIGURE 10.9 US microgrid landscape in 2010 and 2012. (From Smith, M., Ofice of Electricity Delivery and Energy Reliability smart grid R&D program, in 2012 DOE Microgrid Workshop, Summary Report, 30–31 July, Chicago, IL, 2012. For 2010: http://der.lbl.gov/sites/ der.lbl.gov/iles/Smith_2010.pdf; for 2012: http://e2rg.com/microgrid-2012/lnt%27l_USA_ Smith.pdf.) 245 Microgrids allowed, but a net-zero power/energy exchange was demonstrated, or examples of permanently islanded systems (i.e., mini power systems such as the Azores Islands, where ABB/PowerCorp showcased their solution, as reported in [35]). Figure 10.9 shows that the US DOE has been tracking microgrid installations, and as of 2012 it is clear that the numbers have been steady. The United States is the leading microgrid market today, and is expected to remain that way until at least 2020. Another microgrid hotspot is the Asia Paciic market, where efforts have intensiied with the aim of ending energy poverty through remote/ off-grid microgrids in China and India. The market for microgrid technologies in Europe is expected to expand in the near future as the region updates an aging electricity infrastructure and integrates a higher share of renewable energy generation. Over the past 5 years, an increased interest in microgrids in scientiic and industrial arenas has been observed, which can be seen in the upward trending of “Google” hits (Figure 10.10). The same accelerated trends have been identiied in [36] for a total microgrid revenue by forecast scenario in 2013–2020 (Figure 10.11). The main drivers for the increasing number of microgrid installations are as follows (the ranking may differ for different regions): • Increased penetration of local DERs (motivated by government incentives and falling costs of technology) • Increased need for reliability and energy independence, even in the presence of natural disasters—extreme weather events causing massive power outages will continue in the future—or homeland security events, such as military microgrids and data centers • Push from federal and state regulators to develop, implement, and uphold various renewable portfolio standards • Awareness of opportunities to participate in competitive energy markets by selling lexible demand, ancillary services, etc. • Technology push by a large number of companies offering hardware and software control and protection technologies applicable to microgrids One can think of a microgrid as a ixed-boundary virtual power plant, which by itself can participate in the grid management of T&D systems. Due to this active participation, microgrid operators or owners would like to be adequately compensated for their contribution. The business case input and evaluation methodology for these 100 80 60 40 20 2008 2009 2010 2011 FIGURE 10.10 Microgrid-related hits in Google trends. 2012 2013 246 Smart Grids 70,000 Base scenario Average scenario 60,000 Aggressive scenario ($ millions) 50,000 40,000 30,000 20,000 10,000 2013 FIGURE 10.11 2014 2015 2016 2017 2018 2019 2020 Total microgrid revenue by forecast scenario, world markets: 2013–2020. types of microgrids is quite similar to evaluating other interfacing grid technologies, such as utility-scale energy-storage systems. Some of the interesting applications of microgrids mentioned recently are: • Data Centers: where the data center owner would be interested in predominantly net-zero energy operation, but would also have the ability to isolate from the utility supply grid to maintain a connection to the information cloud. • Buildings: owners are interested in net-zero operation and maintaining reliable energy supply to critical loads such as elevators. Disconnection from a grid could be a possibility. • Military Bases: these are interested in reliable power supply for critical loads and reduced dependency on imported fossil fuel (especially for remote and isolated bases). For the bases connected to utility-scale grids, due to policy on enforced renewable portfolio standards, the military aim to self-supply and limit their dependence on the utility grid. • Suburbs or Planned Community Developments: these might want to form their own microgrid in order to survive or get up and running more quickly after storms. This could make sense, seeing that the infeed to the community might be overhead lines, prone to circuit damage from falling trees, whereas inside the community the distribution might be underground and less prone to damage. Hence, after an event it is conceivable that a community with suficient local generation could self-supply while the utility distribution grid is being repaired. • Cities: in China, due to overproduction of solar panels, the government is starting to mandate more local installation of PV and is requesting cities to be designed as microgrids. From the above list, which is by no means exhaustive, it is clear that a microgrid owner or operator would in general aim for net-zero, or, if suficient local capacity is Microgrids 247 available, net-supplier energy operation with the larger grid. Once this is achievable, operators would also aim to disconnect from the grid when necessary, if the larger grids are experiencing issues, in order to increase their overall reliability. It is fair to say that there are indications that microgrids have graduated from a mainly academic topic to one that is of interest to various industries. Technology building blocks exist, and reinements of these technologies are set to continue, but one of the biggest breakthroughs comes when the economics are starting to make more sense and novel business models are emerging. Understanding the economics of developing and operating a microgrid in detail is essential for success in this new and rapidly evolving market. REFERENCES 1. M. MacDonald, Microgrids: Weathering the storm, June 2013. Available at: http:// www05.abb.com/global/scot/scot299.nsf/0/f49786ab08e927ff85257c4000554a40/$f ile/abb-800-wpo_microgrids_weathering_the_storm_11-8-13.pdf. 2. Microgrids: From data centers to developing countries, an enticing solution for self-suficient power, January 2012. Available at: http://www.lead-central.com/ AssetManager/e0bc9e75-84f1-4ca1-9c65-21c74a8b75fe/Documents/New%20WP%20 ile/ABB-73-WPO_Microgrids.pdf. 3. US Congress, National Defense Authorization Act for Fiscal Year 2007, 16 January, 2012. Available at: https://www.govtrack.us/congress/bills/109/s2507. 4. EU More Microgrids Project, Alternative microgrids network structures, Project Deliverable DC1, November 2007. Available at: http://www.microgrids.eu/documents/643.pdf. 5. A. Chowdhury, D. Koval, Current practices and customer value-based distribution system reliability planning, IEEE Transactions on Industry Applications, 40(5), 1174–1182, 2004. 6. A. Dimeas, N. Hatziargyriou, Operation of multiagent system for microgrid control, IEEE Transactions on Power Systems, 20(3), 1447–1455, 2005. 7. Y. Xu, “ORNL Power Systems and Power Electronics for Utility Applications”, ORNL, Workshop on Power Electronics and Distributed Power systems, UTRC, December 2, 2011. Available at: http://web.ornl.gov/sci/ees/etsd/pes/pubs/UTRCWorkshop_ ORNLPowerElectronics_Microgrid.pdf. 8. A. Tsikalakis, N. Hatziargyriou, Centralized control for optimizing microgrid operation, in Proceedings of the IEEE Power and Energy Society General Meeting, pp. 1–8, 24–29 July, San Diego, CA, 2011. 9. GE Digital Energy, Microgrid control system—Optimization solution for permanently Islanded and grid-connected microgrids, Technical Brochure, 2012. Available at: http:// www.gedigitalenergy.com/products/brochures/MicrogridControlSystem.pdf. 10. ABB, Renewable microgrid controller, Technical Brochure, 2012. Available at: http://www05.abb.com/global/scot/scot221.nsf/veritydisplay/d66c5a59934a10b7c1 257c37002cc707/$ile/9AKK100580A2552_Microgrid%20Controller%20600_EN_ LR(Dic2013).pdf. 11. P. Piagi, R. Lasseter, Autonomous control of microgrid, in Proceedings of the IEEE Power Engineering Society General Meeting, 8–22 June, Montreal, QC, 2006. 12. W. Feero, D. Dawson, J. Stevens, Protection issues of the microgrid concept. Available at: http://certs.lbl.gov/pdf/protection-mg.pdf. 13. H. Al-Nasseri, M.A. Redfern, R. O’Gorman, Protecting microgrid systems containing solid-state converter generation, in Proceedings of the International Conference on Future Power Systems, p. 5, 18 November, Amsterdam, The Netherlands, 2005. 248 Smart Grids 14. J. Driesen, P. Vermeyen, R. Belmans, Protection issues in microgrids with multiple distributed generation units, in Proceedings of the 4th Power Conversion Conference, pp. 646–653, 2–5 April, Nagoya, Japan, 2007. 15. H. Nikkhajoei, R.H. Lasseter, Microgrid protection, in Proceedings of the IEEE Power Engineering Society General Meeting, pp. 1–6, 24–28 June, Tampa, FL, 2007. 16. EU More Microgrids Project, Novel protection systems for microgrids, Project Deliverable DC2. Available at: http://www.microgrids.eu/documents/688.pdf. 17. A. Oudalov, A. Fidigatti, Adaptive network protection in microgrids, International Journal of Distributed Energy, 5(3), 201–225, 2009. 18. N. Schaefer, T. Degner, A. Shustov, T. Keil, J. Jaeger, Adaptive protection system for distribution networks with distributed energy resources, in Proceedings of the 10th IET International Conference on Developments in Power System Protection (DPSP), pp. 1–5, 29 March–1 April, Manchester, UK, 2010. 19. F. van Overbeeke, Fault current source to ensure the fault level in inverter-dominated networks, in Proceedings of the 20th International Conference and Exhibition on Electricity Distribution—Part 1 (CIRED), pp. 1–4, 8–11 June, Prague, Czech Republic, 2009. 20. D. Ishchenko, A. Oudalov, J. Stoupis, Protection coordination in active distribution grids with IEC 61850, in Proceedings of the IEEE PES Transmission and Distribution Conference and Exposition (T&D), pp. 1–6, 7–10 May, Orlando, FL, 2012. Available at: http://www.informatik.uni-trier.de/~ley/db/journals/tie/tie58.html. 21. D.J. Becker, B.J. Sonnenberg, DC microgrids in buildings and data centers, in Proceedings of the IEEE 33rd International Telecommunications Energy Conference (INTELEC), pp. 1–7, 9–13 October, Amsterdam, The Netherland, 2011. 22. C. Marnay, S. Lanzisera, M. Stadler, J. Lai, Building scale DC microgrids, in Proceedings of the IEEE Energytech, pp. 1–5, 29–31 May, Cleveland, OH, 2012. 23. J.-F. Hansen, J.O. Lindtjorn, T.A. Myklebust, K. Vanska, On-board DC grid, newest design for marine power and propulsion systems, ABB Review, 2, 28–33, 2012. 24. H. Kakigano, Y. Miura, T. Ise, Distribution voltage control for DC microgrids using fuzzy control and gain-scheduling technique, IEEE Transactions on Power Electronics, 28(5), 2246–2258, 2013. 25. A. Kwasinski, C.N. Onwuchekwa, Dynamic behavior and stabilization of DC microgrids with instantaneous constant-power loads, IEEE Transactions on Power Electronics, 26(3), 822–834, 2011. 26. M.A. Hyams, Microgrids: An assessment of the value, opportunities and barriers to deployment in NY State, in New York State Energy Research and Development Authority (NYSERDA), pp. 77–83, September, 2010. 27. E.N. Dialynas, Evaluating the approximate probability distributions of load point reliability indices in power distribution networks, IEE Proceedings C, 135(5), 450–460, 1988. 28. Certs Microgrid Laboratory Test Bed, prepared for: California Energy Commission Public Interest Energy Research Program, prepared by: Lawrence Berkeley National Laboratory, February 2009, CEC-500-2009-004. 29. M. Stadler, C. Marnay, A. Siddiqui, J. Lai, B. Coffey, H. Aki, Effect of Heat and Electricity Storage and Reliability on Microgrid Viability: A Study of Commercial Buildings in California and New York States, Environmental Energy Technologies Division, Lawrence Berkeley National Laboratory, LBNL-1334E-2009, Berkeley, 2009. 30. E.N. Dialynas, L.G. Daoutis, C. Toufexis, I. Charalambous, Reliability and reserve capacity assessment of isolated power systems with increased penetration of renewable energy sources, in Proceedings of the 7th Mediterranean Conference and Exhibition on Power Generation, Transmission, Distribution and Energy Conversion, pp. 1–7, 7–10 November 2010, Agia Napa, Cyprus (Paper No. MED10/242). Microgrids 249 31. E.N. Dialynas, D.J. Papakammenos, N.C. Koskolos, Integration of nonutility generating facilities into the reliability assessment of composite generation and transmission power systems, IEEE Transactions on Power Systems, 12(1), 464–470, 1997. 32. H. Liang, J. Su, S. Liu, Reliability evaluation of distribution system containing microgrid, in Proceedings of the China International Conference on Electricity Distribution, pp. 1–7, 13–16 September, Nanjing, China, 2010. 33. T.E. Hoff, H.J. Wenger, C. Herig, R.W. Shaw Jr, Distributed generation and microgrids, in Proceedings of the 18th Annual North American Conference of the USAEE, San Francisco, CA, September 1997. 34. R. Lasseter, P. Paigi, Microgrid a conceptual solution, in Proceedings of the IEEE 35th Power Electronics Specialists Conference, vol. 6, pp. 4285–4290, 20–25 June, Aachen, Germany, 2004. 35. M. Smith, Ofice of Electricity Delivery and Energy Reliability smart grid R&D program, in 2012 DOE Microgrid Workshop, Summary Report, 30–31 July, Chicago, IL, 2012. Available at: (a) http://energy.gov/oe/downloads/2012-doe-microgrid-workshop-summary-report-september-2012, (b) http://der.lbl.gov/sites/der.lbl.gov/iles/ Smith_2010.pdf, (c) http://e2rg.com/microgrid-2012/lnt%27l_USA_Smith.pdf. 36. P. Asmus, A. Lauderbaugh, M. Lawrence, Navigant Research, Market data: Microgrids, Executive Summary, 2013. Available at: http://www.navigantresearch.com/wp-assets/ uploads/2013/03/MD-MICRO-13-Executive-Summary.pdf. 37. G.D. Rockefeller, C.L. Wagner, J.R. Linders, K.L. Hicks, D.R. Rizi, Adaptive transmission relaying concepts for improved performance, IEEE Transactions on Power Delivery, 3(4), 1446–1458, 1988. 38. B. Kroposki, C. Pink, J. Lynch, V. John, S.M. Daniel, E. Benedict, I. Vihinen, Flexible DER utility interface system, Final Technical Report NREL/TP-560-39876, August 2006. Available at: http://www.nrel.gov/docs/fy06osti/39876.pdf. 39. EU More Microgrids Project, Decentralized control concepts for microgrids, Project Deliverable DB3, February, 2008. Available: http://www.microgrids.eu/documents/641.pdf. 40. C. Yuen, A. Oudalov, A. Timbus, The provision of frequency control reserves from multiple microgrids, IEEE Transactions on Industrial Electronics, 58(1), 173–183, 2011. Available at: http://www.informatik.uni-trier.de/~ley/pers/hd/t/Timbus:Adrian.html. 41. A. Bagherian, S.M.M. Tafreshi, A developed energy management system for a microgrid in the competitive electricity market, in Proceedings of the IEEE Bucharest Power Tech Conference, pp. 1–6, 28 June–2 July, Bucharest, Romania, 2009. 42. H. Kanchev, D. Lu, B. Francois, V. Lazarov, Smart monitoring of a microgrid including gas turbine and a dispatched PV-based active generator for energy management and emission reduction, in Proceedings of the IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT Europe), pp. 1–8, 11–13 October, Gothenburg, Sweden, 2010. 43. D. Lu, B. Francois, Strategic framework of an energy management of a microgrid with a photovoltaic-based active generator, in Proceedings of the 8th International Symposium on Advanced Electromechanical Motion Systems and Electric Drives Joint Symposium, pp. 1–6, 1–3 July, Lille, France, 2009. 44. N. Hamsic, A. Schmelter, A. Mohd, E. Ortjohann, E. Schultze, A. Tuckey, J. Zimmermann, Increasing renewable energy penetration in Isolated grids using a lywheel energy storage system, in Proceedings of the International Conference on Power Engineering, Energy and Electrical Drives (POWERENG), pp. 195–200, 12–14 April, Portugal, Setúbal, 2007. 45. Y. Rebours, D. Kirschen, What is Spinning Reserve?, The University of Manchester, Manchester, UK, 2005. 46. H. Bindner, Power Control for Wind Turbines in Weak Grids, Riso National Lab, Roskilde, Denmark, 1999. 47. A. Ruddell, G. Schönnenbeck, R. Jones, Flywheel Energy Storage Systems, Rutherford Appleton Lab, UK, 1998. 11 Integrating Consumer Advance Demand Data in Smart Grid Energy Supply Chain Tongdan Jin, Chongqing Kang, and Heping Chen CONTENTS 11.1 Introduction .................................................................................................. 252 11.2 Smart Grid Technology: The State of the Art .............................................. 252 11.2.1 Distribution Automation ................................................................... 253 11.2.2 Distributed Energy Generation and Integration ............................... 253 11.2.3 Optimal Asset Management ............................................................. 254 11.2.4 Advanced Metering Infrastructure ................................................... 254 11.2.5 Demand-Side Management .............................................................. 254 11.3 Principle of Virtual Energy Provisioning ..................................................... 255 11.3.1 Revisiting Manufacturing Supply Chain Model ............................... 255 11.3.2 Virtual Energy Supply Chain Model ................................................ 256 11.4 System Architecture and Operational Conditions ........................................ 257 11.4.1 System Architecture ......................................................................... 257 11.4.2 Operational Conditions ..................................................................... 258 11.5 Beneiting Operation of Smart Grid System................................................. 258 11.5.1 Facilitating Operation of Distributed Generation Systems .............. 258 11.5.2 Optimizing Asset Management ........................................................ 259 11.5.3 Enhancing Load Forecasting ............................................................ 259 11.5.3.1 Discrete Bayesian Prediction .............................................260 11.5.3.2 Continuous Bayesian Prediction ........................................ 262 11.5.3.3 Bayesian Prediction with Normal Distributions ................ 263 11.6 Consumer Performance Assessment ............................................................ 263 11.6.1 Minimizing Mean Square Error .......................................................264 11.6.2 Composite Performance Assessment Metrics ..................................264 11.6.3 Illustrative Example .......................................................................... 265 11.7 Communications and Automation Technology ............................................266 11.8 System Preliminary Design .......................................................................... 269 11.9 Conclusion .................................................................................................... 271 References .............................................................................................................. 272 251 252 11.1 Smart Grids INTRODUCTION Smart grids are enabling technologies featuring high reliability, self-healing, full controllability, and optimal asset management. By combining intelligent control algorithms with advanced communication and information technology, smart grids allow utility suppliers to manage and control energy generation and distribution in real time. As such, the network security, reliability, and eficiency are signiicantly improved, beneiting all stakeholders, including generators, distributors, marketers, and end consumers. A key technology for smart grid deployment is demand-side management (DSM) technology. Leveraging communication and information technology, the purpose of DSM is to allow consumers to directly participate in the grid operation through various means such as load shifting, direct load control, dynamic pricing, and other demand response programs. This chapter will introduce a novel DSM concept, virtual energy provisioning (VEP), to manage the energy supply chain from generation to delivery and inal consumption. In a manufacturing supply chain system, the production typically is driven by advance demand information from the retailers or end customers. Similarly, the power system can be treated as an energy supply chain system in which the decision on electricity generation and distribution is made based on the consumption of end consumers. However, existing energy systems usually do not take into account the consumers’ advance demand information. In other words, electricity is generated according to forecasts based on historical consumption. Since consumers are the ultimate decision makers regarding future consumption, it is imperative for the utility supplier to let consumers participate in the forecasting process. To that end, we propose an online VEP system in which consumers are able to provide advance demand information so that better and more realistic forecasts can be made by the utility supplier. The utility company, after aggregating consumers’ advance demand data, can decide on an optimal generation and distribution scheme. This chapter will discuss the VEP concept, the operational condition, and the preliminary implementation. The remainder of the chapter is organized as follows. Section 11.2 provides an overview on recent advances in smart grid initiatives, including DSM technology. Section 11.3 describes the principle of the VEP concept. Section 11.4 explains the operational condition of the VEP system. In Section 11.5, the system beneits will be discussed, focusing on asset management and distributed generation (DG) integration. Section 11.6 elaborates on the consumer performance assessment metrics. Section 11.7 investigates modern communication technology and its impact on VEP implementation. In Section 11.8, the preliminary implementation is presented. Section 11.9 concludes the chapter with some remarks on future research. 11.2 SMART GRID TECHNOLOGY: THE STATE OF THE ART Smart grids are enabling technologies aiming to resolve major challenges arising from increased energy demands, the aging utility infrastructure, and growing concerns over the use of fossil fuels. Smart grids have been characterized or deined by many organizations as Intelligent Grid (or Intelligrid) or Future Grid [1]. In order Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 253 to build intelligent features into the existing electric network, a comprehensive technological suite comprised of real-time monitoring, forecasting, decision making, controls, and optimization algorithms is necessary [2]. From the implementation perspective, smart grid technologies include (1) distribution automation; (2) distributed energy generation and integration; (3) optimal asset management; (4) advanced metering infrastructure (AMI); and (5) DSM. In conjunction with the deployment of information technology and telecommunication networks, smart grids allow utility companies to monitor and optimize the production and distribution of electric energy in real time. As a result, the system, reliability, power quality, and energy eficiency are consistently guaranteed, beneiting all the stakeholders, including generators, utilities, market traders, and end consumers. 11.2.1 dIstrIButIon automatIon Distribution automation (DA) includes monitoring, control, and communication functions located on the electrical feeder [1]. Protection and switching functions are among the most important aspects when we design an effective DA mechanism. Nowadays, various intelligent DA devices have been deployed in distribution lines to track current and voltage states, exchange device information, and reconigure the network to meet the changing environment. The topology in a smart grid is anticipated to be more lexible and adaptive, allowing the reconiguration of the topology upon the occurrence of faults, excessive load demands, and malicious attacks. Another important DA feature is automated fault diagnosis by monitoring the grid fault, identifying the root cause, and restoring the system. Automated fault and rootcause identiication has also been investigated using cause–effect network, fuzzy rules, artiicial intelligence, or Bayesian inference [3–5]. The deployment of wireless sensor networks, distributed actuators, and ubiquitous communication technologies on smart grids would generate more accurate real-time data about the system states, making automated fault diagnoses feasible and applicable in future grids. 11.2.2 dIstrIButed energy generatIon and IntegratIon DG produces and supplies electricity from many small energy sources, which are often referred to as distributed energy resources (DERs). Typical DER units include wind turbines, solar photovoltaics (PV), fuel cells, microturbines, and internal combustion engines. The capacity of a DER unit ranges from 3 kW to 10 MW, while the capacity of a centralized generation unit usually exceeds 500 MW. Since energy generated from DER units is often consumed locally, there is no need to build long transmission lines to transport the electricity to end consumers, reducing costs signiicantly. Wind power and solar PV emerge as sustainable energy resources to meet the increased electricity demand in the next two decades. It is anticipated that by 2030 about 20% of US electricity will be supplied from wind turbines, compared with the current 2% market share [6]. However, wind and solar energies are quite intermittent, and the power output depends on the weather conditions, the geographical location, and the season of the year. The techniques for integrating renewable 254 Smart Grids energy resources into the distribution networks have been extensively studied in the literature, with the objective of designing robust and adaptive DG systems. For a comprehensive survey, readers are referred to [7]. 11.2.3 oPtImal asset management The strategy of asset management can be perceived as a process to maximize the return on investment of the equipment over its life cycle [8]. Decision makers often leverage operations research and probability theory to identify the best asset management strategy that maximizes the long-term proits or minimizes the equipment life-cycle cost. Utilities have aggressively devised various asset maintenance strategies to monitor, predict, and extend the remaining useful lifetime of the equipment [9,10]. Advanced sensory technology combined with data mining provides ample opportunities to characterize and forecast the remaining useful life of wind turbines, solar PV, and transformers [11,12]. In smart grids, more intelligent sensors will be built into the equipment to monitor the performance parameters so that conditionbased maintenance will be carried out immediately before the occurrence of equipment failures. 11.2.4 advanced meterIng InFrastructure Communication systems play a pivotal role in transmitting metering data, publishing price information, and sending grid protections and control commands. Power line communication, broadband Internet, and wireless meshes are three mainstream communication technologies currently used in the power industry. Ostertag and Imboden [13] propose a hybrid power line communication system to support both the DA and the DSM functions using a priority-based medium access protocol. In [14], it is reported that Texas New Mexico Power is using AT&T’s digital wireless cellular network to collect metering data from 10,000 residential smart meters. Debates still continue with respect to the selection of communication infrastructures for AMI deployment. Power line communication has been demonstrated to be a reliable and secure communication technology, yet the transmission rate is quite limited. Internet-based AMI systems seem very promising because of the maturity of the technology and the ubiquity of network infrastructure. The concern with the Internet is reliability, security, and robustness when faced with high-volume metering transmission and malicious cyberattacks. 11.2.5 demand-sIde management DSM is the key to the effective implementation of the smart grid concept. The purpose of DSM is to develop and deploy new technologies for achieving power eficiency, energy conservation, and demand responsiveness. Many studies have been published to explore various types of DSM concept along with the implementation technology. These technologies include optimal usage scheduling [15], load shifting [16], direct load control [17], dynamic pricing [18], and smart metering and appliances [19], among others. In [20], Zhong et al. reviewed the successes and challenges Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 255 during the deployment of DSM technology in China, one of the fastest-growing countries in the world in terms of energy consumption. Dynamic pricing such as real-time pricing (RTP) and time-of-use (TOU) encourages consumers to alter their energy usage patterns in order to reap inancial beneits. The implementation of AMI makes dynamic pricing possible. Recently, many studies have been carried out to evaluate the impacts of dynamic pricing on electricity consumption patterns. For a comprehensive review, interested readers are referred to [21]. These methods usually apply game theory or develop inancial incentives to guide consumers to reach a delicate balance between generation and consumption in peak hours. 11.3 PRINCIPLE OF VIRTUAL ENERGY PROVISIONING The power system is considered healthy and reliable only when all demands are appropriately satisied or the interruption is kept at a minimum level. It is almost impossible to ind a global optimal algorithm to control such a complex and dynamic system based on today’s knowledge [2]. One viable approach is to leverage information technology and proactively collect future demand information from consumers. The utility company, after analyzing and aggregating consumers’ advance demand information, decides an optimal generation and distribution scheme to meet their actual needs. In fact, the idea of consolidating consumers’ advance demand information is not new in the literature. In a manufacturing setting, the production and distribution of goods are often driven by advance orders placed by end customers or retailers. However, it is quite novel when this order-then-production or make-toorder concept is extended and applied to the energy industry. In the following, we describe an Internet-based demand management system, called On-line Purchase Electricity Now (OPEN) system, which is genially proposed in [22]. Through the OPEN system, consumers are able to send advance demand information to the utility supplier before they actually consume the electricity. 11.3.1 revIsItIng manuFacturIng suPPly chaIn model Figure 11.1 depicts a typical manufacturing supply chain system comprising customer orders, production, inventory, and delivery process. Obviously, the supply chain is driven or pulled by customers’ advance demand information. Upon receipt of customer orders, the manufacturer starts to produce the necessary goods to meet the demand. The inventory is widely used in manufacturing industries for temporarily storing inished goods before they are shipped to the retailers or end customers. Customer order Production Inventory or warehouse Delivery FIGURE 11.1 Manufacturing supply chain driven by customer orders. 256 Smart Grids An inventory usually involves three decision variables: the order quantity, the stocking level (i.e., quantity of goods), and the lead time. The lead time is the time interval between the arrival of the order and the instance of receiving the inished goods by the customer. Obviously, this closed-loop product supply chain mechanism remains stable as long as the inventory is not backordered. This stability can be easily maintained and sustained as long as the lead time is suficiently long. As such, the manufacturer has enough time to produce the goods to meet the customer demand. 11.3.2 vIrtual energy suPPly chaIn model Many similarities can be found between a manufacturing enterprise system and an energy production system. In the former, goods are produced and sold to the end customers, and the inventory is used as a buffer to mitigate any potential backorders in the presence of demand uncertainty. Similarly, electricity is generated by power plants or wind farms, and it is delivered to the end consumers via transmission/ distribution lines. Power interruptions or outages will occur if the load exceeds the generation. The problem can be mitigated or partially resolved if large storage systems can be created between the generation and the distribution lines. These storage systems, such as inventory systems, provide backup electricity to end consumers when the load exceeds the generation. At present, research is under way to develop eficient conversion or large storage technologies using lying wheels, batteries, and supercapacitors [23]. However, commercially large and inexpensive storage systems are still not mature enough for wide-scale applications. We propose a VEP concept to complement and enhance the physical storage technology. Conceptually, the energy provisioning system is operated as follows (see Figure 11.2). The utility company will design and implement an online database, the OPEN system, in which an individual account will be assigned to each consumer. Consumers can conveniently log on and access their web-based account via desktops, laptops, and mobile devices. The system allows consumers to place demand orders based on expected needs. These orders, representing future Supply chain under production-then-consumption Demand forecasting Generation Transmission and distribution Consumption Supply chain under order-then-production Consumer order Utility supplier Virtual energy provisioning Actual generation Transmission and distribution FIGURE 11.2 Generation-then-consumption versus order-then-generation. 257 Integrating Consumer Demand Data in Smart Grid Energy Supply Chain electricity demand, can be speciied on an hourly, daily, weekly, or monthly basis. The utility company, after aggregating all consumers’ orders, can decide an optimal generation and distribution scheme to meet the consumers’ needs. Just like a manufacturing supply chain system, this managed order-then-generation system is able to deliver the required electricity to the end consumers in a timely and reliable manner. In Section 11.6, various incentive programs are devised to motivate the consumers to place orders and then consume exactly the amount of energy that they ordered. The VEP model aims to achieve the same objective as the physical storage system, by delivering reliable and secure energy to end consumers. However, VEP does not store the electricity in physical media such as batteries or capacitors; rather, it leverages information technology to solicit advance demand information from consumers. The lead time between the demand order and the actual generation (i.e., consumption) allows the utility supplier to plan the optimal production scheme to meet the future need. 11.4 11.4.1 SYSTEM ARCHITECTURE AND OPERATIONAL CONDITIONS system archItecture Laptop Internet Power line Loading 1 SM Data mining Database Decision making Power generation Wireless meshes The OPEN system is a realization of the VEP concept described in the previous section. The OPEN system will be built on a hybrid architecture integrating existing utility infrastructure with the Internet platform. Figure 11.3 depicts a conceptual coniguration of the hybrid architecture, duplicated from [22]. Given the ubiquitous Internet technology, the deployment of the OPEN system only requires several costeffective software programs and hardware components. Software programs include an online database, a data mining engine, and a decision-making tool. The purpose of the database is to collect the demand information from end consumers as well as to store the actual consumption transmitted by the smart meter. The data mining engine is capable of forecasting the future load condition by combining consumers’ Smart phone PC Loading 2 Loading 3 (e.g., household) SM SM Smart meter Internet Power line FIGURE 11.3 A hybrid architecture for OPEN system implementation. (From Jin, T. and Mechehoul, M., IEEE Transactions on Smart Grid, 1, 302–310, 2010.) 258 Smart Grids orders with their historical consumption. The decision-making tool synthesizes the information and determines the optimal generation and distribution scheme. The major hardware component is the smart meter, which transmits the data from the consumer site to the utility supplier. The consumption data are primarily in a short message or packet format, representing hourly or daily energy consumption by the consumer. Since the transmission of the discrete data is not time-critical, transmission control protocol/Internet protocol (TCP/IP) should be suficient and reliable enough to undertake the communication task. Research [13,14] shows that such a communication function will be relatively easy to implement by directly integrating the smart meter with the Internet system or the wireless meshes. 11.4.2 oPeratIonal condItIons Consumer participation is the key to the effective operation of the OPEN system. Various incentive packages can be developed so that consumers will participate in the program and provide demand information. Fahrioglu and Alvardo [24] show that inancial incentives, when appropriately implemented, will motivate consumers to participate in the energy management system. For example, monetary discounts or energy credits could be used to reward the consumer whose demand order is equal or close to the actual consumption. In addition, the supplier can design customized usage portfolios to accommodate different users’ consumption behaviors. These incentive programs can be easily posted and distributed to individual consumers via the online database system. Another assumption is that dynamic pricing, such as TOU or RTP, is an effective tool to inluence and guide the consumer’s usage behavior. Studies [19,25] have shown that dynamic pricing is an effective tool to control and minimize demand uncertainty. Like an adaptive controller, dynamic pricing establishes a closed-loop feedback between the suppliers and their consumers, and brings the actual demand close to the generation via game theory [26]. Eventually, a long-term balance between generation and consumption is reached. 11.5 BENEFITING OPERATION OF SMART GRID SYSTEM The utility supplier is able to gain substantial beneits from the deployment of the OPEN system. These beneits include (1) facilitating the integration and operation of DG systems; (2) optimizing asset management; and (3) improving demand forecasting. These beneits will be explained in the following paragraphs, and, for more detailed discussions, readers are referred to [27]. 11.5.1 FacIlItatIng oPeratIon oF dIstrIButed generatIon systems Wind turbines and solar PV have emerged as clean and sustainable resources to meet growing electricity needs in the next 20–30 years. These renewable technologies, however, are quite intermittent due to the stochastic nature of wind speed and solar irradiation. Integrating renewable energy into existing power grids potentially Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 259 creates additional uncertainty on top of the load variability. When planning and operating DG systems, the uncertainties in power generation and load proile must be taken into account at the same time. Otherwise, the grid reliability and stability could be jeopardized, as more variable energy resources are penetrating into the electricity market. The OPEN system can play a substantial role in managing DG systems. Built upon the order-then-production principle, the OPEN system is capable of collecting and consolidating consumers’ future demands ahead of time. The lead time allows the utility supplier to perform demand predictions, plan optimal generation, and determine distribution schemes. Proactive measures can be adopted in advance to avoid the occurrence of power outage or interruption due to lower wind speed or cloudy weather. In this case, the OPEN system serves as an advance information system that allows the supplier to acquire future load conditions to mitigate the power uncertainty in renewable energy resources. 11.5.2 oPtImIzIng asset management The OPEN system creates a truly interactive environment in which the supplier and the consumers can engage in real-time communications. The system is quite beneicial in situations when a sudden change in demand occurs, such as a family leaving for a long vacation or a surge in energy demand due to manufacturing ramp-up. Consumers can easily send their reduced or increased energy demand information through the OPEN system. The OPEN system facilitates the deployment of plug-in hybrid electric vehicles (PHEVs) or other electrical apparatus across the grid system. Charging a PHEV means pumping 20–40 kWh electricity from the grid within 2–4 h, causing a signiicant impact on the stability of the distribution grid [28]. Functions can be built into the OPEN system that allow PHEV users to notify utility suppliers when and where the charging will be performed, and how much energy will be needed. This allows the supplier to take the necessary measures to mitigate the impact of charging on the local distribution network. The OPEN system can assist utility companies in optimizing the asset maintenance time based on consumers’ demand information. Maintenance activities on transformers, power lines, and other utility facilities can be carried out at times when the consumers’ demands are relatively low, based on their placed orders. If an urgent repair must be performed in a timely manner, the supplier can notify the affected consumers about the pending repair actions by broadcasting the message through the OPEN system. Consumers can then temporarily stop or reduce their energy usage to streamline the maintenance activity. 11.5.3 enhancIng load ForecastIng Utility companies are able to perform more realistic load forecasting, as they can combine historical data with advance demand orders. Bayes’ theorem is capable of synthesizing historical usage with advance orders to estimate the future load. Although the theorem has been widely used in various domains, its application in 260 Smart Grids the power industry for predicting future consumption is relatively new. Some studies [29,30] have reported construction of the prediction model using Bayes’ theorem. These works primarily focus on the utilization of Bayes’ theorem to obtain a better model structure. In the OPEN system, the Bayesian forecasting model is adopted to predict future electricity demand without making any assumptions about the model structure. The probability function representing historical usage is referred to as the prior distribution, and the probability function representing future consumption is called the posterior distribution. As shown in Figure 11.4, Bayes’ theorem allows us to directly estimate the posterior distribution by synthesizing the prior information with the demand orders. Two Bayesian prediction methods are available: discrete Bayesian and continuous Bayesian models. In the former, the prior and the posterior distributions are represented by discrete probabilities. In the latter, continuous probabilities are used to model the prior or the posterior distribution or both. 11.5.3.1 Discrete Bayesian Prediction The discrete Bayesian model can be applied in situations in which the demand data are characterized by attributes, or categorical data. The mathematical expression for the model is given in Equation 11.1: P { Xi | Y } = P { X jY } P {Y } = P {Y | Xi } P { Xi } ∑ k j =1 (11.1) P {Y | X j } P { X j } where: k is the total number of events or categories P{Xi} is the prior probability for Xi with i = 1, 2, …, k P{Y|Xj} is the conditional probability for consuming Y given Xj P{Xi|Y} is the posterior probability for Xi given Y The following example, taken from [27], is used to explain the application of this model. Table 11.1 lists the daily energy consumption of a consumer over 15 days. The actual consumption of energy, denoted as X, is listed in the second column, while the forecast or ordered quantity, Y, provided by the consumer is shown in the last column. For example, on Day 10, the consumer’s original order is 13 kWh, while the actual consumption is 14 kWh on that day. Prior distribution Customer demand orders Historical consumption FIGURE 11.4 The Bayesian forecast process. Posterior distribution Future consumption Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 261 TABLE 11.1 Daily Energy Usage versus Prediction Actual Usage (X) (kWh) Consumer Order (Y) (kWh) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 10 11 13 13 12 11 10 14 14 14 12 14 12 12 11 12 12 14 12 10 12 14 12 13 13 13 11 12 15 11 11 Day TABLE 11.2 Conditional Probability of P{Y|X} Y kWh 10 11 X 12 13 14 10 0 0.333 0 0 0 11 0.5 0.333 0.25 0 0 12 0.5 0.333 0.5 0.5 13 0 0 0.25 0 0.25 0.5 14 0 0 0 0.5 0.25 From Table 11.1, we can easily compute the prior probability for X, and the result is given as P{X = 10, 11, 12, 13, or 14} = {0.13, 0.20, 0.27, 0.13, or 0.27}, respectively. Using Equation 11.1, we are able to estimate the conditional probability for Y given X by using the data in Table 11.1, and the results are summarized in Table 11.2. For example, P{Y = 13|X = 14} = 0.5, as shown in the last row. Now the posterior probability for X can be extrapolated based on Table 11.2 along with the prior probability of X. The result is summarized in Table 11.3. For instance, the last row of the table shows P{X = 14|Y = 13} = 0.67, meaning that the probability of consuming 14 kWh electricity is 0.67, given that the consumer’s order is 13 kWh. Given the consumer’s order, we can, further, compute the conditional mean demand and the variance based on the conditional probability in Table 11.3. For 262 Smart Grids TABLE 11.3 Posterior Probability of P{X|Y} Y kWh 10 11 X 12 13 14 10 0 1.0 0 0 0 11 0.33 0.33 0.33 0 0 12 0.17 0.17 0.33 0.17 0.17 13 0 0 0.33 0 0.67 14 0 0 0 0.5 0.5 instance, if the consumer’s order quantity is Y = 13 kWh, then the conditional mean demand and its variance can be estimated as 5 E X Y = 13 = ∑ xP {X = x Y = 13} = 13.33 kWh (11.2) i =1 ( ( ) var X Y = 13 = E 2 X Y = 13 − E X Y = 13 ) 2 = 0.89 ( kWh) 2 (11.3) This numerical example, though relatively simple, clearly demonstrates how the discrete Bayesian inference model can be utilized to predict consumers’ actual consumption by synthesizing the historical data with their demand perspectives. 11.5.3.2 Continuous Bayesian Prediction If the actual consumption data are continuous and the demand order is placed in a discrete format, forecasting can be performed by using the Bayesian model for density functions: ( ) fX x Y = y = { } P Y X = x fX ( x ) P {Y } = { } P {Y X = x} f ( x ) dx P Y X = x fX ( x ) ∫ +∞ 0 (11.4) X When both X and Y are represented by continuous density functions, the Bayesian prediction model can be expressed by ( ) f X |Y x y = ( ) f X ,Y ( x, y ) fY | X y x f X ( x ) = fY ( y ) fY ( y ) where: X is the actual consumption, a continuous random variable Y is the consumer demand order, a discrete random variable P{Y} is the probability for Y f X(x) is the prior distribution for x f X(x|Y = y) is the probability density function for x given y P{Y|X = x} is the conditional probability for Y given x (11.5) 263 Integrating Consumer Demand Data in Smart Grid Energy Supply Chain f Y (y) is the marginal distribution for y f X,Y (x,y) is the joint distribution for x and y f Y|X(y|x) is the conditional density function for y given x f X|Y (x|y) is the posterior density function for x given y The likelihood function f Y|X(y|x) can be derived from consumers’ previous orders along with their actual consumption. All other notations are the same as in Equation 11.4. 11.5.3.3 Bayesian Prediction with Normal Distributions Normal distributions are often used to represent random electricity loads [31]. If X and Y follow the normal distributions with X ∼ N µ x , σ2x , and Y ∼ N µ y , σ2y , their joint density function is given as ( f X ,Y ( x, y ) = 1 2πσ x σ y 1 exp − 2 2 1− ρ 2 1 − ρ ( ) ) ( ) ( x − µ x )2 ( y − µ y )2 2ρxy + − σ2x σ2y σ xσ y (11.6) where ρ is the correlation coeficient with −1 ≤ ρ ≤ 1. By substituting Equation 11.6 into Equation 11.5, the posterior distribution for X is given as f X |Y ( x | y ) = 1 ( σ x 2 π 1 − ρ2 ) 1 exp − 2 2 1 − ρ ( ) ( x − µ x )2 ρ2 ( y − µ y )2 2ρxy + − σ2x σ2y σ xσ y (11.7) In summary, Bayes’ theorem is able to combine historical usage with the demand order to predict the future load. The discrete Bayesian model is preferred if the consumer usage and order are speciied in discrete format. The continuous distributions are more appropriate in applications in which the consumption and demand data are expressed as continuous variables. 11.6 CONSUMER PERFORMANCE ASSESSMENT Incentive programs are implemented to encourage consumers to participate in the OPEN system and place demand orders via their online accounts. Prior to constructing incentive programs, it is important to deine the performance metrics in order to assess the quality of orders. One reasonable approach is to compare the difference between the order and the actual consumption. Then various inancial incentive programs can be implemented, based on whether the actual usage matches the original order. The following set of performance measures to assess the quality of orders is based on [32]. 264 Smart Grids 11.6.1 mInImIzIng mean square error Minimum mean square error (MMSE) is often used to measure the difference between the prediction and the actual result. Let xi be the actual consumption and yi be the demand order for time ti; MMSE can be computed as MMSE = 1 m m ∑( x − y ) i 2 (11.8) i i =1 where m is the sample size. A small MMSE indicates that actual consumption closely matches the demand order. Obviously, energy credits or monetary discounts should be used to reward the consumer who maintains a small MMSE over the consumption horizon. MMSE is a simple statistical tool, which can be easily implemented in the OPEN system. However, MMSE cannot effectively capture the discrepancy between xi and yi for a particular time. In other words, consumers with the same MMSE may have a very different usage proile. Table 11.4 shows a list of 7-day energy orders versus the actual usages collected from three consumers. All consumers have the same MMSE = 4. The variation for consumer C on Day 7 is more pronounced than those of A and B. In addition, the actual computation for A is consistently smaller than the order, while the actual consumption for B is consistently larger than the order. Therefore, additional metrics are required to capture the consumer performance variations. 11.6.2 comPosIte PerFormance assessment metrIcs Three major factors need to be considered when we perform the assessment on the quality of the demand order: (1) the mean error between the order and the actual TABLE 11.4 Demand Order versus Actual Consumption Day Consumer A (kWh) i xi yi 1 2 3 4 5 6 7 18 20 17 20 21 23 16 20 22 19 22 23 25 18 MMSE (xi − yi)2 4 4 4 4 4 4 4 4 Consumer B (kWh) xi yi 12 14 11 9 11 12 13 10 12 9 7 9 10 11 (xi − yi)2 4 4 4 4 4 4 4 4 Consumer C (kWh) xi yi (xi − yi)2 32 31 29 34 32 29 38 32 31 29 33 31 28 33 0 0 0 1 1 1 25 4 Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 265 consumption; (2) the variation between the order and the consumption; and (3) the largest deviation across the entire period. Based on these criteria, the following composite metrics are proposed: q1 = q2 = { 1 m 1 m m ∑( x − y ) i (11.9) i i m ∑( x − y ) i 2 (11.10) i i =1 q3 = max ( xi − yi ) 2 } for i = 1, 2, … , m (11.11) These metrics form a set of composite assessment instruments to address the concerns raised previously. q1 calculates the mean error between the order data and the actual usage. The result is able to indicate whether the order is always below or above the actual usage. q2 measures the deviation between the order and the actual usage. q3 identiies the largest discrepancy between xi and yi for all i. All these performance metrics have a simple analytical form, and together they are able to assess consumer performance by simply comparing the advance demand order and actual usage data. 11.6.3 IllustratIve examPle The following example, taken from [32], is used to demonstrate the application of the performance assessment instrument. Three sets of 15-day orders from three consumers have been collected and listed in Table 11.5. The actual daily consumption is also recorded by the smart meters, and the results are also listed in the table. Table 11.6 summarizes the performance metrics computed using Equations 11.9 through 11.11. According to the mean error q1, Consumer A overestimates the demand while B and C underestimate the needed electricity. In terms of variation, Consumer B has the smallest MMSE, meaning that the actual usage for B is close to what she ordered. Consumer A has the largest MMSE, indicating a high daily variation between the order and the actual usage. Although the MMSE for C is not the largest, the largest daily deviation occurred for Consumer C. By examining the dataset for C, we found that the difference between the order and actual usage is 7 kWh on Day 15, resulting in the largest daily discrepancy among all the consumers. These examples show that, based on the composite assessment instruments, the utility supplier is able to determine consumer performance in a fairly easy way. Since the target value for q1, q2, and q3 is zero, the consumer performance can be easily evaluated in terms of the proximity to the target value. Based on this, various incentive packages can be devised to reward consumers who achieve good performance. In [22], a quadratic price model combining the reward function is proposed to incentivize the consumer to participate in the OPEN system. The model allows a 266 Smart Grids TABLE 11.5 Demand Order versus Actual Consumption Consumer A (kWh) Order (yi) Consumer C (kWh) Actual (xi) Order (yi) Actual (xi) Order (yi) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 24.5 22.1 26.9 24.8 27.3 28.9 32.6 30.2 38.5 29.3 31.0 27.1 32.6 29.5 28.6 25.7 24.3 27.9 32.0 32.0 33.8 25.9 34.9 33.0 29.1 32.7 29.8 31.3 25.2 30.3 22.5 26.7 18.7 20.3 24.3 18.8 24.0 31.3 24.6 24.6 19.7 26.5 25.7 27.3 21.4 25.9 18.3 19.1 25.3 17.4 24.1 29.4 26.4 25.0 18.0 27.0 38.4 41.7 36.4 39.6 33.9 39.8 41.4 43.1 43.8 38.2 37.8 40.2 40.2 39.4 15 26.5 26.1 26.0 26.5 50.0 39.2 40.6 37.5 40.3 35.1 40.6 40.8 41.7 41.5 37.5 40.1 40.4 40.2 37.8 42.6 Day (i) Actual (xi) Consumer B (kWh) TABLE 11.6 Performance Assessment for Three Consumers Metrics q1 q2 q3 A B C Target Value −1.0 11.4 32.0 0.5 1.7 9.3 0.5 5.0 54.8 0 0 0 consumer to save up to 30% of the monthly bill if all performance metrics meet the targets. Notice that other options are available, and they can be built into the reward functions. Interested readers are referred to [33,34]. 11.7 COMMUNICATIONS AND AUTOMATION TECHNOLOGY To help utility companies and consumers to generate a power usage proile, the actual power consumption has to be measured and available to both the utility company and the consumers. There are many advantages if the real-time power consumption is available: • The transparency of power usage between utility companies and consumers is increased. Hence, the utility companies have a better understanding of consumers’ demand. Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 267 • The utility companies can schedule their power supply according to consumers’ demand. Thus, the power generation cost can be reduced and pollution can be decreased. • Consumers will be aware of their real-time power usage rather than obtaining a monthly utility bill, or at even longer intervals. Thus, their power consumption can be altered and controlled. • If dynamic pricing is available, consumers can schedule their power consumption according to the price proile. For example, they can schedule their washing machine when the price is lowest if they are not in a hurry to wash their clothes. To measure the actual power consumption and make it available to both utility companies and consumers, smart meters and data communication networks have to be implemented. Since the Ethernet is widely used at both consumer sites and utility companies, using the existing communication infrastructure reduces the infrastructure cost and eliminates many implementation-related issues. Figure 11.5 shows a power consumption data collection and distribution scheme. In Figure 11.5, the local smart sensors form a local network. The local servers are installed at the consumer sites. The utility servers are installed at the utility companies. As computers are becoming more powerful and have many different built-in interfaces, such as wireless, Ethernet, USB, and serial ports, it is convenient to connect the smart sensors with different communication interfaces to the computers at the consumer sites. Therefore, there are many possible ways for the local smart sensor network to communicate with the local server. The local servers exchange the data with the utility servers via Ethernet. The main reason for isolating the smart meters with their network from the local servers is security considerations. As the Smart sensor Smart sensor Smart sensor Smart sensor Smart sensor Smart sensor Local server Local server Local server Utility server FIGURE 11.5 Utility server Utility server A sensor network interconnecting utility companies with consumers. 268 Smart Grids smart sensor network becomes more complex, network security will become a major issue in the near future. Since the local servers and the utility servers in the utility companies are connected via Ethernet, they can communicate with each other without sacriicing the security of smart meters. As the amount of data collected increases with time, servers require vast storage capacity, especially in the utility companies, because there are many smart meters connected via the local servers. There are several problems with storage of vast amounts of data: (a) extra hardware is needed to save the data; (b) data mining is becoming more dificult; and (c) the processing time is becoming longer. To deal with these problems, distributed data storage and processing techniques are being developed to reduce the amount of data transmitted to the utility servers. Under the distributed processing framework, data from each smart meter are collected and saved in the smart meter, if it has a storage device, or in the local server. Since the storage capacity of a smart meter is typically limited, the data should be transferred to the local server for storage. As the local server is only connected to one or several local smart meters, the storage capacity of a local server should be large enough to store the data from the local smart meters. After the data are transferred to the local server, they are processed to reduce the data size. Here we propose a B-spline curve-itting method. Suppose we have m power consumption data points Pi (i = 1, 2, …, m) during a cycle (e.g., 1 h or a day), a B-spline curve can be itted to the m points: S (t ) = m−n ∑C b i i,n ( t ), t ∈ t n , t m − n (11.12) i =1 where: t S(t) Ci bi,n(t) is the sampling time during a cycle is the B-spline curve are the control points is the B-spline basis function, which is deined as if t j ≤ t ≤ t j +1 otherwise 1 b j ,0 ( t ) = 0 b j ,n ( t ) = t − tj t −t b j +1,n −1 ( t ) b j ,n −1 ( t ) + j + n +1 t j+n − t j t j + n +1 − t j +1 (11.13) For example, if a uniform cubic B-spline is used, the function can be represented as Sk ( t ) = 1 6 t 3 t2 t −1 3 1 −3 1 3 −6 0 4 −3 3 3 1 1 Ck −1 0 Ck 0 Ck +1 0 Ck + 2 (11.14) Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 269 Therefore, instead of transmitting m data points, only four points are needed to represent the power consumption during a cycle if a common basis function is used. This will greatly reduce the data storage requirement and decrease the data processing time. For the local server, the control points have to be computed. Here we developed a least square method to obtain the optimal control points: f = min ∑ || S(t ) − P || i i (11.15) i After the objective function f is minimized, the control points are obtained. Since this optimization is performed in the local servers, this will greatly mitigate the data congestion issue. 11.8 SYSTEM PRELIMINARY DESIGN Based on the concept proposed in Section 11.3, we have started system implementation in the laboratory setting. This section will summarize our preliminary designs and development of the user-friendly interface that is an integral part of the OPEN system. More detailed discussions are given in the recent work in [27]. In our VEP system, the user interface behaves like a bridge between users and service provider. Similarly to an online shopping web page, users can easily browse their historical usage data and input new usage proiles and other related consumption information, such as number of persons living in a house, appliance information, and daily activities. Hence, designing an interactive, functional, and user-friendly interface is very important. Figure 11.6 shows a web-based interface allowing consumers to specify the power capacity of major household appliances. Our designed interface consists of four components: (1) User information. When the new user logs into the system, he or she needs to set up an account (user name and password) and input personal information, including the number of persons in the house and major appliances, for example, air conditioner model. This information will be stored in the database, and will be very useful for the OPEN system in predicting the future electricity load. The above example is taken from residential homes, but the same principle can be applied to industrial and business entities. (2) Visualization. In order for consumers to plan well, it is necessary for them to visualize past electricity usage (e.g., hourly, daily, weekly, or monthly), usage patterns of the same period in previous years, and usage patterns of similar households. Figures 11.7 and 11.8 graphically show the daily consumption proile by apparatus type and the cumulative usage, respectively. (3) Search. The major function of the “Search” module is to help users ind the electricity consumption information for their appliances, such as air conditioners or refrigerators. For example, consumers can input their refrigerator model and the search module will ind the average electricity consumption of this type of refrigerator. This information will be very helpful for them to make an accurate prediction. (4) Feedback. Through the feedback component, the system will provide the user with a predicted usage amount. Consumers can also input their estimated amount and give feedback to the system. For example, based on previous usage information, the system will automatically generate an 270 FIGURE 11.6 Smart Grids Interface for specifying power capacity of apparatus. Watts 800 700 600 500 400 300 [Total] Sum of consum_AC Sum of consum_TV Sum of consum_music Sum of consum_oven 200 100 Days 1 FIGURE 11.7 Consumer online account. 2 3 4 5 6 271 Integrating Consumer Demand Data in Smart Grid Energy Supply Chain Watts 600 500 400 300 200 100 Sum of max_consum Sum of min_consum Sum of actual_consum 1 2 3 4 5 6 Day FIGURE 11.8 Consumer’s actual usage proile. estimated demand. It is up to the consumer to modify the recommended usage and place the order through the OPEN system. If the consumer will be out of town next month, for instance, he or she can input a very small amount of electricity usage. This advance demand information, once appropriately analyzed by the data mining engine at the utility side, can be used as the baseline to plan future generation. 11.9 CONCLUSION This chapter presented a novel DSM technology, VEP, which allows smart grid consumers to predict and order electricity via the broadband Internet as if performing an online purchasing transaction. This new energy management concept intends to shift the electricity delivery service from production-then-consumption mode to an order-then-production paradigm. The system architecture, operational conditions, and potential beneits to energy stakeholders are elaborated from various aspects, including generation, marketing, delivering, and end consumption. We also investigated several statistical instruments to assess the quality of demand orders. The purpose is to evaluate consumers’ performance and further develop incentive programs to increase their participation. The preliminary research shows that the OPEN system is technically feasible, theoretically justiiable, and inancially achievable. The ultimate goal of the VEP concept is to make the consumers “order the electricity they need, and consume exactly what they ordered.” As a result, the stochastic load proile is turned into a deterministic demand curve, reducing operational uncertainty and improving system robustness. Many research opportunities will be opened up 272 Smart Grids by implementing this new DSM concept. These areas include, but are not limited to, dynamic pricing, optimal generation and distribution planning, and consumer incentive management. Future efforts will be focused on software development and laboratory testing, and will eventually move to real-world applications. REFERENCES 1. Brown, R.E., Impact of smart grid on distribution system design, in Proceedings of the IEEE Power and Energy Society General Meeting, pp. 1–4, July 20–24, 2008, Pittsburgh, PA. 2. Venayagamoorthy, G.K., Potentials and promises of computational intelligence for smart grids, in Proceedings of the IEEE Power and Energy Society General Meeting, pp. 1–6, July 26–30, 2009, Calgary, AB. 3. Chen, W.-H., C.-W. Liu, and M.-S. Tsai, On-line fault diagnosis of distribution substations using hybrid cause-effect network and fuzzy rule-based method, IEEE Transactions on Power Delivery, 15, 710–717, 2000. 4. Xu, L. and M.-Y. Chow, A classiication approach for power distribution systems fault cause identiication, IEEE Transactions on Power Systems, 21, 53–60, 2006. 5. Cai, Y. and M.-Y. Chow, Exploratory analysis of massive data for distribution fault diagnosis in smart grids, in Proceedings of the IEEE Power and Energy Society General Meeting, pp. 1–6, July 26–30, 2009, Calgary, AB. 6. US DOE, The 20% wind energy by 2030: Increasing wind energy’s contribution to the U.S. electricity supply, US Department of Energy Report, p. 7, 2008. 7. Jin, T. and J. Jimenez, A review on planning and automation for distributed generation systems, in Proceedings of Conference on Automation Science and Engineering, pp. 269–274, August 21–24, 2010, Toronto, ON. 8. Tor, O. and M. Shahidehpour, Power distribution asset management, in Proceedings of the IEEE Power Engineering Society General Meeting, pp. 1–7, June 18–22, 2006, Montreal, QC. 9. Garrity, T.F., Innovation and trends for future electric power systems, in Proceedings of the Power Systems Conference (PSC ‘09), pp. 1–8, March 10–13, 2009, Clemson, SC. 10. Gao, Z., J. McCalley, and W. Meeker, A transformer health assessment ranking method: Use of model based scoring expert system, in Proceedings of North American Power Symposium (NAPS), pp. 1–6, October 4–6, 2009, Starkville, MS. 11. Nilsson, J. and L. Bertling, Maintenance management of wind power systems using condition monitoring systems—Life cycle cost analysis for two case studies, IEEE Transactions on Energy Conversion, 22(1), 223–229, 2007. 12. Tian, Z., T. Jin, B. Wu, and F. Li, Condition-based maintenance optimization for wind power generation systems under continuous monitoring, Renewable Energy, 36, 1502– 1509, 2011. 13. Ostertag, M. and Ch. Imboden, High data rate, medium voltage powerline communications for hybrid DA/DSM, in Proceedings of the IEEE Transmission and Distribution Conference, vol. 1, pp. 240–245, April 11–16, 1999, New Orleans, LA. 14. Heidell, J. and H. Ware, Is there a case for broadband utility communications networks? Valuing and pricing incremental communications capacity on electric utility smart grid networks, The Electricity Journal, 23(1), 21–33, 2010. 15. Molderink, A., V. Bakker, M.G.C. Bosman, J.L. Hurink, and G.J.M. Smit, A three-step methodology to improve domestic energy eficiency, in Proceedings of the Innovative Smart Grid Technologies (ISGT), pp. 1–8, January 19–21, 2010, Gaithersburg, MD. 16. Babu, C.A. and S. Ashok, Peak load management in electrolytic process industries, IEEE Transactions on Power Systems, 23(2), 399–405, 2008. Integrating Consumer Demand Data in Smart Grid Energy Supply Chain 273 17. Ramanathan, B. and V. Vittal, A framework for evaluation of advanced direct load control with minimum disruption, IEEE Transactions on Power Systems, 23(4), 1681–1688, 2008. 18. Moholkar, A., P. Klinkhachorn, and A. Feliachi, Effects of dynamic pricing on residential electricity bill, Power Systems Conference and Exposition, 2, 1030–1035, 2004. 19. Sui, H., H. Wang, M.-S. Lu, and W.-J. Lee, An AMI system for the deregulated electricity markets, IEEE Transactions on Industry Applications, 45(6), 2104–2108, 2009. 20. Zhong, J., C. Kang, and K. Liu, Demand side management in China, in Proceedings of the IEEE Power and Energy Society General Meeting, pp. 1–4, July 25–29, 2010, Minneapolis, MN. 21. Zeng, S., J. Li, and Y. Ren, Research of time-of-use electricity pricing models in China: A survey, in Proceedings of the IEEE International Conference on Industrial Engineering and Engineering Management, pp. 2191–2195, December 8–11, 2008, Singapore. 22. Jin, T. and M. Mechehoul, Ordering electricity via Internet and its potentials for smart grid systems, IEEE Transactions on Smart Grid, 1(3), 302–310, 2010. 23. Kusko, A. and J. DeDad, Stored energy—Short-term and long-term energy storage methods, IEEE Industry Applications Magazine, 13(4), 66–72, 2007. 24. Fahrioglu, M. and F.L. Alvardo, Designing incentive compatible contracts for effective demand managements, IEEE Transactions on Power Systems, 15(4), 1255–1260, 2000. 25. Tsoukalas, L.H. and R. Gao, From smart grids to an energy Internet: Assumptions, architectures and requirements, in Proceedings of the Third International Conference on Electric Utility Deregulation and Restructuring and Power Technologies, pp. 94–98, April 6–9, 2008, Nanjing. 26. Guan, X., Y.-C. Ho, and D.L. Pepyne, Gaming and price spikes in electric power markets, IEEE Transactions on Power Systems, 16(3), 402–408, 2001. 27. Jin, T., L. Lu, and C. Kang, Energy management system for smart grid consumers using advance demand information, in Proceedings of the IASTED European Conference on Power and Energy Systems, pp. 300–306, June 22–24, 2011, Greece. 28. Clement-Nyns, K., E. Haesen, and J. Driesen, The impact of charging plug-in hybrid electric vehicles on a residential distribution grid, IEEE Transactions on Power Systems, 25(1), 371–380, 2010. 29. Barghinia, S., S. Kamankesh, N. Mahdavi, A.H. Vahabie, and A.A. Gorji, A combination method for short term load forecasting used in Iran electricity market by NeuroFuzzy, Bayesian and inding similar days methods, in Proceedings of the 5th International Conference on European Electricity Market, pp. 1–6, May 28–30, 2008, Lisbon. 30. Niu, L., J. Zhao, and M. Liu, Application of relevance vector regression model based on sparse Bayesian learning to long-term electricity demand forecasting, in Proceedings of the International Conference on Mechatronics and Automation, pp. 2363–2367, August 9–12, 2009, Changchun. 31. Couchman, P., B. Kouvaritakis, M. Cannon, and F. Prashad, Gaming strategy for electric power with random demand, IEEE Transactions on Power Systems, 20(3), 1283–1292, 2005. 32. Jin, T. and Y. Yu, Online virtual electricity provisioning system and its preliminary design, Automation of Electric Power Systems, 35(4), 35–40, 2011. 33. Ferrero, R.W. and S.M. Shahidehpour, Short-term power purchases considering uncertain prices, IEE Proceedings—Generation, Transmission and Distribution, 144(5), 423–428, 1997. 34. Mohsenian-Rad, A.H., V. Wong, J. Jatskevich, and R. Schober, Optimal and autonomous incentive-based energy consumption scheduling algorithm for smart grid, in Proceedings of the Innovative Smart Grid Technologies (ISGT), pp. 1–6, January 19–21, 2010, Gaithersburg, MD. 12 Photovoltaic Energy Generation and Control for an Autonomous Shunt Active Power Filter Ayman Blorfan, Damien Flieller, Patrice Wira, Guy Sturtzer, and Jean Mercklé CONTENTS 12.1 Introduction .................................................................................................. 275 12.2 Consequences of Harmonic Pollution .......................................................... 277 12.3 Principle of Sliding Mode Control ...............................................................280 12.3.1 Choice of Sliding Surface ................................................................. 282 12.3.2 Determination of the Reference Current .......................................... 285 12.4 Characteristics of the System ....................................................................... 286 12.4.1 The Electrical Equations .................................................................. 286 12.4.2 Projection in the Rotation Reference ................................................ 288 12.4.3 Control Vectors ................................................................................. 288 12.4.4 Description of the Control Laws....................................................... 289 12.4.5 Sliding Mode Existence Conditions ................................................. 290 12.4.6 Choice of Vector Field ...................................................................... 294 12.5 The Photovoltaic Source ............................................................................... 294 12.5.1 Solar Cell Model Using Orcad.......................................................... 296 12.5.2 Maximum Power Point Tracking ...................................................... 297 12.5.3 Direct Current Voltage Optimization with the Photovoltaic Source .... 301 12.5.4 Control of the Switching Frequency .................................................302 12.6 Simulation Results ........................................................................................304 12.7 Conclusion .................................................................................................... 305 12.8 Nomenclature ................................................................................................308 References ..............................................................................................................309 12.1 INTRODUCTION The past several decades have seen a rapid increase of power electronics-based loads connected to the utility system [1]. However, the proliferation of these nonlinear loads has raised concern with regard to the effect of the resulting harmonic distortion 275 276 Smart Grids levels of the supply current on the power system [2]. Passive iltering is the traditional method of harmonic iltering. However, it is well known that the application of passive ilters creates new system resonances that are dependent on the speciic system conditions. Although this solution is simple, it gives rise to several shortcomings. Furthermore, since the harmonics to be eliminated are of low order, large ilter components are required [3]. The active power ilter (APF) is an eficient solution to solve the problems created by nonlinear loads; the concept is shown in Figure 12.1. We will use it in this work. Hybrid APFs (HAPFs) have been developed in which a passive ilter is connected parallel to a conventional APF [4]. The HAPF coniguration is effective in improving the damping performance of high-order harmonics. Our system consists of the following ive parts: 1. Source: alternating current (ac) line sinusoidal, balanced or not 2. Photovoltaic (PV) array: a nine-module (85 W each) PV array with 1000 W/m2 insolation and a boost direct current-direct current (dc-dc) converter 3. Nonlinear load: a bridge diode rectiier feeding a resistive inductive load 4. Three-phase inverter 5. Resistor/inductor/capacitor (RLC) passive power ilter Recently, there has been increasing concern about environmental pollution. The need to generate pollution-free energy has triggered intensive efforts to ind alternative sources of energy. Solar energy, in particular, is a promising option. Some researchers have worked on developing a combined APF and PV system [5,6]. However, existing HAPF conigurations, with all their beneits, are not yet implemented for PV applications. iS iL Source Active power filter Photovoltaic array FIGURE 12.1 Circuit diagram of a shunt APF connected to a nonlinear load. Nonlinear load iinf Photovoltaic Energy Generation and Control for Active Power Filter 12.2 277 CONSEQUENCES OF HARMONIC POLLUTION Many of the effects of harmonics on electrical equipment and installations can be cited [5]. The greatest effects are overheating and interference with networks. • Heating: Total losses due to the joule effect are the sum of the losses due to the fundamentals and harmonics; it is deined by Equation 12.1: ∞ PLosses = ∑I R 2 h (12.1) h =1 • • • • where Ih is the harmonic current, which represents the fundamental range for h = 1, and R is the resistance traversed by the current. Harmonics also increase iron and copper losses. They become more important in devices based on magnetic circuits (motors, transformers, etc.). In power electronic equipment, these harmonics can shift the voltage zero crossing and lead to malfunctioning of medical instruments and other electronic devices. Furthermore, losses and malfunctions caused by the presence of harmonic currents markedly reduce the performance of equipment such as motors, transformers, and so on. Aging of insulation: Insulation aging often results from tensile stress due to the presence of harmonics and thus a local increase in leakage current, or to excessive heating in the conductors. The most spectacular example of this type of effect is the destruction of equipment (condenser, breaker, etc.). Interference with telecommunication systems: The electromagnetic coupling between electric and telecommunication networks can cause interferences. The importance of the interference depends on the amplitude and the frequency of electric currents and the importance of electromagnetic coupling between the networks. For example, parts of the telecommunication networks may be rendered unusable in the case of resonance. The unintentional generation, propagation, and reception of the electromagnetic energy and the effect of these harmonics on the equipment items or systems must be analyzed and characterized. This is referred to as electromagnetic compatibility. Malfunctioning of electrical equipment: In the presence of harmonics, the voltage, the current, or both can change sign several times in a half period; therefore, sensitive equipment at the zero crossing of these electrical parameters is perturbed. Risk of resonance excitation: The instantaneous effects of resonance are due to capacities or inductances that can have resonant frequencies close to harmonic frequencies. This is the case when the batteries are connected to the network capacity to meet the power factor; resonance frequencies can be quite low, and thus coincide with those of the harmonics generated by static converters. In this case, there will be an ampliication of harmonic phenomena; surges occur and damage the cables, transformers, protection systems, battery capacity, and so on. 278 Smart Grids Various quantities are deined to quantify this disturbance: • The total harmonic distortion (THD) rate deined by: THDh = Ih I1 (12.2) where Ih is the h-order harmonic component and I1 represents the fundamental component • The overall harmonic distortion: we characterize the pollution of a whole network by the THD in the voltage or current as THD ( % ) = 100 ∞ ∑ h=2 I h2 I12 (12.3) Generally, the harmonics included in electrical networks are lower than 2500 Hz. They correspond to harmonic terms from rank 2 to 50 and are considered low-frequency disturbances. The higher-frequency harmonics are greatly attenuated by the skin effect and the presence of the inductances of lines. In addition, the devices that generate harmonics mostly have an emission spectrum below 2500 Hz; this is why the harmonics study ield generally extends from 100 to 2500 Hz, that is, from rank 2 to 50 [1]. • The power factor: For a signal wave, the power factor is given by the ratio between the active power and the apparent power S. Generators, transformers, transmission lines, and control devices inluence the measurement of voltage and current ratings. A low value of the power factor often results from poor use of these facilities. In the case of harmonics, an additional power called the deformans power is deined mostly when the power is produced by a nonsinusoidal source. The active power in periodic nonsinusoidal signals is deined by P= ∑ P = ∑U I cos φ n n n n n (12.4) n where Un and In are the root mean square values of the voltage and current harmonic of order n, and ϕn is the phase shift between them. The reactive power can then be deined by Q= ∑ Q = ∑U I sin φ n n n n n (12.5) n However, these deinitions do not conform to the equation of the power triangle: S = P 2 + Q2 (12.6) 279 Photovoltaic Energy Generation and Control for Active Power Filter The apparent power is deined by S 2 = U 2I 2 = ∑U ∑ I 2 n n 2 n (12.7) n Then S = 2 + ∑U ∑ I ≥ ( ∑U I cos φ ) ∑ 2 n n 2 2 n n n n n n U n I n sin φn 2 (12.8) The quantity D is added using the following equation: D2 = S 2 − P 2 − Q2 (12.9) We can write the apparent power as follows: (12.10) S 2 = P 2 + Q2 + D2 The Fresnel diagram of the powers is shown in Figure 12.2. Distorting power consists mainly of the cross-product of the voltage and current harmonics of different orders and will be reduced to zero if the harmonics are reduced to zero, that is, under sinusoidal conditions: ∞ D= ∑U I 2 2 i j (12.11) i, j =2 i≠ j If only the current is nonsinusoidal, then: ∞ D = U1 ∑I 2 k (12.12) 2 S D Q γ φ φ1 FIGURE 12.2 Fresnel diagram of the power deformans. P 280 Smart Grids The major drawback of this deinition is that it is not certain that the power factor is equal to unity if the reactive power (by deinition) is reduced to zero and the reactive power can be fully compensated by inserting inductive or capacitive components. • The power factor (F.P.) becomes: F.P. = P 2 P + Q2 + D2 = cos ϕ ⋅ cos γ (12.13) It is clear that the harmonics affect the power factor. 12.3 PRINCIPLE OF SLIDING MODE CONTROL The sliding mode design approach is one of the effective nonlinear robust control laws. Sliding mode control has been widely applied to power converters due to its characteristics, such as speed, robustness, and stability to large load variations. It provides system dynamics with an invariance property to uncertainties once the system dynamics are controlled in the sliding mode. It consists of two components. The irst involves the design of a switching function so that the sliding motion satisies the design speciications. The second is concerned with the selection of a control law that will make the switching function attractive to the system state. It can be seen that this control law is not necessarily discontinuous. Let the nonlinear system be in the following form with a scalar control: Xɺ = f ( X , U ) (12.14) where X is the system error and U is the control issued by the corrector with variable structure. The second member of the equation is discontinuous due to the discontinuity imposed on the control side of the surface σ(x) = 0. Depending on the order applied, the system will take one of the following two forms, as shown in Figure 12.3: ( X = f − X ,U − Sliding surface σ= 0 ) if σ( x) < 0 σ>0 U = U+ σ<0 U = U− FIGURE 12.3 System behaviors on the sliding surface. (12.15) 281 Photovoltaic Energy Generation and Control for Active Power Filter ( X = f + X ,U + ) if σ( x) > 0 (12.16) A nonlinear system is given by the differential equation: Xɺ = f ( X , t ) + g ( X , t ) ⋅ U n m n (12.17) n× m where X ∈ ℜ, U ∈ ℜ f ( X , t ) ∈ ℜ g ( X , t ) ∈ ℜ . How should the variable structure for the command be applied? Two questions can be asked: 1. What is this control? 2. How can it be synthesized? The synthesis of a variable structure, represented by Figure 12.4, involves two phases [5]: • The irst is to deine the switching surface σ(x) = 0 of the desired dynamics. • The second aims to synthesize the control function u(x, t), in order to make sure that [6] any state outside the switching surface will be joined to the states inside the sliding tranche (with bandwidth 2δ surrounding the sliding surface; see Figure 12.5) during a inite time (limited by the switching frequency of the insulated gate bipolar transistor). Once it Umax Actuator Controlled system Umin Switching controller S (X) FIGURE 12.4 Schematic diagram of variable structure control. 2.δ σ>0 σ>0 0 0 σ<0 σ<0 Surface S σ=0 FIGURE 12.5 Trajectories of sliding mode, ideal and real. Surface S σ=0 282 Smart Grids reaches them, a sliding mode occurs. The system adopts the dynamic of the surface and the point representing the evolution of the system reaches the equilibrium point. The dynamic slip is independent of the control, which only serves the evolution of the intersection of surfaces and ensures their attractiveness. From any initial condition, the representative point of the movement meets the hypersurface slip in a inite time. 12.3.1 choIce oF slIdIng surFace Generally, the sliding surfaces are chosen as hyperplanes through the origin of space, and their control is shown in Figure 12.4. The general shape of these surfaces is given by the following equations: σ ( X ) = CX = 0 σ(X ) = (12.18) n ∑C X i (12.19) i i =1 Although the surface may be nonlinear, the linear surfaces are more practical for the design of a corrective variable structure sliding mode in the case of a multivariable system. We consider the following nonlinear system: Xɺ = f ( X , t ) + g ( X , t ) ⋅ U n m n (12.20) n ×m where X ∈ ℜ, U ∈ ℜ f ( X , t ) ∈ ℜ ( X , t ) ∈ ℜ . The problem is more complex due to the existence of several switching surfaces (σ1, σ2, σ3, …, σm). The sliding mode occurs on the intersection of these surfaces. The two switching surfaces are chosen as σ ( X ) = CX (12.21) σ ( X ) = σ1 ( X ) , σ2 ( X ) , … , σm ( X ) = 0 T C = c1, c2 , … , cn T (12.22) (12.23) The switching surface σ(x) = 0 is a surface of dimension m, and σi(x) is the ith switching surface. The intersection of the surface σ(X) is a sliding surface dimension (n−m). Each entry has the form Ui: U + with σi ( X ) > 0 U ( X, t ) = − i = 1, … , m U with σi ( X ) < 0 (12.24) Photovoltaic Energy Generation and Control for Active Power Filter 283 The switching surface is chosen such that, when the system is in sliding mode, the behavior is meeting the desired objectives. After selection of the sliding surface, an important phase remains to be considered. Can we ensure the existence of a sliding mode? If the tangents to the trajectories or velocity vectors near the surface are directed toward the switching surface σ(x) = 0, then a sliding mode occurs on this plane. Let us consider the canonical form: xɺi = xi +1 i = 1, … , n − 1 n ɺ = − x ai xi + f ( t ) + U n i =1 ∑ (12.25) where U is the command with discontinuities in the vicinity of σ(x) = 0. It should be noted that the disturbance f(t) and the parameters may be unknown. The switching surface is of the form: n σ( x) = ∑C X T i i (12.26) i =1 where the coeficients are constants. The following pair of inequalities constitutes the conditions to the existence of the sliding mode: lim σɺ > 0 and lim+ σɺ < 0 s →0− s →0 (12.27) To prove the invariability with respect to the coeficient ai and f(t), solve Equation 12.26 to equal 0, and substitute it into Equation 12.25. Once the system is in sliding mode, the closed-loop system becomes xɺi = xi +1 i = 1, … , n − 2 n −1 ɺ = − x ci xi 1 − n i =1 ∑ (12.28) The system no longer depends on f(t) or on the ai coeficients; on the contrary, it depends on the parameters ci, which gives the system an invariance to disturbances and parametric variations. Let the system be described by Equation 12.15, and a surface σ separate the hypersurface into two parts, G+ and G−: dX = F ( t , x1, …, xn ) dt (12.29) with X = [x1, …, xn]T; F = [f1, …, f n]T. The functions f1, …, f are piecewise continuous and have discontinuities on the hypersurface σ of equation σ(x) = 0. 284 Smart Grids f N+ and f N− are the f-projections on the normal orientation of σ. These projections are made of both sides of the surface; we can show that if, at any point on the surface σ, ∂fi/∂xj is bounded, f N+ − f N−( f N+ < 0 and f N− > 0) is negative and σσɺ < 0 . The sliding condition can take two equivalent forms, given by Equations 12.30 through 12.32: σσɺ < 0 (12.30) f N+ < 0 being σ > 0 and σɺ < 0 (12.31) f N− > 0 > 0 being σ < 0 and σɺ > 0 (12.32) ( f N+ < 0 and f N− > 0) ⇒ σσɺ < 0 (12.33) where In the ideal sliding mode case, the switching frequency is assumed to be ininite. In practice, the frequency is limited by the power electronic components. As a consequence, some hysteresis effects are introduced. They increase the phenomenon of “chattering” embodied by the oscillations around the sliding surface, but do not call into question the theory of sliding. Figure 12.5 shows the effect of hysteresis bandwidth 2δ. The balance point can be arbitrarily chosen, but an origin allows faster convergence to the sliding surface. Commutations move freely within the bandwidth around the surface. An oscillating phenomenon appears around the equilibrium point. These trajectories are shown in Figure 12.5. Consider the monovariable state system described by Equation 12.20; the equivalent control of the ideal sliding mode (neither threshold, nor delay, nor hysteresis) is obtained on σ(x, t) = 0 with σɺ = ( x, t ) = 0. T T ∂σ ∂σ dx ∂σ ∂σ σɺ = ( x, t ) = dt + dt = ∂x f ( x, t ) + g ( x, t ) U eq + dt x ∂ ∂σ T , U eq ( x t ) = − g ( x, t ) ∂x −1 ∂σ T ∂σ f ( x, t ) + ∂t ∂x (12.34) (12.35) The condition of the sliding mode is −1 ∂σ T g ( x, t ) ≠ 0 ∂x (12.36) Photovoltaic Energy Generation and Control for Active Power Filter 285 This condition is also called the transversality condition. In reality, the fact that the product is zero means that the vector ield g is orthogonal to the surface normal vector. So g has no vertical component and system status cannot be brought to the surface. This condition is necessary but not suficient to ensure the sliding mode. By replacing the expression of the equivalent control in the system of Equation 12.20, we obtain the following expression: −1 −1 ∂σ T ∂σ T ∂σ T ∂σ T xɺ = I − g ( x, t ) g ( x, t ) ∂x f ( x, t ) − g ( x, t ) ∂x g ( x, t ) ∂x ∂x (12.37) The equivalent control shown in Figure 12.6 maintains the system state on the sliding surface σ = 0. The average value that takes control is deined as follows { } { } min U − ( x, t ) , U + ( x, t ) 〈ueq 〈 max U − ( x, t ) , U + ( x, t ) (12.38) The necessary and suficient condition for the sliding mode’s existence is given by { } { } U min = min U − ( x, t ) , U + ( x, t ) (12.39) U max = max U − ( x, t ) , U + ( x, t ) 12.3.2 determInatIon oF the reFerence current In an APF application, the dynamic performance of the sliding mode control will also depend on the algorithm used for identifying the current harmonics [7]. The method adopted to extract the reference currents and the synthesis of different regulators (control loops of voltage and current) will ensure a sinusoidal current on the network. Different methods of identifying the harmonic (or the fundamental) current components of a nonlinear load can be used. In the following study, we will present and analyze, respectively, the p−q method of instantaneous powers and the direct method. Their use will depend on the targets: harmonic compensation, its Umax Umin FIGURE 12.6 Representation of the equivalent control. Ueq Voltage of the source 286 Smart Grids Vsα L.P.F Vs1 Vs2 T32t Vs3 p iref1 p Vsβ ~ p p=? irefα = ? irefα iref2 Current of the load T32 icα iL1 iL2 iL3 T32t L.P.F q=? icβ q irefβ = ? q irefβ iref3 ~ q FIGURE 12.7 Overall system coniguration and control block diagram. application to three-phase installations (under balanced or unbalanced conditions), or single-phase installations. The p−q technique is illustrated in Figure 12.7. In order to generate the compensation current that follows the reference current, a ixed-band hysteresis current control method is adopted. The proposed scheme is designed to inject the reactive and harmonic current of the nonlinear load and the reactive current of the passive ilter. Furthermore, a current must be drawn from the utility source to maintain the voltage across the dc-bus capacitor to a preset value that is higher than the amplitude of the source voltage. Under normal operation, the PV system will provide active power to the load and the utility. Under poor PV power generation conditions, the utility source supplies the active power to the load directly. 12.4 CHARACTERISTICS OF THE SYSTEM The power circuit of the proposed HAPF [7] consists of a PV system that is in parallel with a nonlinear load and supplied by a source voltage connected at the point of common coupling. The proposed HAPF is a passive high-pass ilter (HPF) associated with a shunt APF constructed by a three-phase full-bridge voltage source inverter connected to a dc-bus capacitor, and PV arrays in parallel with the dc-bus capacitor. In the proposed scheme, the low-order harmonics are compensated using the shunt APF, while the high-order harmonics are iltered with the passive HPF. This coniguration is effective to improve the iltering performance of the high-order harmonics. The inverter structure that is adopted is shown in Figure 12.8. 12.4.1 the electrIcal equatIons The voltages supplied by the transformer on the secondary side are assumed to be a sinusoidal three-phase balanced system with a constant frequency of 50 Hz. They are expressed by 2⋅π Vsi = Vm ⋅ cos ωt − ( i − 1) 3 with i = 1, 2, 3 and ω = 2·π·F. (12.40) 287 Photovoltaic Energy Generation and Control for Active Power Filter Rf Lf Rf Lf Rf Lf Vc FIGURE 12.8 Structure of the inverter. If we denote the output voltages of the inverter, the following model provides an initial pattern of the inverter’s behavior: dI ond1 L f dt + R fI ond1 = Vs1 − Vond1 dI ond2 + R fI ond2 = Vs2 − Vond2 L f dt dI ond3 L f dt + R f I ond3 = Vs3 − Vond3 C dVc = dt 3 ∑ (u I i ondi ) (12.41) (12.42) i =1 ui is deined as the switching function on the state of the switch on or off of the inverter. A switch is characterized by the value 1 if it is closed and 0 if it is open. The output voltages of the inverter Vondi based on the dc voltage can be expressed by the following relationship: Vond1 Vc Vond2 = 3 Vond3 2 −1 −1 −1 2 −1 −1 u1 −1 u2 2 u3 (12.43) The differential system of Equations 12.20 through 12.22 is strongly coupled. So, we apply a change of variables on all orders: uɶ1 2 ɶ 1 u2 = 3 ⋅ −1 uɶ3 −1 −1 2 −1 −1 u1 −1 u2 2 u3 (12.44) 288 Smart Grids This allows us to obtain a decoupled model of Equation 12.20; each switching function uɶi affects only the current I ondi: dI ond1 ɶ1 − RfI ond1 L f dt = V1 − uV dI ond2 = V 2 − uɶ2V − RfI ond2 L f dt dI ond 3 L f dt = Vs 3 − uɶ3Vc − Rf I ond 3 (12.45) The relationship of the inverter dc side remains unchanged: C 12.4.2 dVc = dt 3 ∑ (uɶ I i ondi ) (12.46) 1 ProJectIon In the rotatIon reFerence The projection of Equation 12.45 in the Park reference frame (dq) leads to the following equation: dI ondd Vsd = R f I ondd + L f dt − L f ω I ondq + Vcud dI ondq V = R I + L f ω I ondd + Vcuq f ondq + L f sq dt (12.47) Considering the rotating reference linked to the voltage of phase 1, we can write Vsd = Vm Vsq = 0 (12.48) The inverter model expressed in the Park reference frame is given by Vc dI ondd Vm R f dt = L − L I ondd + ω I ondq + L ud f f f dI ondq Rf Vc I ondq − ω I ondd − =− uq Lf Lf dt dVc 2 I ondd * ud + I ondq * uq = d 3 t C ( 12.4.3 (12.49) ) control vectors The logic states of the inverter are given in Table 12.1. There are six vectors of control. These vectors form the vertices of a hexagon, shown in Figure 12.9. Photovoltaic Energy Generation and Control for Active Power Filter 289 TABLE 12.1 Inverter States (uɶ1, uɶ2, uɶ3) (u1, u2, u3) (uα, uβ) Vector A (0,0,0) (0,0,0) (0,0) (0,0,1) (−1/3,−1/3,2/3) (−1/3,−1/ 3) B (0,1,0) (−1/3,2/3,−1/3) (−1/3,1/ 3) C (0,1,1) (1,0,0) (−2/3,1/3,1/3) (−2/3,−1/3,−1/3) (−2/3,0) (2/3,0) D E (1,0,1) (1/3,−2/3,1/3) (1/3,−1/ 3) F (1,1,0) (1/3,1/3,−2/3) (1/3,1/ 3) G (1,1,1) (0,0,0) (0,0) H β 1 C: (0,1,0) D: (0,1,1) A: (0,0,0) G: (1,1,0) 3 H: (1,1,1) 1 E: (1,0,0) α 3 B: (0,0,1) F: (1,0,1) FIGURE 12.9 Representation of different control vectors. 12.4.4 descrIPtIon oF the control laWs In order to insert the control in the (dq) reference frame, we set the following functions: * σd = Iondd − Iond d * σ = I − I q ond ond q q (12.50) * * The components I ond and I ond represent the reference currents that will serve as q d inputs for the inverter. The two corresponding switching surfaces are thus deined by * Sd : σd = I ondd − I ond d * S : σ = I − I ondq ondq d q (12.51) 290 Smart Grids They can be calculated from Equation 12.34. In our case, the derivatives of the reference currents are not zero and the model in the Park reference frame gives dI fd VSd R f Vc dt = L − L I fd + ωI fq − L ud f f f dI fq VSq R f V I fq − ωI fd − c ud = − d t L L L f f f Vc 3 I fd ud + I fq ud = Lf 2C ( (12.52) ) The equivalent command is then given by udeq = L f dI fd VSd Rf I fd + ωI fq + − − Vc dt L f L f L dI f VS R uqeq = f − q + q − f I fq − ωI fd Vc dt L f L f 12.4.5 (12.53) slIdIng mode exIstence condItIons The convergence of the sliding surface refers to the following conditions: σd σq σd 〈0 dt σq 〈0 dt (12.54) Now, according to relations 12.51 through 12.53, we have Vc σd = L ( udeq − ud ) s σq = Vc ( uqeq − uq ) Ls (12.55) The convergence condition is equivalent to the following equations: sgn ( σd ) = sgn ( ud − udeq ) sgn ( σq ) = sgn ( uq − uqeq ) (12.56) The expression of controls ud and uq is given by the Park transformation: ud = cos ( θ ) ⋅ uα + sin ( θ ) ⋅ uβ uq = − sin ( θ ) ⋅ uα + cos ( θ ) ⋅ uβ (12.57) Photovoltaic Energy Generation and Control for Active Power Filter 291 Given that the equivalent commands must be bounded, leaving the minimum and the maximum, one can deduce them from Equation 12.58 in order to regulate the dc voltage of the inverter: dVc2 ( t ) 1 Pf = − ⋅ C f ⋅ 2 dt (12.58) The orders of uα and uβ are conined to ± 23 ; udq = ud2 + uq2 must be less than or equal to the standard uαβ = uα2 + uβ2 . The vector ueq must therefore be inside the circle inscribed in the hexagon in Figure 12.10. The suficient condition for the existence of sliding regimes therefore requires that the two commands are equivalent bounded by the two following equations: ud 〈 23 2 uq 〈 3 (12.59) This condition depends on the values chosen for the dc-bus voltage Vc and the coupling inductance of the inverter network (Lf). The worst case occurs between two phases of the rectiier conduction, because it is during this phase that the gradient of the load current is highest. Consider, for example, the switching phase between phases 2 and 3, which occurs during the time interval [0, time]. q β q′ C: (0,1,0) G: (1,1,0) d′ d θ D: (0,1,1) E: (1,0,0) α B: (0,0,1) F: (1,0,1) FIGURE 12.10 Example of choice of control vectors. 292 Smart Grids Assuming that the currents change linearly during the encroachment, the load currents are deined by the following system: Ich = I 1 d Id t Ich2 = − I d + time Id t Ich3 = − time (12.60) According to the Concordia transform, this leads to Ichα = I d 2 * Id 1 Ichβ = 3 − I d + t ime t (12.61) If the current Id is assumed to be perfectly smooth, then the derivatives of the currents Ichα and Ichβ are dIchα dt = 0 dIchβ = 2 * I d dt 3time (12.62) They can be written in the Park reference frame as 2 * Id dIchd dt = ω Ichq + 3t sin ( ωt ) ime d Ich 2 Id * q = −ω Ichd + cos ( ωt ) dt 3time (12.63) From Equation 12.63, we deduce dIfd* = ωIfq* − d t * dIfq = −ωIf * − d dt 2Id sin(ωt ) 3time 2Id sin(ωt ) 3time (12.64) Photovoltaic Energy Generation and Control for Active Power Filter 293 Assuming that Ifd* = Ifd and Ifq* = Ifq, and Rf can be neglected, the equivalent commands are expressed as Equation 12.65: Vm 2 * Lf * I d sin(ωt ) udeq 0 = V + Vm c + Vc 3time Vc Lf 2 * I d cos(ωt ) uqeq 0 = V 3 t c ime (12.65) The phase duration is [0, time] if t = time and τ = time/ω: Vm 2 * Lf * ω * I d sin(ωtime ) udeq 0 = V + ω * time Vc 3 c Lf 2 * ω * I d cos(ωt ime ) u = qeq 0 Vc ωtime 3 (12.66) We consider the following simpliication with sin ( ω t ime ) ≈1 ω * t ime Vm 2 * Lf * ω * Id udeq0 = V + V0 3 c L 2 * ω * I d cos(τ) u = f qeq0 Vc τ 3 (12.67) (12.68) The conditions of Equation 12.59 give 2 * L f * ω* Id 3 Vc 〉 Vm + 2 3 V 〉 3 L * 2 * ω * I d cos(τ) c 2 f τ 3 (12.69) To ensure sliding mode, the value of Vc must take the maximum of these two values. 294 12.4.6 Smart Grids choIce oF vector FIeld The appropriate vector to ensure the convergence to the desired sliding surfaces must be chosen. We recall that the convergence condition is subject to the direct: sgn ( σd ) = sgn ( ud − udeq ) sgn ( σq ) = sgn ( uq − uqeq ) (12.70) To represent the magnitude (ud − udeq) and (uq − uqeq), we work in another rotating reference frame (dʹ, qʹ) that is deduced from (d, q) by the translation vector ueq. The axes of the rotating reference frame are oriented according to the sign of (ud − udeq) and (uq − uqeq). Thus, in the rotating reference frame (dʹ, qʹ), the sign of (u d − u deq) and (uq − uqeq) is directly determined by the quadrant or located (see Figure 12.10). For example, if the surfaces σd and σq are both positive and if the position θ is given arbitrarily, then the convergence condition (Equation 12.59) requires that the terms (u d − u deq) and (uq − uqeq) must be positive. This is only possible for a vector located in the irst quadrant of the reference frame (dʹ, qʹ), and only the vector G can guarantee this. We set the two following equations: λ = arc sin ( 23 uqeq ) and µ = arc cos ( 23 udeq ) (12.71) The different values of θ are determined by the passage of the axis dʹ or qʹ in one of the six vertices of the hexagon, which provides a boundary of θ. The choice of switching controls is directly determined by the choice of vector B, C, D, E, F, or G, which depends on three parameters: the sign of σd and σq, the vector ueq (and there* * fore the instructions Iondd and I ondq), and position θ. Finally, Table 12.2 provides the values of the switching vector. Table 12.2 allows the switching sector to be selected in the ideal case, that is to say, when π 3 − µ + λ = 0. However, this condition is valid only for a particular operating point. In the general case, two situations may arise, depending on the values of λ and μ. The operating point to be determined in the working space depends at the same time on the values of λ and μ. Actually, two lookup tables are used, depending on the sign of π 3 − µ + λ = 0. The regions of θ, mod 2π, are given in Table 12.3. 12.5 THE PHOTOVOLTAIC SOURCE Recently, there has been increasing concern about environmental pollution. The need to generate pollution-free energy has triggered intensive efforts in search of alternative sources of energy. In particular, solar energy is a promising option. 295 Photovoltaic Energy Generation and Control for Active Power Filter TABLE 12.2 Table of Switching Vectors Region θ Module 2π σd > 0 and σq > 0 σd > 0 and σq < 0 σd < 0 and σq > 0 σd < 0 and σq < 0 θ 2π 2π − λ, 3 − µ G E D B θ1 π 3 − λ, π − µ C G B F θ2 4π 2π 3 − λ, 3 − µ D C F E θ3 5π π − λ, 3 − µ B D E G θ3 4π 3 − λ, 2π − µ F B G C θ4 7π 5π 3 − λ, 3 − µ E F C D θ5 Some researchers have worked on developing a system combining an APF and a PV panel [8]. However, the existing HAPF coniguration has not yet been implemented in an experimental PV application. The power output of a solar PV module changes with the solar elevation angle, changes in the solar insolation level, and varying temperature [9]. Hence, the maximization of power improves the utilization of the solar PV module. The simplest solar cell model consists of a diode and a current source connected in parallel. Source current is directly proportional to the solar radiation. The diode represents the PN junction of a solar cell. iPV = I sc − id (12.72) i = I 0 (eeV / KT − 1) − I L ⇒ i = I 0 (eVD /VT − 1) − I d (12.73) The proposed model of a PV cell using MATLAB® is shown in Figure 12.11. The PV system consists of: 1. A nine-module (85 W each) PV array with full sun (1000 W/m2 insolation) 2. A boost dc-dc converter that steps up a dc input voltage 3. A capacitor C under dc voltage that provides the necessary energy storage to balance instantaneous power delivered to the grid 296 Smart Grids TABLE 12.3 Improved Vectors of Choice for Switching π −µ + λ < 0(disunion) 3 12.5.1 π −µ + λ > 0(overlap) 3 7π 2π − µ , − µ min −λ, 3 3 7π 2π − µ , − µ max −λ, 3 3 2π π − µ , π − µ min − λ, 3 3 2π π − µ , π − µ max − λ, 3 3 2π 4π − λ, π − µ , − µ min 3 3 2π 4π − λ, π − µ , − µ max 3 3 4π 5π − µ , − µ min π − λ, 3 3 4π 5π − µ , − µ max π − λ, 3 3 5π 4π − λ, − µ , 2π − µ min 3 3 5π 4π − λ, − µ , 2π − µ max 3 3 5π 7π − λ, 2π − µ , − µ min 3 3 5π 7π − λ, 2π − µ , − µ max 3 3 solar cell model usIng orcad The characteristics of the photodiode are provided by the producer of the PV array. Figure 12.12 shows the characteristics of the photodiode that will be used. Id = f(Vd) Looking at Figure 12.12, we can conclude that a geometric approximation represented by four segments will be suficient for our needs, so the new model will be as shown in Figure 12.11. The values of the resistance are calculated using Equation 12.72 [10]. Ri = 1 Ei mi − mi −1 m0 = 0 (12.74) Vi = Ei The new approach is modulated using Orcad 15.7, which is considered a new version of PSpice with a high ability to produce an electronic board from the simulated circuit. Figure 12.11 shows this new model with four parallel branches. The characteristic of the PV is represented by the four segments, the irst graph is the current (Ipv) as a function of the voltage of the PV (Vpv), and the second is the power (P) as a function of the same variable (Vpv). 297 Photovoltaic Energy Generation and Control for Active Power Filter m4 rox im atio nb y se Id (A) gm en t s Diode I-V curve m3 Ap p m2 m1 E1 E2 E3 E4 Vd (V) FIGURE 12.11 Real model of the PV. PV power Vpv X ppv Product 1 Insolation 1–9 ∗(e(u/26∗1 )–1) PN-junction characteristic Isc – 1 + Ipv To workspace I-V characteristic –3 Vpv PV Id Ipv Insolation to Isc current gain FIGURE 12.12 Model of PV cell using Simulink. 12.5.2 maxImum PoWer PoInt trackIng The power output of a solar PV module changes with the changing direction of the sun, with the solar insolation level, and with varying temperature. Hence, in order to get the maximum power from the solar PV, an interface should be added. A maximum power point tracker (MPPT) is used for extracting the maximum power from the solar PV module and transferring that power to the load. A dc/dc converter (step up/step down) serves the purpose of transferring the maximum power from the solar PV module to the load. 298 Smart Grids We choose the hill climbing method, which consists of observing the current and voltage at the output of the PV generator. Multiplying these two signals allows the power to be obtained. The value of the power is stored in two integrators in order to be used as two successive values in memory. Two successive values of the power can thus be compared to determine the power’s derivative. Then, using a JK toggle with a third integrator and a triangular signal allows the pulse width modulation control to be created and sent to the driver of the transistor (we use a P-channel metal–oxide semiconductor ield-effect transistor [MOSFET]). Figure 12.13 shows a diagram of this technique. The MPPT controller represents a good compromise between high performance and robustness. However, to achieve convergence in good conditions, whatever the power range and the algorithm, it is necessary that the power curves issued by the generator are constant or slowly varying. If this assumption is not respected (i.e., if there are some abrupt changes in operating conditions), then the system can diverge. The principle of the MPPT controller is often based on the “elbow” of the PV characteristics. This is more or less trial and error, as shown by Figure 12.14. Indeed, the algorithm starts from a point located on the curve, let us say X1. After a while, the actual value on the curve (X2) is compared with the previous one (X1). If X2 > X1, then we move to X2, and this is repeated until the next term (Xn) is lower than the previous (Xn−1). At this time, we take an interval value from every lower point, and repeat from (Xn−1) to obtain the maximum power point (MPP) (X). The MPPT controller consists of two distinct parts: 1. The control part, whose purpose is to determine the operating point, where the solar cell can transfer the power toward the batteries 2. The power part, which transfers energy between the solar panels and batteries These two parts are also shown in Figure 12.15. Now it is easy to transfer to the electronic schematic from the Orcad library, as shown in Figure 12.16. Idc {Idc} R1 5.88 V1 9 R2 12.5 V2 13.5 R3 0.86 V3 15 R5 0.02 D1N4001 D4 D1N4001 D3 D1N4001 D2 Parameters: Rs = 10 D1N4001 D1 Parameters: Idc = 3.1 R4 0.38 Rp 50 V4 18 0 FIGURE 12.13 Block diagram of the proposed MPPT controller. Rs {Rs} 299 Photovoltaic Energy Generation and Control for Active Power Filter Model PV Converter dc-dc R1 Rs Rsh Driver MOSFET R2 Quick integrator T1 K + JK Compare Multiplier Quick integrator T2 Inverter amplifier – Integrator (R0,C0) Compare Clear Triangular generator FIGURE 12.14 Principle of operation of MPPT. X (MPP) Xn P (PV) Xn–1 X2 X1 V (PV) FIGURE 12.15 Block diagram of this type of control. Some techniques for implementing an MPPT depend on the changing bias point and measuring changes in the output, while others use predetermined PV models. Table 12.4 shows all the MPPT methods that can be employed. The perturbation and observation (P&O) method is one of the most popular MPPT methods because of its simplicity. The P&O method operates by making small incremental changes in voltage and measuring the resulting power change. By comparing the current power measurement with the previous power measurement, the P&O method selects the direction for the next perturbation. The P&O MPPT method can be implemented using a minimal number of components; however, its 300 Smart Grids P Solar array P0 Power part Battery PWM P Control part V0 FIGURE 12.16 Proposed controller circuit diagram. TABLE 12.4 Table of Various MPPT Methods MPPT Method 1 2 3 4 5 6 7 8 9 10 P&O Hill climbing Incremental conductance Power feedback control Fractional short-circuit current Fractional open-circuit voltage Computational and lookup table Current sweep Fuzzy logic Array reconiguration speed is limited by the size and the period of the perturbation [11]. Figure 12.17 shows the lowchart of this algorithm. In a situation where the irradiance changes rapidly, the MPPT also moves along the right-hand side of the curve (see Figure 12.14). The algorithm takes this as a change due to perturbation, and in the next iteration it changes the direction of perturbation, hence moving away from the MPP, as shown in Figure 12.18. The proposed MPPT scheme employing the P&O algorithm has been modeled and simulated using the MATLAB/Simulink® software. Figure 12.19 shows a block diagram of the developed PV system, consisting of a PV array and a boost converter circuit connected to a load and an MPPT controller. The PV array, with a capacity of 1 kW, consists of six PV modules associated with an MPPT controller. All this is simulated with conventional P&O under operating conditions of temperature 25°C and with varying insolation (0–1000 W/m2). The MPPT controller is based on the P&O method. 301 Photovoltaic Energy Generation and Control for Active Power Filter Parameters: Parameters: Idc = 25 D17 MUR150 Rs = 50 R1 5.88 V1 9Vdc D4 R2 12.5 R3 0.86 R4 0.38 V2 13.5Vdc V3 15Vdc V4 18Vdc D13 D1N4001 1 IRFP9140 M1 L2 2 1mH 20m D1N4001 D3 D1N4001 D2 D1N4001 D1 Idc {Idc} D1N4001 R5 D15 Rp 50 Rs {Rs} C6 3.5u MUR150 12 V11 100 C5 R19 0.1 R27 0 100 V 4 3 6 R11 -VCC AD633/AD VCC 220k 220n 0 C3 R33 VCC 6 V+ 7 220n C2 1k + U12A 1 LM139 OUT 1 2 – 12 12 8 0 0 R10 22k 8.2k V-VCC 3 U15 J Q CLK K Q JKFF 5 R12 V 1k 7 C4 R22 1k 5 R21 1k 15 1 R34 1k -VCC 0 – 7 LM324 OUT Jcc 5 15 OUT + 3 U13A V+ V- 6 LM239 VCC 3.3u 0 R20 1k 15 – 4 11 0 + 4 U6C V+ VCC 5 10 – R16 LM324 OUT 8 5k -VCC R39 VCC VCC X1 X2 7 Y1 W Y2 Z 7 + 8 U19B V+VCC 0 10k V- U16 V+ 11 9 1k 1 2 3 4 6 V- R29 OUT 5 V- VCC R17 22k 4 2.2k + 8 U18A V+ VCC R38 0 -VCC V- 1 3 R28 6n TL082 6 – R37 OUT 0 100k -VCC C8 -VCC V- TL082 2 – + 4 U7B 0 V+ VCC 0 0 -VCC FIGURE 12.17 Flowchart of the P&O algorithm. The whole PV system associated with an APF, as shown by Figure 12.19, has been simulated. The performance of the PV system using the P&O algorithm under a constant temperature of 25°C and a changing irradiance is shown by Figure 12.20. The irradiance values are 0, 400, 850, 950, 1000, 950, 850, 400, and 0 W/m2 at times 0, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, and 0.08, respectively. It can be seen that the power issued by the converter is close to the PV power (Figure 12.21). 12.5.3 dIrect current voltage oPtImIzatIon WIth the PhotovoltaIc source The shunt APF and the PV array are connected back-to-back with a dc-bus capacitor. The dc-bus voltage controller maintains the average voltage across the dc-bus capacitor constant against variations in distribution source. As a result, the reference current injected into each phase allows the grid to gain more useful active power. The voltage across the capacitor must be constant and ixed at a predetermined value. Active power losses in the active ilter (switches and output ilter) are the main factors that may change the voltage. The average voltage across the capacitor energy storage must be regulated by adding the current fundamental asset in the reference currents. 302 Smart Grids Begin P&O algorithm Measure V(k), I(K) P(k) = V(k)×I(k) ∆P = P(k)–P(k–1) No Yes ∆P>0 V(k)–V(k–1)>0 V(k)–V(k–1)>0 No Decrease module voltage Yes Yes No Increase module voltage Decrease module voltage Increase module voltage Update history V(k–1) = V(k) P(k–1) = P(k) FIGURE 12.18 Curve showing incorrect tracking of MPP by P&O algorithm under rapidly varying irradiance. This stabilizes the dc voltage of the capacitor C of the three-phase voltage power inverter close to a continuous value. This has been veriied by simulations, and the result can be seen in Figure 12.22. 12.5.4 control oF the sWItchIng Frequency The switching frequency strongly depends on the coupling inductance Lf and on the width of the hysteresis band. The objective of this part is to propose a method to limit the frequency switching and make it independent of the parametric variations. For this, we propose to add to the expression of the surface a triangular signal with constant frequency and ixed amplitude. The principle of the method is illustrated in Figure 12.23. Photovoltaic Energy Generation and Control for Active Power Filter 160 1000 W/m2 D 140 Module output power (W) 303 120 750 W/m2 C ∗ G 100 ∗ F 80 500 W/m2 B ∗ E 60 250 W/m2 40 A 20 0 0 5 10 15 20 25 30 35 Module voltage (V) 40 45 50 FIGURE 12.19 Curve showing wrong tracking of MPP by P&O algorithm under rapidly varying irradiance. Photovoltaic array Vpv + Boost converter Input Output Vx Vin Vo Active power filter Vout Cpv Ipv Idc Iref Perturb and observe FIGURE 12.20 Functional block of the whole system using the P&O technique. 500 400 300 200 100 0 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 FIGURE 12.21 Power of the converter Pout-Boost, photovoltaic Ppv, and the ideal output Pideal. 304 Smart Grids 0.1 0.05 0 FIGURE 12.22 Evolution of the dc-bus voltage around its reference value at 300 V. σ+ u + m –m 1 Fdec 2δ FIGURE 12.23 Principle of limiting the switching frequency. 12.6 SIMULATION RESULTS The proposed HAPF was simulated using the MATLAB/Simulink software. The system parameters that were used are summarized in Table 12.5. In the simulation, a diode rectiier with a dc-link capacitor and a smoothing inductor was used as a nonlinear load producing harmonics. The complete scheme of the control structure that has been developed in this study to govern the APF supplied by a PV array and a capacitor is shown in Figure 12.23. Furthermore, the high-order harmonics are iltered from the source voltage and the source current. The simulation results show that the source currents after compensation are almost sinusoidal and in phase with the voltage at the connection point. The harmonic spectra of the source current before and after compensation are presented in Figures 12.24 and 12.25, respectively. The THD of the current source (calculated on the irst 30 harmonic orders) has decreased from 30.09% to 1.95%, and the load current harmonic components are greatly reduced. The dc-bus voltage is regulated to 300 V. All results are shown in Figure 12.26 (the source, the current of load before and after the compensation, the current of the ilter, and the dc voltage of the capacitor). Figure 12.27 shows the spectrum of the source current Is in decibels. This spectrum allows us to see that the switching frequency varies between 5 and 10 kHz (see Figure 12.25). The average switching frequency is equal to 9.07 kHz. Low-frequency peaks are located between −40 and −60 dB. Above 4 kHz, the amplitudes of the Photovoltaic Energy Generation and Control for Active Power Filter 305 TABLE 12.5 Matlab/Simulink Simulation Parameters Utility voltage 220 2 (50 Hz) Source inductance 0.65e-3 H Source resistance 5−3 Ω Rectiier smoothing inductor 1−3 H Rectiier smoothing resistance 1.5 Ω Load inductance 100−3 H Load resistance 29*1.5 Ω Load capacitance 2000−6 F Current control band 1.8 A Sampling time 10−6 s APF inductor 2−3 H APF resistor 0.005 Ω APF dc-bus capacitor 450 V HPF capacitor 5.01e-6 F HPF resistor 20.46 Ω PV 9 arrays Insolation 1000 W/m2 Short-circuit current Isc 5.45 A Open-circuit voltage Voc 22.2 V Rated current IR at MPP 4.95 A Rated voltage VR at MPP 17.2 V Capacitance of the PV output 2−3 F Resistance of PV output 72.78 Ω lines are increasing, and once the frequency 7.5 kHz is reached, they are decreasing. Beyond 16 kHz, the spectrum is more regular, and rays have relative amplitudes below −50 dB. 12.7 CONCLUSION This chapter has detailed a new technique for improving the eficiency of the sliding mode controller that is used to manage the PV energy in an autonomous shunt APF. In order to improve the reference current for compensating harmonics, a method 306 Vs123 (V) Smart Grids 200 0 –200 0.2 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 –20 0.2 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 702 700 698 0.2 0.205 0.21 0.215 0.22 Time (s) 0.225 0.23 0.235 0.24 iL (A) 20 0 vc (V) if, irefi (A) is (A) –20 0.2 20 0 –20 0.2 20 0 FIGURE 12.24 Compensation of harmonic currents with the sliding mode control. 30 Harmonic spectrum 25 20 15 10 5 0 0 500 1,000 Frequency f (Hz) 1,500 FIGURE 12.25 Harmonic spectrum of source current before compensation. to limit the switching frequency has been added and veriied. The THD has been improved and the average switching frequency has only slightly increased. To reduce the switching frequency without changing the system settings, a technique based on the addition of a triangular signal to the switching surface has been proposed and tested (see Figure 12.23). 307 Photovoltaic Energy Generation and Control for Active Power Filter Reference value Vc Regulator of voltage Pl Low pass filter u1,2,3 Inverter λ µ = arccos 3 2 udeq Vαβ Equivalent control λ = arcsin 3 2 Source of voltage Vs123 µ Park uqeq θ PLL Hysteresis bloc Ifa,1,2,3 Ifad,q Park Gissant surface Ĩchd,q Harmonic references calculation Nonlinear load Ich,1,2,3 Vs123 Source FIGURE 12.26 Complete diagram of the control structure. A P&O MPPT has been implemented to maximize the PV output power. Through a boost dc-dc converter combined with the fastness and robustness of sliding mode control, the HAPF has good robustness and stability. A loop for capacitor voltage regulation is introduced into the sliding mode control; a good regulation effect is obtained for load variation in the dc bus. The simulation results confirm the effectiveness of the proposed scheme for wideband harmonics compensation and PV power handling capability (see Figures 12.27 and 12.28). 308 Relative amplitude in decibels Smart Grids 10 0 –10 –20 –30 –40 –50 –60 –70 –80 –90 –100 0 0.5 1 1.5 2 Frequency f (Hz) 2.5 3 × 104 FIGURE 12.27 Spectrum of the source current Is in decibels. 30 Harmonic spectrum 25 20 15 10 5 0 0 500 1,000 Frequency f (Hz) 1,500 FIGURE 12.28 Harmonic spectrum of source current after the compensation. 12.8 HAPF APF PV σ(x) X u U 2δ f, g C NOMENCLATURE Hybrid active power ilter Active power ilter Photovoltaic Surface of sliding System error Control variable Control vector Bandwidth around the sliding surface Functions of the nonlinear system Coeficients of σ(x) Photovoltaic Energy Generation and Control for Active Power Filter 309 REFERENCES 1. H. Akagi, A. Nabae and A. Satoshi, Control strategy of active power ilters using multiple voltage-source PWM converters, IEEE Transactions on Industry Applications, IA-22, 460–465, 1986. 2. D. Ould Abdeslam, P. Wira, J. Merckle, D. Flieller and Y.-A. Chapuis, A uniied artiicial neural network architecture for active power ilters, IEEE Transactions on Industrial Electronics, 54, 61–76, 2007. 3. S. Rahmani, A. Hamadi, N. Mendalek and K. Al-Haddad, A new control technique for three-phase shunt hybrid power ilter, IEEE Transactions on Industrial Electronics, 56, 2904–2915, 2009. 4. A. Hamadi, S. Rahmani and K. Al-Haddad, Sliding mode control of three-phase shunt hybrid power ilter for current harmonics compensation, in Proceedings of the IEEE International Symposium on Industrial Electronics (ISIE), pp. 1076–1082, 4–7 July, Bari, 2010. 5. A. Blorfan, J. Mercklé, D. Flieller, P. Wira and G. Sturtzer, Sliding mode controller for three-phase hybrid active power ilter with photovoltaic application, Smart Grid and Renewable Energy, 3, 17–26, 2012. 6. B. Cheng, P. Wang and Z. Zhang, Sliding mode control for a shunt active power ilter, in Proceedings of the Third International Conference on Measuring Technology and Mechatronics Automation (ICMTMA), pp. 282–285, 6–7 January, Shanghai, 2011. 7. D. Wenjin and Yu. Wang, Active power ilter of three-phase based on neural network, in Proceedings of the Fourth International Conference on Natural Computation, ICNC ‘08, pp. 124–128, 18–20 October, Jinan, 2008. 8. A. Blorfan, D. Flieller, P. Wira, G. Sturtzer and J. Merckle, A new approach for modeling the photovoltaic cell using Orcad comparing with the model done in Matlab, in International Review on Modelling and Simulations (IREMOS), vol. 3, pp. 948–954, October 2010. 9. G. Sturtzer, Projet Simulation Panneaux Photovoltaïques, p. 8, 2009. 10. S. Chowdhury, S.-P. Chowdhury, G.A. Taylor and Y.H. Song, Mathematical modelling and performance evaluation of a stand-alone polycrystalline PV plant with MPPT facility, in Proceedings of the 2008 IEEE Power and Energy Society General Meeting—Conversion and Delivery of Electrical Energy in the 21st Century, pp. 1–7, 20–24 July, Pittsburgh, PA, 2008. 11. M. Veerachary, T. Senjyu and K. Uezato, Voltage-based maximum power point tracking control of PV system, IEEE Transactions on Aerospace and Electronic Systems, 38, 262–270, 2002. 13 Self-Tuning and SelfDiagnosing Simulation Jin Ma CONTENTS 13.1 Introduction .................................................................................................. 311 13.2 Framework of Self-Tuning and Self-Diagnosing Simulation ........................ 312 13.3 Integrate the Synchronous Measurements with Digital Simulation ............. 315 13.4 Model Data Calibration ................................................................................ 320 13.5 Locate the Disturbance Source for Low-Frequency Oscillation .................. 325 13.6 Conclusion .................................................................................................... 331 Acknowledgments.................................................................................................. 331 References .............................................................................................................. 331 13.1 INTRODUCTION The bulk power system is usually divided into the transmission network and the distribution network. The transmission network is responsible for transferring the bulk electrical energy from the generator side to the demand center, while the distribution network connects the end users. There are two ways to design the generation layouts. One is to build a large power plant in an area with abundant raw energy, for example, near the coal base or the hydrobase. Consequently, the electrical energy produced by this kind of plant is transferred long distance on the high and the ultrahigh voltage levels, usually above 110 kV. The other way to plan the generation is to disperse small-scale generation facilities in the demand center, which not only minimizes the risk of blackout for the users, but also serves as a backup for the transmission network. One of the main targets of the smart grid is to provide clean and secure electrical energy with high quality. But how? The transmission network will be strengthened; advanced electronic technologies will be applied to better integrate the distributed generation into the demand side; more lexible and interactive methods of energy utilization will be attempted. But, when we do this, we have to understand how this system is operated and what will happen under our smart controls. The power grid, as built so far by human beings, is too complex to be easily understood. So, one may ask why this giant machine has been operated for such a long time although we do not understand it very well. Actually, it is not operated very well. We manage it very cautiously and operate it very conservatively because we do not know the capability limits of this machine. Yet we have witnessed several 311 312 Smart Grids major blackouts in the last decade. We have to say NO to this traditional operation mode with the advent of the smart grid, since it requires high eficiency in energy utilization without sacriicing security. So, a compromise between security and economy has to be made. At least, we have to know how much potential transfer capability this network has before we take full advantage of it. But, again, we have to ask ourselves how. Performing tests on this system is certainly one solution. For things that we do not know, we have to employ tests. However, there are three major challenges to performing tests in the real power system. First, system security is always top priority. Therefore, a major category of tests, especially those that may cause serious damage, such as short-circuit tests, cannot easily be applied to the main power system. Second, the operation scenarios are always changing. The load is different at any time. The generation units and the controllers in operation vary with time. One cannot design a test case covering all the scenarios. The operation case is always different from the test case. Third, the possible contingencies are numerous, and it is impractical to perform all these contingency tests in the grid. Consequently, digital simulation becomes a key tool for the operators in understanding the power system operation mechanisms and designing control strategies. Reliable simulation is crucial to system security, which is especially important to the modern, largescale power system. It is the prior condition for inding the balance point between economy and security in the coming smart grid era. The traditional simulation is one-way. The smart grid is equipped with an advanced global measurement system in the whole grid consisting of accurate and smart meters. Based on this, a two-way simulation is possible, that is, self-tuning and self-diagnosing simulation. It would be possible to combine advanced measurement technology with digital simulation, greatly enhancing the simulation reliability; thus, the security of the whole power system could be ensured. This chapter will present the framework of the self-tuning and self-diagnosing simulation and propose a speciic approach to combining wide-area measurements (WAMs) with the digital models of the power system component. This is called the variable impedance method. This chapter also introduces two cases of applying this self-tuning and self-diagnosing simulation in the real power grid. One concerns simulation validation and model data correction, while the other concerns locating the disturbance source to minimize the possibility of resonant oscillation in a large power grid. The chapter is organized as follows: the general framework of the self-tuning simulation is given in Section 13.2; the speciic methods to incorporate the real measurements into the simulation are discussed in Section 13.3; two application ields, as stated before, are discussed in Sections 13.4 and 13.5; and Section 13.6 concludes the chapter. 13.2 FRAMEWORK OF SELF-TUNING AND SELF-DIAGNOSING SIMULATION Power system simulation can be deined as using models, either digital or physical, to simulate the real operation scenarios of the power system. Due to the rapid development of computer technologies and the greatly increased complexity of the power 313 Self-Tuning and Self-Diagnosing Simulation Model data preparation WAMs Selected contingency list Recorded contingency list Digital simulation CM/M&M simulation Results output Match? Y N End Model parameters calibration FIGURE 13.1 Flowchart of closed-loop power system simulation. grid, digital simulation has been playing an increasingly important role. The usual way of carrying out power system digital simulation consists of four steps, as shown in the left part of Figure 13.1. Before starting the simulation, it is necessary to collect all the model data from either the ield tests or the manufacturers. Usually, this is done repeatedly when the new devices are put into operation in the utility. The data are then stored in the database and will be retrieved whenever the simulation is to be implemented. The second step of the digital simulation is to build the contingency lists based on the operators’ knowledge of the power grid. Since numerous possible faults exist, the operator will select the most serious ones based on various system operation scenarios. The digital simulations are carried out case by case for all the contingencies in the list to ensure that the system is securely operated under all possible serious faults. For those faults that may cause serious damage to the system, such as blackouts, additional digital simulations must be performed to ind proper control measures to minimize the risks. All the information obtained from the simulation will be sent to the system operator to aid decision making on the system operation. This is a typical open-loop simulation, starting from the data preparation and ending with the output of results. However, one key element of the simulation is missing, that is, the simulation validation. Sometimes it is far too easy to trust the output of the simulation, but think too little about its idelity. The computer may give a completely incorrect image of the operated system, partly due to inappropriate algorithms and incorrect model parameters. False simulation results could also come from the uncertain components of the power system, such as the load. Since 314 Smart Grids the load is always changing, both in its composition and in its amplitude, the difference between the real load and its simpliied model will affect the simulation results. Typical lessons in unreliable simulation can be learned from the 1996 Western Systems Coordinating Council (WSCC) blackouts [1]. After cascading contingencies, the WSCC system witnessed undamped oscillations and consequently broke into several islands. The postfault simulation analysis based on WSCC’s operation database indicated that the blackout could not be reproduced. Even worse, the simulation showed a very stable system, in contrast to the real situation. It is dangerous to make an optimistic simulation, which gives the system operator false conidence and puts the power system under unknown potential risks. On the other hand, if the simulation is too pessimistic and exaggerates the fault hazards, the power system will be operated very conservatively. The eficiency of the energy transfer will drop and the utilization rate of the power lines and power transformers will be low, which causes great waste. To enhance the simulation validity, the model and its parameters have to be calibrated frequently. However, due to various reasons, including obstacles on the operation side, the technique side, and the management side, routine ield tests in order to correct the model parameters are often delayed. For those models with great uncertainty, for example, the load model, their parameters should be updated even more frequently. As the recorded data accumulate, the law behind the supericial randomness may be revealed gradually. By continually analyzing these recorded data, we can ind the laws behind the uncertainty and apply them to adjust the model parameters. Currently, however, this mostly relies on offline data collection and analysis. The updating rate for the model parameters cannot keep up with the change of the component to be modeled. Thanks to the advent of advanced measurement technologies and communication technologies, which will be further developed in the coming smart grid era, the once open-loop digital simulation can be closed, as shown in Figure 13.1. The deployment of WAMs [2–4] provides new opportunities for the traditional simulation and widens its application ields. As shown in the other chapters of this book, WAMs could provide synchronized measurements, which are just like taking snapshots of the dynamics for the whole power system, whatever its scale. It is even more important that the WAMs make this closed-loop simulation more adaptive. This helps to ind the real problem of the power system in operation, by locating either the erroneous model parameters or the disturbance source, as will be illustrated in detail in the following sections. By fully utilizing the newly emerged measurement technologies, this closed-loop simulation could increase the reliability of the power system in operation and provide more authentic guides to the system operators. So, the simulation becomes smart; and, in turn, it helps to make a smart power grid. This self-tuning and self-diagnosing simulation has two parallel branches. The conventional simulation is still carried out based on the prepared contingency list. However, when the WAMs record the real faults or the disturbances in the power grid, the other branch of the simulation initiates itself. This simulation consists of two aspects: the complete model (CM) simulation and the model and measurement (M&M) simulation. CM simulation is based on the computer models in the operation database. Since it is carried out on the speciic fault recorded by the WAMs, the sequence of events (SOE) has to be set to be the same as the recorded one, which Self-Tuning and Self-Diagnosing Simulation 315 is the basic difference between this CM simulation and the conventional digital simulation. To fulill this, the recorded SOE must be programmed into the computer codes for piecewise simulations. This can be realized by interactive simulation packages, which are often equipped with user programming tools and automatic simulation functionality. The basic target of the CM simulation is to make a comparison with the WAMs. Thus, the validity of the current model database for system operation can be evaluated. The error criteria have to be carefully designed to relect not only numerical errors but also the dynamic characteristic differences between the simulation and the WAMs. So, this evaluation could facilitate searching for the error sources to the greatest extent. The M&M simulation can be classiied as a type of hybrid simulation. It mixes the computer model data and the measurements to carry out the simulation. It is a newly proposed method to utilize the measurements in the digital simulation. The principle and the advantage of the M&M simulation will be elaborated further in the following section. If the CM simulation or the M&M simulation matches the measurements, the models and the respective data will be reliable for use in the simulation to aid the system operation. However, if it does not, the models or the parameters, or both, have to be calibrated based on the WAMs. For power components with uncertainty, the WAMs can record their inputs and outputs, which therefore can be used to decouple these components from the power system. Then the model parameters can be identiied and updated to adapt to the changing situation. The sensitivity analysis method can be used to detect erroneous parameters. If the discrepancy between the simulation and the WAMs is caused by an unknown disturbance in the system, the WAMs can be utilized to ind the disturbance source at this stage, which may reduce disturbance hazards in future system operation. On all occasions, the M&M simulation may have to be performed repeatedly to minimize the searching space and alleviate the computational burden caused by the complexity of the system. It should be noted here that, although the self-tuning simulation and self-diagnosing simulation improve the simulation reliability and reduce the operation risks, this cannot be done once and for all. The power system is a highly nonlinear dynamic system with a great number of models and parameters. Therefore, many sets of parameters exist to make the simulation meet the WAMs, but not all of them are correct ones that can be used in the system operation. The best way to discriminate between all these candidate parameters is to take advantage of their sensitivities to the changing operation scenarios. The true set of parameters is robust to the various operation situations, both in the steady state and in the presence of faults, while the false parameters are sensitive to them. So, with increased WAMs, the self-tuning simulation can converge gradually to the correct parameter values and the real system operation scenarios. 13.3 INTEGRATE THE SYNCHRONOUS MEASUREMENTS WITH DIGITAL SIMULATION One key step in self-tuning simulation is the M&M simulation, which combines the model and the measurements to simulate the power system dynamics. Before the wide application of WAMs, this was impossible, because the measurements are not 316 Smart Grids all synchronized and consequently cannot be put into the digital simulation at the same time. The basic idea of the M&M simulation is to replace part of the variables in the digital simulation with the real measurements; therefore, the simulation data will consist of the model data and the real measurements, which can be called the hybrid data dynamic simulation. The power system experiences transients when it changes its operation state due to either faults or operation actions. The transients are usually classiied as electromagnetic transients and electromechanic transients. The electromagnetic transients involve the current and the voltage dynamics, while the electromechanic transients involve the dynamics of the generator rotor position and the rotation speed. Obviously, because of the considerable inertia of the generator rotors, the electromechanic transients last much longer than the electromagnetic transients. So these two kinds of transients are usually decoupled and analyzed individually. Since WAMs only record the electromechanic dynamic states, we will apply them here to the electromechanic dynamic simulation. The mathematical representation of the power system electromechanic dynamics can be described as xɺ = f ( x, y, p, u ) 0 = g ( x, y, p, u ) (13.1) where: x denotes the electromechanic dynamic states, such as the rotor angle and the rotor speed y denotes the algebraic states, such as the bus voltage and the phase angle p denotes the parameters u denotes the controls In Equation 13.1, the transients of the transmission network are neglected, since they fall into the domain of the electromagnetic transients and last for a much shorter time; consequently, they are represented by a set of algebraic equations, as shown in the second equation of Equation 13.1. The irst equation of Equation 13.1 represents the power system electromechanic transients. In the M&M simulation, since parts of the state variables are available from the WAMs, they will be replaced by the measurements, so that the power system dynamic equation takes the form of xɺ1 = f ( x1, y1, p, u, x2 , y2 ) 0 = g ( x1, y1, p, u, x2 , y2 ) (13.2) Compared with Equation 13.1, the dynamic states x and the algebraic states y are decomposed into two parts. The unknown states that have to be solved by the digital simulation are represented by x1 and y1, respectively, while the known measured states are represented by x2 and y2, correspondingly. Equation 13.2 represents dimension-reduced power system electromechanic dynamics compared with Equation 13.1, since only a subset of the power system states has to be solved by Self-Tuning and Self-Diagnosing Simulation 317 the step-by-step integration. Naturally, there is less computational burden in solving Equation 13.2 than in solving Equation 13.1. The other advantage of this hybrid data dynamic simulation is that the measured states x2 and y2 are independent of the system models and parameters. Thus, if the WAMs data are reliable and well calibrated, x2 and y2 represent true images of the power system dynamics. This fact is very important, since it provides a possible way to decouple a very big system model into various small pieces with the real measurements at the boundaries. The one machine ininite bus (OMIB) system can be taken as an example to illustrate the idea of the hybrid data dynamic simulation. Assuming the phasor measurement unit (PMU) is installed at the high-voltage bus side, the whole OMIB system can be decoupled into two subsystems. One of them models the generator as well as its controllers in detail, while the system outside the machine is equalized by PMU measurements; on the other hand, the whole system outside the machine will be reserved in the simulation database while the machine is equalized by the PMU measurements. Each subsystem has fewer dimensions than the original whole system. The advantages of this measurement-based decoupling work lie not only in this dimension reduction, but also in the independence of the measurements from the models, as well as their parameters. For example, when simulating the subsystem in which only the generator is modeled in detail, we do not need to consider the validity of the models and their parameters outside the generator; nor do we worry about unknown disturbance in the outside system that may affect the simulation results. This is because the synchronized measurements from the WAMs represent the real dynamics of the outside system. The conigurations to decouple a big system into small subsystems vary, but rely on the deployment of WAMs. With the wide application of modern measurement technologies, the alternatives to decouple the system become more lexible. Three typical characteristics of the power system can be utilized in decoupling the original large system. The irst characteristic of the power system is its hierarchy. The power system consists of the transmission system and the distribution system. The transmission system is operated on higher voltage levels, while the distribution system is operated on lower voltage levels. Thus, the big system can be decoupled based on the voltage levels by injecting the measurements from one side of the transformers. If the measurements are injected from the lower-voltage bus of the transformer, the grid on the lower voltage level is equalized by the WAMs. If they are injected from the higher-voltage bus of the transformer, the grid on the higher voltage level is equalized by the WAMs. The second characteristic of the power system that can be used in the hybrid data dynamic simulation is its geographic composition. The large power system consists of different utilities connected by interties. In general, the interties are equipped with advanced meters. If measurements from these meters are synchronized, for example, the WAMs, the whole system model can be decomposed into individual utility models by replacing the electrical states on the interties with these measurements. So, the simulation on a very large system becomes several simulations on small subsystems. In real applications, it is not necessary to simulate all these subsystems, but only the system that is causing trouble. This will be further illustrated later. The third characteristic of the power system is that it is composed of various components, and these components have clearly deined interfaces. 318 Smart Grids So, if the inputs and outputs of each component are well recorded and synchronized, the measurements can be utilized to decouple them from the large system, and the hybrid data dynamic simulation, therefore, can be implemented on each individual component. To realize the hybrid data dynamic simulation, the interfaces of the model data and the real measurements have to be well designed. The core of the algorithm is how to inject the measurements into the digital simulation model. There have been three state-of-the-art methods: the fast-response generator method [5], the ideal tapchanger transformer method [6,7], and the variable impedance method [8,9]. The fast-response generator method realizes the injection of the measurements by an assumed generator. The WAMs are coded as a preset sequence of the reference values to the assumed generator controller. The dynamic response of the generator is fast enough that the outputs of the generator follow the control reference values immediately. For example, the bus voltage measurements from the WAMs are input to the generator automatic voltage regulators as the target voltage references. Due to the fast response characteristics of the generator and its excitation system, the terminal voltage of this assumed generator takes the exact preset reference values at each simulation point. In this way, the measurements and the model parameters are merged for the hybrid data dynamic simulation. The ideal phase-shifting transformer method uses the changeable ratio between the primary winding and the secondary winding to realize the hybrid data dynamic simulation. The primary winding of the phase-shifting transformer connects to the ininite bus with constant voltage and constant frequency. The measurements from the WAMs are taken to modulate the tap ratio. At each simulation step, the tap ratio will be corrected by the quotient of the constant ininite bus voltage with respect to the measured voltage phasor. In this way, the voltage of the secondary-winding bus takes the exact values of the measurements. Among the three methods to implement the hybrid data dynamic simulation, the variable impedance method takes the simplest form. It applies a value-changed impedance to replace the system outside the boundary, as shown in Figure 13.2. The equivalent complex impedance Z or Y = 1/Z can be calculated from the measurements as follows: P (t ) 1 = Y ( t ) = G ( t ) + jB ( t ) = 2 + U (t ) Z (t ) Q (t ) j− 2 U ( t ) Boundary bus Subsystem Z FIGURE 13.2 The variable impedance method. (13.3) Self-Tuning and Self-Diagnosing Simulation 319 where P(t), Q(t), and U(t) represent the boundary bus voltage, the transferred active power, and the reactive power. Since these measurements change with time, the value of Z is not constant, but is also changed at every simulation time step. That is why it is called the variable impedance method. Since the value of the impedance Z is changed according to the WAMs, the digital simulation integrates the model data and the real measurements together. It should be noted here that the simulation step and the sampling period of the measurements may not be consistent. Thus, the WAMs must irst be synchronized with the simulation before being injected into the digital simulation system using any of the above three methods. No standards have been speciied for the digital simulation step length, or the time interval when the WAMs data are expected to be sent to the data server in the control center. In general, the typical electromechanic transient simulation package takes half a cycle as the simulation step, that is, 0.01 s for a 50 Hz system, while the time interval for two consecutive measured phasors is one cycle, that is, 0.02 s for a 50 Hz system. Thus, the measurements cannot be directly incorporated into the reserved digital model to implement the hybrid data simulation. One solution to this problem is to interpolate the measured points so that a one-to-one correspondence can be set up between the measured point and the simulation point [8]. We have applied the variable impedance method in many benchmark test systems and real power grids. Its eficiency has been well proved [9]. A typical three-area system in a Power System Simulator for Engineering (PSS/E) package [10] is used here to exemplify the hybrid data dynamic simulation. Assume that the synchronized PMUs are installed at the interties among Area 1 and the other two areas. Since this is a test system, we take the simulation results as the pseudomeasurements in order to derive the variable impedances connected to Area 1’s boundaries. In this way, we have a case of hybrid data dynamic simulation, with Area 1 represented in detailed models and parameters while its boundaries are injected by the timediscrete measurements. An assumed three-phase short circuit on one load bus is simulated based on the CMs and parameters of this three-area system. The active power dynamics on one generator bus is shown in Figure 13.3. The simulation results from the reduced network are also shown in Figure 13.3. It can be seen clearly that the M&M simulation using the variable impedance method produces the same dynamic trajectories as the CM simulation does. However, if we change some parameter values in the reduced network, Area 1 in this case, such as the parameters of Gen. NUC_A and Gen. NUC_B, and make them different from their values in the original complete network, the hybrid data dynamic simulation results will become erroneous compared with the simulation results for the full network, which has also been clearly shown in Figure 13.3. The hybrid data dynamic simulation curves in both cases of Figure 13.3 verify that the hybrid data dynamic simulation can relect the real dynamics of the whole network from a reduced network model. If the hybrid data dynamic simulation does not match the real measurements, it can be concluded that there is a discrepancy between the reserved models of the scale-reduced network, in this case the models in Area 1, and the real situation of the complete network. Since the injected 320 Smart Grids 780 Active power on bus 101 (MW) 760 740 720 700 680 660 Simulation from the whole network Hybrid simulation from the reduced network Hybrid simulation with wrong parameters in the reduced network 640 620 600 0 0.5 1 1.5 2 Time (s) 2.5 3 3.5 4 FIGURE 13.3 Active power on one generator bus. measurements are independent of the models and parameters in the simulation database, the discrepancy between the measurements and the hybrid data dynamic simulation results could only be caused by the differences between the reserved network and the corresponding part of the real power system. This information is very important, since it reduces the error location ranges, which greatly enhances the simulation eficiency, especially when simulating a very large system with many components. The hybrid data dynamic simulation presented in this section is the core technique in self-tuning and self-diagnosing simulation, since it provides a realizable approach to constituting a feedback closed-loop simulation using measurements, which makes self-correcting parameters possible in the simulation. 13.4 MODEL DATA CALIBRATION As stated in the prior section, when the hybrid data dynamic simulation produces different results from the measurements, there must be errors in the simulated scalereduced system. There are two major error sources. One is incorrect models or parameters, or both; the other is the unknown disturbance that happens in the real power system but without being simulated in the computer. This section addresses the irst problem, that is, the false model and parameters. The second problem is addressed in the next section. There are few incorrect model structures in the power system simulation database, since all the models are built based on the components’ physical mechanisms and conirmed by academic researchers as well as electrical engineers. Although there are electric components with uncertainty, such as the load, widely accepted models of these are used in the operation database. As time goes on and the measured data accumulate, the models can be veriied. If the hybrid data dynamic simulation is applied to an individual component, with the entire external network equalized by Self-Tuning and Self-Diagnosing Simulation 321 the variable impedance, and indicates signiicant errors between the simulations and the measurements, the validity of the model structure should be suspected. With the synchronously recorded inputs and outputs of this component, its operation mechanism can be reanalyzed and consequently its model structure can be reselected. After the parameters of the new model have been identiied from the measurements, the simulation database will be updated with the newly built model and will wait for the next recorded contingency for the model validation. In most cases when the hybrid data dynamic simulation does not match the measurements from the WAMs, this is due to incorrect parameters or changes of parameter values under different operation situations. The best solution to ind these key parameters is to use the proposed hybrid data dynamic simulation tool to decouple the system into as many subsystems as possible. This greatly reduces the searching space of each subsystem for the incorrect parameter. To cut down the searching scale further, the sensitivity method can be utilized to ind the parameter that has the greatest effect on the system simulation results. Corrections to these most sensitive parameters have the highest eficiency in itting the simulation results to the measurements. As pointed out in Section 13.2, due to the high nonlinearity of the power system and large number of components, the parameter corrections made may not be right even if they do match the simulations with the measurements. Thus, the self-tuning simulation has to repeat itself continually whenever WAMs records new dynamics, and validate the parameter constantly. A real application case in a very large power system will be presented here to illustrate the hybrid data dynamic simulation as well as the model parameter correction work [9]. The Northeast (NE) power grid is a very large interconnected power system with more than 45,000 MW installed generation capacity in 2006. The north part of the NE grid has an abundance of generators, while the south part of the NE grid is load-concentrated. So, a considerable amount of energy has to be transferred from the north to the south through long transmission lines. Meanwhile, the NE grid also outputs the energy to another large power grid through 500 kV tie-lines. So, the available transfer capability calculation is important for the system’s security and economic operation, which relies on an authentic simulation. In order to enhance the transfer capability of the transmission corridors and validate the models as well as their parameters, artiicial three-phase short circuits at the 500 kV buses were intentionally created four times in the NE power grid during 2004 and 2005. The WAMs in the NE power grid recorded the disturbed dynamics reliably during these four short-circuit tests, and therefore provided the measured database for the model parameter tuning work. Following the procedure in Figure 13.1, CM simulations are performed on the whole NE power system for the tested fault scenarios. The results show that the system dynamics cannot be reproduced by the computer simulation. However, it is not easy to locate the erroneous parameters that cause the mismatch between the simulation and the WAMs for such a large power grid. To solve this problem, an M&M simulation is carried out. The WAMs are deployed on all the 500 kV buses in the NE power grid; thus the whole system is decomposed into several subsystems from the 500 kV buses. Each subsystem has fewer components than the original large system, which minimizes the parameter correction space to a great extent. 322 Smart Grids Take the YM plant as an example. It is a key fossil power plant in the NE power grid, equipped with two 500 MW generators and producing more than 4 trillion kWh electric energy each year. The terminal voltage of these two generators is 20 kV, which is transformed to 500 kV at 2M51 buses where both generators are connected to the 500 kV power grids. PMUs are installed at the 500 kV 2M51 buses to monitor the energy transfer from the YM plant to the NE grid and the local network. Applying the variable impedance method, the whole NE grid and the local network can be replaced by the equivalent admittances with G ( t ) + jB ( t ) = P1 ( t ) + P2 ( t ) + P3 ( t ) + V 2 (t ) Q1 ( t ) + Q2 ( t ) + Q3 ( t ) j− V 2 (t ) (13.4) where V denotes the voltage of bus 2M51; and P1(t), Q1(t), P2(t), Q2(t), P3(t), and Q3(t) represent the active power as well as the reactive power recorded by PMU1, PMU2, and PMU3. The active power outputs of the YM plant from both the WAMs and the hybrid data dynamic simulations are shown in Figure 13.4. It can be clearly seen that the simulated active power cannot match the measurements. This means that some model parameters in the YM power plant, especially those that have great effects on the active power dynamics, need to be calibrated. In this way, the searching space for the incorrect parameters is limited to the YM plant, and the effects of the incorrect parameters in the YM plant on the CM-based simulation can also be evaluated very clearly. To ind the incorrect parameters and correct them, the sensitivity method is applied. Considering P0 = Vd I d + Vq I q (13.5) 1,200 Active power (MW) 1,000 800 600 400 Real measurements 200 Hybrid dynamic simulation before model validation work 0 0 0.5 1 1.5 2 Time (s) 2.5 3 3.5 FIGURE 13.4 Real measurements and hybrid data dynamic simulation of active power. 4 323 Self-Tuning and Self-Diagnosing Simulation Q0 = Vq I d − Vd I q (13.6) V0 = Vd2 + Vq2 (13.7) where V0 denotes the generator terminal voltage Vd, Vq and Id, Iq are the projections of the generator terminal voltage and the stator current at the d-axis and the q-axis, respectively P0 and Q0 are the generator active power output and the reactive power output The trajectory sensitivities P0 and Q 0 with respect to the model parameter λ can be derived as ∂I ∂V ∂P0 ∂I ∂V = Vd d + Vq q + I d d + I q q ∂λ ∂λ ∂λ ∂λ ∂λ (13.8) ∂I ∂V ∂Q0 ∂I ∂V = Vq d − Vd q − I q d + I d q ∂λ ∂λ ∂λ ∂λ ∂λ (13.9) ∂V0 Vd ( ∂ Vd ∂λ ) + Vq ( ∂ Vq ∂λ ) = ∂λ Vd2 + Vd2 (13.10) since V, P, Q on bus 2M51 can be calculated as P ≈ 2 P0 (13.11) P 2 + Q2 Q ≈ 2 Q0 − 0 2 0 XT V 0 (13.12) QX V = V0 − 0 T V0 K (13.13) where K is the transformer voltage ratio and XT is the transformer leakage reactance. The trajectory sensitivities of PMU measurements with respect to the parameter λ can be calculated as ∂P ∂P ≈2 0 ∂λ ∂λ (13.14) ∂Q Q X ∂Q P X ∂P P 2 + Q2 ∂V = 2 1 − 2 0 2 T 0 − 2 0 2 T 0 + 2 0 3 0 X T 0 ∂ ∂λ V V V ∂ ∂λ λ λ 0 0 0 (13.15) 324 Smart Grids Q X ∂V ∂V X ∂Q0 = K 1 + 0 2 T 0 − K T ∂λ V0 ∂λ V0 ∂λ (13.16) Substituting Equations 13.8 through 13.10 into Equations 13.14 through 13.16, the sensitivities of the measured active power, the reactive power, and the voltage from PMUs with respect to the generator parameters can be calculated. Equations 13.8 through 13.10 indicate that it is very important to calculate the partial derivatives of the d-axis and q-axis components of the generator terminal voltage and the current with respect to the selected parameters; nevertheless, it should be noted here that the result is dependent on the model structure. So this is again a repeating model tuning process. The model in the operation database will be taken as the initial model to start the sensitivity calculation. If the corrected parameters based on this sensitivity information cause the simulation to match the measurements, the conidence level of the current model structure will increase; if not, a more detailed model should be considered to replace the initial model. In this way the simulation reliability is improved gradually, allowing more credible operation decisions on the real power system. The generator model in the YM plant for the power system operation and control includes the rotor dynamics, the dynamics of the ield winding, and the dynamics of three damping windings. The discrepancy between the PMU measurements and the simulation is caused mainly by parameters. After the most sensitive parameters were corrected based on the sensitivity analysis, the simulations closely itted the measurements. To verify the parameter corrections, the calibrated model parameters were applied to other recorded faults. Figure 13.5 shows the PMU measurements in another threephase short-circuit test in the NE power grid, and the active power simulated using the hybrid data dynamic simulation. 1,100 Active power (MW) 1,000 900 800 700 PMU 600 500 Model before validation work Model after validation work 0 0.5 1 1.5 2 Time (s) 2.5 3 FIGURE 13.5 Real measurement and model outputs on active power. 3.5 4 Self-Tuning and Self-Diagnosing Simulation 325 From the igures, it can be seen clearly that the simulation based on the calibrated models its the PMU measurements more closely. The model’s adaptive capability to the different faults indicates good reliability of the model database. 13.5 LOCATE THE DISTURBANCE SOURCE FOR LOW-FREQUENCY OSCILLATION As well as incorrect parameters, unknown disturbances happening in the power grid can also cause inaccurate simulation results. Finding the disturbance is very important to the system operators, so that they can eliminate the disturbance source in time to avoid it happening again. Generally, a large disturbance can be detected and separated from the grid by the relays. However, it is sometimes very hard to ind and eliminate a small disturbance in a very large network. One of the most widely encountered problems caused by a small disturbance in a large power system is low-frequency oscillation. This is a long-standing problem. Theoretically, a small disturbance can excite low-frequency oscillation in two general ways. If the system is marginally stable or unstable, the small disturbance can push the operation states away from their equilibriums. In this case, the voltage, the current, and the power transferred in the grid begin to oscillate in increasing magnitudes or equal magnitudes, depending on whether the system equilibrium lies on the imaginary axis or the right half plane. This type of low-frequency oscillation is known as negative damping low-frequency oscillation, since the real part of the unstable equilibrium is greater than zero. This phenomenon was irst noticed during the 1960s after the wide application of fast high-gain exciters. These increase the power system transient stability by offsetting the power imbalance caused by the fault; however, they shift the eigenvalues of the linearized power system to the right half plane. The best way to quench the negative damping oscillation is to compensate the damping so that the total damping of the dynamic system becomes positive. The power system stabilizer (PSS) is designed as an auxiliary exciter control strategy to make up for system damping, and has been widely applied in the power grid worldwide. The second way in which a small disturbance excites the low-frequency oscillation of the whole power grid is resonance. Resonance-raised oscillation has become a serious problem with the increased interconnections of the power grids. As the power grid grows larger, the chances of being affected by the disturbance increase. On the other hand, the intrinsic oscillation frequency of the interconnected power system becomes smaller, which makes it very close to the oscillation frequency existing in the disturbance. Unlike negative damping low-frequency oscillation, resonant low-frequency oscillation appears and disappears in a very short time. It can reach its oscillation peak in several oscillation cycles, and constitutes a great danger to system security. The key to quenching resonant low-frequency oscillation lies in inding the oscillation source, that is, the disturbance that excites the resonant oscillation, and preventing the disturbance from happening again. However, it is not easy to ind an unknown disturbance in the large power grid. In practice, the digital simulation is performed repeatedly in a trial-and-error way. The pseudodisturbance is initiated at various locations of the power grid to match the simulation with the 326 Smart Grids PMU measurements. Due to the large scale of the power system and the number of electrical components involved in the simulation, great effort is required to search for the small disturbance. A self-diagnosing simulation could take the measurements from WAMs and apply hybrid data dynamic simulation to minimize the searching space for the disturbance. In the following, a real resonant oscillation case will be presented to show how modern measurement technologies and analysis tools are applied to minimize the computational burden of locating the disturbance source [11]. Figure 13.6 is a schematic diagram of the South Power Grid of China, interconnecting the provincial power grids GX, GZ, GD, and YN. It is a very large power grid, covering 1,000,000 km2, and nearly 2000 km long from east to west. Hydropower and fossil power are concentrated in the west of the South Power Grid; however, the load is mainly located in the east. Consequently, the electric energy has to be transferred from the west to the east through long transmission lines. Low-frequency oscillations tend to happen on these heavily loaded long transmission lines, which constitutes a major threat to the secure operation of the South Power Grid. One low-frequency oscillation happened in the South Power Grid of China in 2005. The oscillation spread quickly to the whole grid and affected the energy transfer on the interties. Based on the recorded dynamic trajectories, it can be seen that the low-frequency oscillation happened twice between 00:55:00 and 00:59:30. The irst oscillation started from 00:54:54 and ended at 00:57:37, lasting about 4 min. It took about 30 s for the oscillation to reach its peak. Although the oscillation peak-to-peak value reached 140 MW in the irst 30 s, the average power transfer remained nearly constant over this period. The oscillation amplitude of the active power remained unchanged. Meanwhile, the average power transfer increased during the period of the oscillation, which exacerbated the loading of this intertie. Nevertheless, it can be observed that the increased loading of the transmission line had few effects on GZ ASH Guiyang GX Qingyan Nayong GD GZH Anshun Beijiao Hechi Anshun Yantan Hezhou TSQ Pingguo TSQ 2 Luoping Shatang Luodong Wuzhou Lubuge YN TSQ 1 Baise ZHQ Nanning Xijiang Yulin Maoming Hydro power plant Power plant Jiangmen Converter station FIGURE 13.6 Schematic diagram of 500 kV network of South Power Grid, China. 327 Self-Tuning and Self-Diagnosing Simulation the oscillation. The oscillation continued until 00:56:48. Then it disappeared very quickly from 00:56:48 to 00:57:15, taking about 30 s, as it had begun. Meanwhile, the grid control center reduced the power generation in the west of the South Power Grid to alleviate the loading of the grid transfer interfaces. Even though the transferred energy was continually decreasing, the second low-frequency oscillation still happened. The second low-frequency oscillation took place from 00:57:40 to 00:59:30. During the oscillation, the average energy transferred continuously decreased. It also took about 30 s to reach its oscillation peak. The peak-to-peak value of the second low-frequency oscillation was about 100 MW, which was a little lower than that of the irst oscillation. Following this, the active power oscillated with equal amplitude, as had happened in the irst oscillation. The equal amplitude oscillation continued until 00:59:00. Then it quenched very quickly, also in 30 s, from 00:59:00 to 00:59:30. By linearizing the power system model and calculating its eigenvalues, we can ind the major electromechanic oscillation modes during this contingency, which are shown in Table 13.1. The oscillation frequency ranges from around 0.55 to around 1.21 Hz. It should be noted here that all these oscillation modes are very well damped. The analysis directly on the PMU measurements via the Prony method conirms the results derived from the linearized system model. The Prony analysis decomposes the signal into the summation of a series of exponential functions; thus the principal oscillation mode can be found. When the steady-state intertie power oscillation undergoes Prony analysis, the real measurements from PMU and the Prony analysis results in both the irst and the second oscillation it very well. Based on the Prony analysis, the principal frequency mode of the irst oscillation is 0.854 Hz and that of the second oscillation is 0.9117 Hz. They are slightly different, but both can be found in Table 13.1, which indicates that these two oscillation TABLE 13.1 Electromechanic Oscillation Modes of South Power Grid, China No. 1 2 3 4 5 6 7 8 9 10 11 12 Real Image Frequency (Hz) Damp −5.1679 −5.4668 −4.3498 −8.8456 −8.8883 −9.1937 −9.2359 −0.3420 −0.2911 −4.3851 −0.7818 −1.7864 3.4812 3.7609 4.3003 4.8572 5.0454 5.2298 5.3949 5.7577 5.9420 6.1588 7.0740 7.6212 0.5541 0.5986 0.6844 0.7730 0.8030 0.8323 0.8586 0.9164 0.9457 0.9802 1.1259 1.2130 0.8294 0.8239 0.7111 0.8766 0.8697 0.8692 0.8635 0.0593 0.0489 0.5800 0.1098 0.2282 328 Smart Grids modes are intrinsic oscillation modes of the South Power Grid, China. To sum up, this low-frequency oscillation has the following characteristics. First, it reaches its peak oscillation magnitude in a very short time and disappears in the same manner. The time duration is 30 s for both. This is typically different from negative damping oscillation, which usually takes a much longer time to reach its peak value or quench, as observed from several prior low-frequency oscillations in the same power grid. Second, there is no obvious oscillation start level for this low-frequency oscillation. Neither increased nor decreased average power transfer level on the transmission interties could quench the oscillation. In contrast, negative damping oscillation has, in general, an interchange power threshold. When the transferred power of the interties is above the threshold, low-frequency oscillation with increasing amplitude starts; on the other hand, if the transferred power of the interties is below this threshold, the oscillation disappears. To prevent low-frequency oscillation in the South Power Grid, China, PSS has been widely applied in this system; however, this time it was unable to stop the lowfrequency oscillation. Posterior simulation after the fault could not reproduce the oscillation. Since the oscillation modes are intrinsic to the South Power Grid, it is suspected that some unknown disturbance in the grid coincidentally has a frequency component close to the intrinsic oscillation frequency of the grid, which therefore excites forced low-frequency oscillation in the whole grid. Because the location is unknown, the disturbance cannot be properly simulated. Consequently, the simulation could not ind the root of the oscillation problem and effectively erase it. To ind the oscillation cause, the pseudodisturbance is assumed at different points in the South Power Grid. Due to the large scale of the grid, the computational burden is quite heavy. In addition, another problem closely related to locating the disturbance source by the trial-and-error method is that the simulation based on the CM data of the South Power Grid could possibly produce similar simulation curves even when the disturbance is assumed at different locations. For example, simulations based on the imaginary disturbance either in a power plant of GZ or in another power plant of GX produce nearly the same power oscillations in the Luoma intertie. So we cannot judge the location of the disturbance from the simulations alone. This aggravates the dificulty of inding the disturbance source. Let us take a two-area system to explain the idea of applying hybrid data dynamic simulation to locate the disturbance source for the forced low-frequency oscillation. Assuming there is a disturbance in Area 1, we could not judge where the disturbance lies from PMU measurements alone. But the PMU measurements could be used to decompose the original two-area system into two one-area subsystems by hybrid data dynamic simulation, as shown in Figures 13.7 and 13.8. The dashed line in Area 1 indicates that the disturbance is actually in Area 1 but is not simulated because we do not know where it is. Analyzing Figure 13.8, we ind that its simulation scenario is completely the same as for the two-area system, except that Area 1 in the original two-area system is replaced by the real measurements. Since the measurements relect the real dynamics of the power system and are irrelevant to the system conigurations in Area 1, the unknown disturbance in Area 1 does not affect the hybrid data dynamic simulation results of Figure 13.8. So, the hybrid data dynamic simulation on Figure 13.8 would 329 Self-Tuning and Self-Diagnosing Simulation Area 1 PMU FIGURE 13.7 Subsystem with Area 1 modeled in detail. PMU Area 2 FIGURE 13.8 Subsystem with Area 2 modeled in detail. be expected to reproduce the PMU measurements. This match between the simulation and the measurements indicates that the disturbance does not lie in Area 2. However, the scenario of the hybrid dynamic simulation in Figure 13.7 is different from the true scenario. In Figure 13.7, PMU measurements dynamically equalize the models and parameters of Area 2, but they cannot represent the unknown disturbance in Area 1. Due to the lack of simulation of the unknown disturbance in Area 1, the hybrid data dynamic simulation could not reproduce the PMU measurements, which indicates that the disturbance lies in this area. In this way, the searching range of the disturbance shrinks to Area 1 only, greatly reducing the computational burden. Since the PMU measurements are applied to decouple the very large system into small ones, the probability of very similar simulation results from different pseudodisturbance locations is also minimized. Self-diagnosing simulation to ind the disturbance source for the forced lowfrequency oscillation is applied to the South Power Grid, China. Due to data conidentiality agreements with the utility, simulated phasors will replace the PMU measurements in the following to illustrate the idea. However, the procedure is completely the same as when the real measurements are applied. Assume that a small disturbance in one power plant in the South Power Grid, China, has excited the resonance in the whole grid and we are trying to ind it from the pseudomeasurements. The South Power Grid is divided into four subsystems, as shown in Figure 13.6 by the dashed lines. Each subsystem is actually an administrative province in south China and is connected to the others by heavily loaded interties. The hybrid dynamic simulation can be applied to each subsystem as well as any combination of the subsystems. If the subsystem GZ is replaced by the measurements and we simulate the remaining three subsystems (GX, YN, and GD) together using the hybrid data dynamic simulation method, the measurements and the simulation results match each other very well. However, if the GX, GD, and YN subsystems are substituted by the measurements on the boundaries to GZ and we simulate the GZ power grid individually, the voltage measurement on 500 kV bus Anshun and the respective hybrid data dynamic simulation result will be quite different. Therefore, the disturbance must lie in the GZ power system. Then the searching space can be reduced to GZ power system, with the power grids outside GZ 330 Smart Grids equalized by the phasor injections. The hybrid data dynamic simulation is further applied to the individual power plant in GZ. The active power outputs of the two power plants in the inner power grid of GZ are shown in Figures 13.9 and 13.10, respectively, with the hybrid data dynamic simulation results. Since, in Figure 13.9, the measurements and the hybrid data dynamic simulation results match each other, the disturbance is not in the NY power plant. On the other Measurement Simulation 301.0 Active power (MW) 300.5 300.0 299.5 299.0 298.5 115 116 117 118 119 120 121 122 Time (s) FIGURE 13.9 Active power outputs of NY power plant in GZ power grid. 235 Measurement Simulation 230 Pe (MW) 225 220 215 210 205 200 115 116 117 118 119 120 Time (s) FIGURE 13.10 Active power outputs of QB power plant in GZ power grid. 121 122 Self-Tuning and Self-Diagnosing Simulation 331 hand, Figure 13.10 shows obvious differences between the measurements and the hybrid data dynamic simulations on the active power outputs in the QB power plant. It is, therefore, likely that the disturbance lies in the QB power plant. Then the hybrid data dynamic simulation is carried out in the QB power plant with a small disturbance initiated in its prime mover. In this way, the forced oscillation is reproduced by the hybrid data dynamic simulation, which veriies the disturbance location. 13.6 CONCLUSION Simulation has always played a very important role in power system operation and control. On some occasions it is the only tool for understanding the fault mechanisms of the power system, due to the large scale and complexity of the system as well as the security concerns. The traditional simulation is an open loop. With the wide application of modern and smart measurement technologies in the power grid, the simulation loop can be closed; in this way, simulation will assume even more responsibility for aiding system operation. The smart grid has revolutionized the concepts and the visions of the future power system. As one important block in building the smart grid, digital simulation will also make great advances in adaptability and lexibility. Together with fast-growing computation technologies, self-tuning and selfdiagnosing simulation is coming. ACKNOWLEDGMENTS The author would like to dedicate this chapter in memory of Professor Alain Germond of Ecole Polytechnique Federale de Lausanne, Switzerland, in thanks for all his encouragement and selless help to the author. This work is supported in part by the National Science Foundation of China (Grant No. 51077049 and No. 50707009), in part by the Beijing Nova Program, and in part by the ABB Research Center in China. Part of the research was carried out when the author was with North China Electric Power University, Beijing. The author would like to thank his postgraduate students in this research ield: Mr Sheng Wen-Jin, Mr Bo Bo, and Ms Zhang Pu. The author wants to express his hearty thanks to Professor Dong ZhaoYang at the University of Sydney, Australia, for his insightful suggestions and discussions. The author appreciates the discussions on smart grid development with Professor Rehtanz of the University of Dortmund and Dr Yue Cheng-Yan of the ABB Research Center in China, and would like to thank them for their support and suggestions. REFERENCES 1. D.N. Kosterev, C.W. Taylor, W.A. Mittelstadt, Model validation for the August 10, 1996 WSCC system outage, IEEE Transactions on Power Systems, 14(3), 967–979, 1999. 2. R.O. Burnett Jr, M.M. Butts, T.W. Cease, V. Centeno, G. Michel, R.J. Murphy, A.G. Phadke, Synchronized phasor measurements of a power system event, IEEE Transactions on Power Systems, 9(3), 1643–1650, 1994. 3. J. Hauer, D. Trudnowski, G. Rogers, B. Mittelstadt, W. Litzenberger, J. Johnson, Keeping an eye on power system dynamics, IEEE Computer Applications in Power, 10(4), 50–54, 1997. 332 Smart Grids 4. M. Zima, M. Larsson, P. Korba, C. Rehtanz, G. Andersson, Design aspects for wide-area monitoring and control systems, Proceedings of IEEE, 93(5), 980–996, 2005. 5. D. Kosterev, Hydro turbine-governor model validation in Paciic Northwest, IEEE Transactions on Power Systems, 19(2), 1144–1149, 2004. 6. Z. Huang, R.T. Guttromson, J.F. Hauer, Large-scale hybrid dynamic simulation employing ield measurements, in Proceedings of the IEEE Power Engineering Society (PES) General Meeting, June 5–10, 2004, Denver, CO. 7. Z. Huang, D. Kosterev, R.T. Guttromson, T. Nguyen, Model validation with hybrid dynamic simulation, in Proceedings of the IEEE Power Engineering Society (PES) General Meeting, June 18–22, 2006, Montreal, QC. 8. J. Ma, W.-J. Sheng, R.M. He, Research on WAMs-based power system dynamic simulation strategy, Automation of Electric Power Systems, 31(18), 11–15, 2007. 9. J. Ma, D. Han, W.-J. Sheng, R.M. He, C.Y. Yue, J. Zhang, Wide area measurements-based model validation and its application, IET Generation, Transmission and Distribution, 2(6), 906–916, 2008. 10. Shaw Power Technologies, Program Operation Manual: PSS/E-30, 2004. 11. J. Ma, P. Zhang, H.-J. Fu, B. Bo, Z.-Y. Dong, Application of phasor measurement unit on locating disturbance source for low-frequency oscillation, IEEE Transactions on Smart Grid, 1(3), 340–346, 2010. 14 A Consensus-Based Fully Distributed Load Management Algorithm for Smart Grid Yinliang Xu, Wei Zhang, and Wenxin Liu CONTENTS 14.1 Introduction .................................................................................................. 333 14.2 Consensus-Based Global Information Discovery Algorithm ....................... 334 14.2.1 Introduction to Average Consensus Theorem................................... 335 14.2.2 Consensus-Based Information Discovery Algorithm Design .......... 337 14.2.3 Convergence Speed of Information Discovery Algorithm ............... 339 14.2.4 Considerations during Implementation.............................................340 14.2.4.1 Distributed Synchronization ..............................................340 14.2.4.2 Plug and Play ..................................................................... 341 14.3 Applications .................................................................................................. 343 14.3.1 MAS-Based Load Management ....................................................... 343 14.3.2 Load Shedding ..................................................................................344 14.3.2.1 Normal Operating Conditions ........................................... 345 14.3.2.2 Reliability........................................................................... 347 14.3.3 Load Restoration ............................................................................... 350 14.4 Conclusion .................................................................................................... 356 References .............................................................................................................. 356 14.1 INTRODUCTION Global concern over greenhouse gas production, increasing demand for better reliability and security, widespread use of intermittent energy resources, charging of plug-in hybrid electric vehicles, and so on, has led to new requirements being proposed for the power grid. However, the aging power grid, designed according to decades-old techniques, can hardly address these new requirements. Driven by these new requirements, the concept of the smart grid was proposed to upgrade current power grids for the next century. The smart grid is the electric delivery network integrated with the latest advances in digital and 333 334 Smart Grids information technology to improve the electric system’s reliability, eficiency, security, and resiliency. The power system is probably the most complex engineering system ever to have existed. In [1–3], various kinds of centralized methods were proposed to deal with problems such as load shedding, load restoration, optimal power low, and so on. It is known that a centralized control scheme requires a complicated communication network to collect global operating conditions and a powerful central controller to process the huge amount of information. Thus, centralized schemes are costly to implement and susceptible to single-point failures. Due to the uncertainty of intermittent renewable energy resources, unintentional changes of system structure will further increase the burden on a centralized control scheme. Since distributed control solutions are lexible, reliable, and low cost to implement [4,5], they are a promising choice for the next generation of power systems. As one of the most popular distributed control solutions, multiagent systems (MAS) have been widely applied for the operation and control of power systems. Advantages of MAS include the ability to survive single-point failures [6] and decentralized data processing, which leads to eficient task distribution and eventually a faster operation and decision-making process [7]. In artiicial intelligence, agent-based technology has been hailed as a promising paradigm for conceptualizing, designing, and implementing software systems. Recently, MAS-based solutions have been proposed for power system reconiguration and restoration [5,8,9]. However, it is still dificult to apply existing algorithms to systems with complex structure. Most importantly, most methods were validated through simulation instead of rigorous analysis. The convergence and stability of these algorithms have not been suficiently investigated. To address the needs of the smart grid and the problems with existing research, this chapter introduces a stable, fully distributed, multiagent-based solution. “Fully distributed” means that each bus in a power system has an associated bus agent, and two agents communicate with each other only if their corresponding buses are electrically coupled. The design of the distributed algorithm is based on a consensus-based stable global information discovery algorithm. The information discovery algorithm can guarantee convergence for power grids of any size and topology, and is robust against various faults. The distributed optimization algorithm has been successfully applied to load management (load shedding [10] and load restoration [11]). The information discovery algorithm is used to discover net active power and load connection information. Based on the information, centralized optimization algorithms, such as greedy algorithms and dynamic programming (DP), can be used to maximize energy utilization. The rest of the chapter is organized as follows. In Section 14.2, a consensus-based global information discovery algorithm is introduced. In Section 14.3, the proposed algorithm is applied to load management. Section 14.4 concludes this chapter. 14.2 CONSENSUS-BASED GLOBAL INFORMATION DISCOVERY ALGORITHM According to the fully distributed solution, calculation and distribution of global information are achieved through an average-consensus-based information discovery A Consensus-Based Fully Distributed Load Management Algorithm 335 algorithm. For convenience, the average consensus theorem is briely introduced as follows. 14.2.1 INTRODUCTION TO AVERAGE CONSENSUS THEOREM The average consensus theorem relies on the local information of agents to guarantee that important information is shared in a distributed way [12]. Recently, the consensus algorithm has attracted great interest in various areas; a recent survey can be found in [13]. According to the consensus theory, the overall information discovery process is modeled as a discrete time-linear system, as in Equation 14.1. X k +1 = DX k (14.1) where Xk and Xk + 1 are the vectors of discovered information at the kth and (k + 1)th iterations, respectively, and D is a sparse iteration matrix. ∑ 1 − a1 j j∈N1 ⋯ D = ai1 ⋯ an1 ⋯ a1i ⋯ ⋯ ⋯ ⋯ ⋯ 1− ∑a ⋯ ij j∈Ni ⋯ ⋯ ⋯ ⋯ ani ⋯ ⋯ ain ⋯ 1− anj j∈N n a1n ∑ If all elements of D are nonnegative and the sums of each row and column are all ones, that is, D*e = e and DT*e = e, with e = [1, 1, …, 1]T, then D is called a doubly stochastic matrix [14]. According to the theorem of Gerschgorin’s Disks [15], it can be concluded that all eigenvalues of D are ≤1. Based on the Perron–Frobenius lemma [15], we have J = lim D k = k →∞ eeT n (14.2) where n is the dimension of D. According to Equations 14.1 and 14.2, the system will reach consensus as k approaches ininity, as represented in Equation 14.3. lim X k = lim D k X 0 = k →∞ k →∞ eeT 0 X n (14.3) From Equation 14.3, one can see that the speed of the consensus-based information discovery algorithm is determined by D, and the coeficients aij need to be 336 Smart Grids properly designed to guarantee stability. For bilateral networks such as power systems, D can be designed to be symmetrical, like Ybus and Zbus, that is, aij = aji. For convergence analysis, a positive deinite Lyapunov function is deined as follows: T ( ) Vk = Xk ∆V = V k +1 − V k = ⎡ X k ⎣⎢ T ( ) (D D − I ) X T n ≤− ∑∑( aij + aij2 )( x kj − xik ) i =1 j∈Ni 2 + (14.4) ⎤ ⎦⎥ k 1 2 Xk n ∑ ∑{( n + n − 2 ) a ( x i i =1 j∈Ni n ≤− ∑ ∑ a ⎡⎣1 − ( n + n ) a ij i ij j i =1 j∈Ni 2 ij j 2 ⎤⎦ x kj − xik ( ) 2 k i − x kj )} 2 (14.5) where ni and nj are the numbers of agents in the neighborhood of agents i and j, respectively. From Equation 14.5, it can be seen that ΔV ≤ 0, provided that aij satisies the following condition: 0 < aij < 2 ni + n j (14.6) Since it may take too long to reach the exact equilibrium, it is necessary to deine a terminating criterion. For this purpose, error at iteration k can be deined as in Equation 14.7 [16]: E k = X k − X ∞ = Dk − J X 0 ≤ D − J ( ) k X 0 − JX 0 (14.7) where Ek is the error at the kth step. If Ek satisies a predeined precision requirement, it can be concluded that a consensus has been reached and the number of steps needed for convergence is approximately determined according to Equation 14.8 [16]. K= log1/ E 1 1 D−J ( ) = −1 log E 1 λ 2 ( ) (14.8) where λ2 is the second largest eigenvalue of D and E is the error tolerance. Equation 14.8 indicates that λ2 decides the converging speed and gives the asymptotic number of steps for the error to decrease by a factor of E. Thus, λ2 can be used for evaluation of algorithm speed and comparison of algorithms. As will be introduced later, the optimal converging speed can be obtained by designing a D with minimum λ2. A Consensus-Based Fully Distributed Load Management Algorithm 337 Since only the second largest eigenvalue is of interest to us, power method with delation can be used to calculate the second largest eigenvalue directly. During the process, calculation of the other eigenvalues is avoided. For convenience, the delation method is introduced as follows. The n eigenvalues of D can be arranged as 1, λ2, …, λn, in decreasing order of magnitude. One is the largest eigenvalue of D, and e is the corresponding eigenvector. Now consider the following matrix: D1 = D − λ1 e1e1T e1e1T D = D−J = − e1T e1 n (14.9) The eigenvectors of a symmetrical matrix are orthogonal, that is, eiT e j = 0, for i > 1, D1ei = Dei − λ1 e1e1T ei = Dei e1T e1 (14.10) Therefore, the eigenvalues of (D−J) are λ2, …, λn, 0. Now λ2 becomes the largest eigenvalue of (D−J) and can be calculated directly according to the power method [17]. 14.2.2 CONSENSUS-BASED INFORMATION DISCOVERY ALGORITHM DESIGN In this chapter, it is assumed that each bus in a microgrid is assigned a node agent (NA). Each NA knows local generation and load information, but does not have direct access to global information. An agent can only communicate with its neighboring agents if they are connected through distribution lines during normal operating conditions. The consensus algorithm is used to design the communication law for global information discovery. The information discovery process for agent i is represented as Equation 14.11. xik +1 = xik + ∑a (x ij k j − xik ) (14.11) j∈Ni where: i n xi0 xik and xik +1 aij Ni = 1, 2, …, n is the number of agents (buses) in the microgrid is the predicted local net power of agent i are the information discovered (average net power or average agent index) by agent i at iteration k and k + 1, respectively is the coeficient for information exchanged between neighboring agents i and j (0 < aij < 1 if agents i and j are connected through a distribution line, otherwise, aij = 0) is the number of indexes of agents that are connected to agent i 338 Smart Grids The global information discovery process of the whole microgrid can be represented using matrix format, as in Equation 14.12. X k +1 = X k + A ∗ X k = ( I + A ) X k = DX k (14.12) where: X k = [ x1k , …, xik , … , xnk ]T, I is an identity matrix. By making aij = aji, the number of independent coeficients in D is equal to the number of distribution lines in the microgrid. To achieve good converging speed, the independent variables aij need to be properly designed. For practical power systems, the iteration matrix D is usually a sparse matrix. According to Equation 14.8, by minimizing |λ2|, the optimal coeficient setting can be obtained [17]. If the system seldom suffers from loss of distribution lines, the coeficients can be optimized off-line. However, since off-line optimization requires global information and is time-consuming, the method is not suitable for online applications during changes of system coniguration. Thus, there is a need for algorithms that can adjust the coeficients automatically and achieve near-optimal performance. The Uniform and Metropolis method is used to set the coeficients [16]. An improved algorithm called Mean Metropolis is proposed, which is inspired by the Lyapunov stability analysis. Similarly to the Metropolis method, the new algorithm has the properties of stability and adaptivity. The improvement is in the near-optimal converging speed, as will be demonstrated later through simulation studies. The updating rule of the method is represented in Equation 14.13. 2 ( ni + n j + ε ) aij = 1 − 2 ( ni + n j + ε ) j∈Ni 0 ∑ j ∈ Ni i= j (14.13) otherwise where ɛ is a very small number (for large complex systems, ɛ can be set to zero). The above information discovery algorithm can be applied to calculate Pnet indirectly. If xik in Equation 14.11 is initialized with local net power (PGi−PLi), where PGi and PLi are the local active power generation and load, respectively, of agent i, xik will converge to average total net power through information exchange. However, the numerator, that is, the total net power, is the value that is going to be used. Thus, the denominator n, which is the number of agents participating in the information exchange, needs to be determined. According to the proposed algorithm, n is discovered together with the discovery of indexes of loads that are connected in the power system. An agent in the power system is assigned a unique index beforehand. Figure 14.1 illustrates the index discovery process during normal operating conditions. Initially, agent i’s N-dimensional index vector vi only has one nonzero element at (i, 1) with the value of i. By applying the above information discovery algorithm to the same elements of the index vectors, the ith elements of the index vectors will converge to i/N. 339 A Consensus-Based Fully Distributed Load Management Algorithm Index: 1 2 1 0 i 0 N 0 1 1/N 2 1/N i 1/N N 1/N 0 2 0 0 2/N 2/N 2/N 2/N 0 0 i 0 i/N i/N i/N i/N 0 0 0 N N/N N/N N/N N/N Initial index vectors Converged index vectors FIGURE 14.1 Initial and inal index vectors during normal operating conditions. 14.2.3 CONVERGENCE SPEED OF INFORMATION DISCOVERY ALGORITHM The proposed information discovery algorithm is independent of system coniguration and size. Theoretically, the algorithm can be applied to large-scale power systems. The authors limit the application to microgrids because of the following considerations. First, microgrids are smaller and distribution lines are shorter. Thus, the algorithm has a better chance to provide real-time solutions, depending on how “real time” is deined. Second, because of the intermittent renewable energy resources, the need for load shedding and restoration is more frequent, especially when microgrids are operating in the islanded mode. The proposed algorithm is tested with the 162-bus power systems to investigate the possibility of applications to large-scale power systems. The coeficients are designed based on the Mean Metropolis method. The second largest eigenvalue of the corresponding iteration matrices D for the 162-bus systems is 0.9792. According to Equation 14.8, the estimated number of steps needed for achieving consensus with tolerance 0.03 is 165, which is supported by Figure 14.2. According to Equation 14.8, one can see that the speed of information discovery process is decided not by the size of the system, but by how the buses are connected and how the coeficients are decided. However, Equation 14.8 is not accurate enough, because some assumptions had to be made during the analysis. For example, the estimated number of steps (71 steps) is less accurate for the 16-bus system. This is Average net power 50 0 –50 0 20 40 60 80 100 120 Number of iterations 140 160 180 FIGURE 14.2 Average total net power discovery process of the IEEE 162-bus system. 200 340 Smart Grids because λ2 is not large enough, such that the remaining eigenvalues have more impact on the convergence. From the algorithm design, one can see that the converging speed of the information discovery algorithm is independent of the initial values. This property can also be observed from Equation 14.8. For veriication, the local net power of each agent is set to a random value as opposed to the rated value. Simulation results show that the converging speed of the system with random loads is the same as that with rated loads. This property makes converging speed analysis and implementation easy. The speed of the information discovery algorithm is evaluated using number of iterations instead of time. Even though the actual time is decided by speciic implementation of hardware and software, it is preferable to make a rough estimation. By neglecting the time used for information update, the time needed for information discovery can be estimated according to the following equation: T= na ∗ ( 3 ∗ nb ) ∗ nc nd (14.14) where: na is the number of iterations needed for convergence (3*nb) is the size of the information matrix nc is the number of bits used to represent each element of the information matrix nd is the communication speed in terms of bits per second For example, assume that the information discovery process of the Institute of Electrical and Electronics Engineers (IEEE) 162-bus system requires 160 iterations to converge and each element of the information matrix is represented using 16 bits. For a network speed of 10 Mbit/s, the time used for information discovery can be calculated as 165(3 × 162)16/10,000,000 = 0.1283 s. Thus, even if the algorithm is not fast enough for real-time control, it is suficient for most planning and optimization problems in large-scale power systems. It should be noted that the above estimation is not accurate, but it can give a rough idea of the speed of the information discovery process. With the development of network techniques, the algorithm can be implemented faster and faster, so that application to large-scale power systems will no longer be a problem. 14.2.4 CONSIDERATIONS DURING IMPLEMENTATION 14.2.4.1 Distributed Synchronization It should be noted that synchronization of autonomous agents is a challenging problem. For a fully distributed MAS solution, no centralized coordinative mechanism can be assumed, not even a common clock signal. The synchronization has to be accomplished through communication between distributed agents. Thus, it is impossible to rigorously synchronize the activities of autonomous agents in terms of time. Due to the randomness of autonomous activities, the speed of the presented algorithm is evaluated using number of iterations. Below are the preliminary solutions for three A Consensus-Based Fully Distributed Load Management Algorithm 341 major synchronization problems: coordinating the information update of neighboring agents, terminating information discovery for decision making, and terminating decision making for information discovery. To synchronize the information updates of agents, each agent maintains an index for its current step of information discovery. Two neighboring agents will exchange and update information only if their indexes are the same. If their indexes are not the same, the agent with the larger index will wait till the agent with the smaller index catches up. After a successful information update, both agents increase their indexes by one. In this way, the information exchange and update activities of autonomous agents can be synchronized. Due to unexpected changes of operating conditions, the required number of iterations for a successful information discovery is unpredictable. Thus, the upper limit of the index can be set to a very large value. Alternatively, to save memory requirement, a smaller upper limit can be used and the index will be reset to one once it hits the upper limit. Since the information discovery instances of agents may converge at different times, there is a need to terminate the overall MAS’s information discovery process and start the load-shedding process as soon as possible. Similarly, the information discovery process needs to be restarted once a load management decision has been deployed. The above operations can be accomplished by distributing signs of convergence and decision making, respectively. Once agent i concludes that it has found the global information based on Equation 14.11, the agent will set its sign of convergence to one and propagate the nonzero value to other agents. If agent j notices that the ith element of the convergence status vector is nonzero, agent j knows that the information discovery process of agent i has converged. Finally, if an agent notices that all elements of the convergence status vector are nonzero, it concludes that the overall system’s information discovery process has converged. The agent will start the load-shedding decision-making process immediately. The last converged agent will make the load-shedding decision based on the discovered global information, deploy its corresponding local decision, and set its sign of decision making to one. Once an agent notices a nonzero value in the decisionmaking vector, it will set its own sign of decision making to one while making and deploying the load-shedding decision. The process proceeds until an agent inds that all elements in the decision-making vector are nonzero. After deploying its local load-shedding decision, the agent will restart the information discovery process by initializing the convergence status vector to all zeros and the index for information exchange to one. 14.2.4.2 Plug and Play The proposed algorithm can be applied to power systems with capability for plug-and-play operation. If a new generator or load is attached to an existing agent, only local net power needs to be updated. Thus, some redundant buses and agents can be designed for the purpose of plug-and-play operations. Since no new agent is involved during plug-in operation, no ID needs to be dynamically assigned. Similarly, IDs of existing agents do not need to be updated during plug-outs. 342 Smart Grids However, special attention is required during changes of network topology. If a new bus is added to a power system, the corresponding new agent will require a unique ID to interact with existing agents. If an existing agent is removed, indexes of the remaining agents may need to be updated to simplify the plugging-in of new agents. The agent ID acquisition and updating process is introduced in the following paragraphs. First, let us consider the scenario of a new agent/bus being plugged in. The new agent will start with a zero vector to participate in the information discovery with other agents. The size of the zero vector is the same as the sizes of existing index vectors, which are larger than N to accommodate future network expansion. For simpliication, only useful information is illustrated in the following igures. Since N + 1 agents participate in information discovery, the ith element of the index vector will converge to i/(N + 1) instead of i/N as before. Based on the converged vector, the new agent will know that its ID should be N + 1. The ID discovery process of the new agent is illustrated in Figure 14.3, and the information discovery process after ID acquisition is shown in Figure 14.4. Second, let us consider the scenario in which agent i was disconnected from the system. To simplify the future ID acquisition process, the IDs of the remaining agents may need to be updated. The detection of agent disconnection and the updated information discovery processes are illustrated in Figures 14.5 and 14.6, respectively. 1 2 1 0 i 0 New N agent 0 0 0 2 0 0 0 0 i 0 0 0 New agent 1 1/(N + 1) 2 1/(N + 1) i 1/(N + 1) N 1/(N + 1) 1/(N + 1) 0 2/(N + 1) 2/(N + 1) 2/(N + 1) 2/(N + 1) 2/(N + 1) 0 0 i/(N + 1) i/(N + 1) i/(N + 1) i/(N + 1) i/(N + 1) N 0 N/(N + 1) N/(N + 1) N/(N + 1) N/(N + 1) N/(N + 1) FIGURE 14.3 ID discovery process of the new plugged-in agent. 1 2 i N N+1 1 2 i N N+1 1 0 0 0 0 1/(N + 1) 1/(N + 1) 1/(N + 1) 1/(N + 1) 1/(N + 1) 0 2 0 0 0 2/(N + 1) 2/(N + 1) 2/(N + 1) 2/(N + 1) 2/(N + 1) 0 0 i 0 0 i/(N + 1) i/(N + 1) i/(N + 1) i/(N + 1) i/(N + 1) 0 0 0 N 0 N/(N + 1) N/(N + 1) N/(N + 1) N/(N + 1) 0 0 0 0 N+1 (N + 1)/(N + 1) (N + 1)/(N + 1) N/(N + 1) (N + 1)/(N + 1) (N + 1)/(N + 1) (N + 1)/(N + 1) FIGURE 14.4 Information discovery process after ID acquisition. 343 A Consensus-Based Fully Distributed Load Management Algorithm 1 2 i–1 i+1 1 0 0 0 N 1 2 i–1 i+1 N 0 1/(N – 1) 1/(N – 1) 1/(N – 1) 1/(N – 1) 1/(N – 1) 0 0 i–1 0 0 0 0 0 0 0 0 0 0 i+1 0 0 0 0 0 N (i – 1)/(N – 1) (i – 1)/(N – 1) 0 (i + 1)/(N – 1) (i + 1)/(N – 1) N/(N – 1) (i – 1)/(N – 1) (i – 1)/(N – 1) 0 N/(N – 1) 0 0 (i + 1)/(N – 1) (i + 1)/(N – 1) N/(N – 1) (i – 1)/(N – 1) 0 (i + 1)/(N – 1) N/(N – 1) N/(N – 1) i+1 N FIGURE 14.5 Detection of the disconnection of agent i. i+1 N N–1 1 2 i–1 i N–1 1 0 0 0 0 1/(N – 1) 1/(N – 1) 1/(N – 1) 1/(N – 1) 1/(N – 1) 0 2 0 0 0 2/(N – 1) 2/(N – 1) 2/(N – 1) 2/(N – 1) 2/(N – 1) i–1 0 0 1 2 0 0 i–1 i 0 0 0 i 0 0 0 0 0 N–1 (i – 1)/(N – 1) (i – 1)/(N – 1) i/(N – 1) i/(N – 1) (N – 1)/(N – 1) (N – 1)/(N – 1) (i – 1)/(N – 1) (i – 1)/(N – 1) i/(N – 1) i/(N – 1) (N – 1)/(N – 1) (N – 1)/(N – 1) (i – 1)/(N – 1) i/(N – 1) (N – 1)/(N – 1) FIGURE 14.6 Updated information discovery process. 14.3 14.3.1 APPLICATIONS MAS-BASED LOAD MANAGEMENT The introduced algorithm is implemented using NAs, the number of which is equal to the number of buses. The function modules in an NA and the MAS-based load management operation are illustrated in Figure 14.7. In Figure 14.7, the initialization module will initialize local information vectors based on the operating condition of the corresponding load. The information exchange module will exchange the current information vectors with its neighboring agents and provide required information to the information update module. The information update module will update aij during change of conigurations and keep updating the information vector till it converges. The converged information vector is then sent to the information analysis module to obtain global operating conditions (global net power and operating conditions of loads) for load management decision making. A negative total net power means that the microgrid cannot support the load demand, and load shedding is required to avoid further damage. On the other hand, a positive average total net power means that the microgrid may be able to satisfy the demands of some unfaulted but out-of-service loads. In that case, the load restoration process will be activated. 344 Smart Grids Load controller Load management decision Node agent i Loads selection Information analysis System operating condition Initialization Consensus xi∞ = xi0 Neighboring agents j j ∈Ni xik+1 xjk, j ∈Ni Information exchange 1 n n ∑ xi0 i=1 xik+1 Information update xjk, j ∈Ni xik+1 = xik +∑ aij xik–xik n j=1 ( ) FIGURE 14.7 Function of NA in MAS load management. The optimal load management decision is then deployed by load controllers. An optimal load management decision can be obtained by satisfying the following two constraints: • Maximizing the amount of loads while ensuring loads with higher priorities are served irst, that is, N max ∑P ⋅P Li i (14.15) i =1 where Pi is the priority index associated with the ith load. • Loads to be served must be less than or equal to excessive generation, that is, N ∑ i =1 N PGi − ∑P Li ≥0 (14.16) i =1 Identifying the loads to be served can be modeled as a 0–1 knapsack problem, which is an NP-hard combinatorial optimization problem. DP is one of the commonly used techniques for this kind of problem [18]. For large-scale systems, it is very dificult to ind the optimal solution within a limited time. Greedy algorithms, such as greedy by proit, greedy by weight, and greedy by proit density, can be applied to reduce the computational burden [19]. According to these greedy algorithms, items with higher proits, larger weights, and larger ratio of proit over weight have higher priorities. Greedy algorithms are usually faster than DP but cannot guarantee optimal solutions. Performances of these algorithms will be compared through an example in this section. 14.3.2 LOAD SHEDDING Faults, sudden dramatic load change, and insuficient generation can create power mismatch between generation and loads. If generation in power systems is insuficient A Consensus-Based Fully Distributed Load Management Algorithm 4 2 3 5 d23 d24 G d35 d12 345 G d13 1 FIGURE 14.8 IEEE 5-bus power system. TABLE 14.1 Data of the IEEE 5-Bus System Neighbors PGi PLi X i0 Priority 4 2 100 0 100 — 5 3 150 0 150 — 3 1, 2, 5 0 150 −150 P1 1 2, 3 0 50 −50 P2 2 1, 3, 4 0 100 −100 P2 No. to power all loads, eficient load-shedding operations may need to be deployed to maintain the supply–demand balance. Load shedding is the process of tripping a certain amount of load with lower priority to maintain the stability of the rest of the system. The proposed MAS-based load-shedding algorithm [10] is tested with the IEEE 5-bus shown in Figure 14.8. The original data on system conigurations are modiied and summarized in Table 14.1. 14.3.2.1 Normal Operating Conditions The ive independent variables a12, a13, a23, a24, and a35 are the coeficients for information exchange between neighboring agents. Eigenvalues of D after optimization are summarized in Table 14.2. The net power discovery process is shown in Figure 14.9. From the igure, it can be seen that the agents will reach consensus within 10 iterations. From Figure 14.9, it can be seen that all agents’ discovered information converged to the same value of −10, which is the average net power of the system. To ind the total net power in the system, the number of agents that are currently connected has to be found. The process is illustrated in Table 14.3. During the process, every agent is initialized with a 5 × 1 index column vector. For agent i, the initial index vector only has one nonzero element at (i, 1) with a value of i. Since there is no agent/bus loss, the ith elements of all index vectors will converge to i/5. According to the converged index vector, each agent can ind out that 346 Smart Grids TABLE 14.2 Independent Coeficients and Eigenvalues after Optimization Independent Coeficients Eigenvalues of D 1/3 1/3 1/3 1/2 1/2 a12 a13 a23 a24 a35 1.0000 0.7071 0.4082 λ1 λ2 λ3 λ4 λ5 −0.4082 −0.7071 150 Agent 1 Agent 2 Agent 3 Agent 4 Agent 5 Net power 100 50 0 –50 –100 –150 0 2 4 6 8 10 12 14 16 18 20 Number of iterations FIGURE 14.9 Average total net power discovery process of the IEEE 5-bus system. the number of agents in the system is ive and their indexes are 1–5. Thus, the total net power can be calculated as −10 × 5 = −50. Now all necessary information for load shedding has been obtained. In Table 14.1, the three loads are classiied into two priority levels, with P1 > P2 in terms of priority. In the same priority level, the loads are sorted according to the amount of local net power. The results of different load-shedding algorithms are shown in Table 14.4. Compared with DP, the greedy algorithms may not be able to guarantee the best energy utilization but are more computationally eficient. Thus, DP is a better choice for smaller systems and greedy algorithms are a better choice for large-scale systems. Since all agents maintain the same load table and use the same load-shedding algorithm, the load-shedding decisions of distributed agents will be the same and the load-shedding activities of the overall system will be coordinated. The information discovery process after load 1 has been disconnected is shown in Figure 14.10. From Figure 14.10, one can see that the initial local net power of Agent 1 is set to zero because its corresponding load has been disconnected. Since the newly discovered net power is 0, no additional load shedding is necessary. 347 A Consensus-Based Fully Distributed Load Management Algorithm TABLE 14.3 Discovery of Number of Agents in Power System Agent Initial Index Vector Converged Index Vector Number of Agents 1 2 3 4 5 [1, 0, 0, 0, 0]T [0, 2, 0, 0, 0]T [0, 0, 3, 0, 0]T [0, 0, 0, 4, 0]T [0, 0, 0, 0, 5]T [0.2, 0.4, 0.6, 0.8, 1.0]T [0.2, 0.4, 0.6, 0.8, 1.0]T [0.2, 0.4, 0.6, 0.8, 1.0]T [0.2, 0.4, 0.6, 0.8, 1.0]T [0.2, 0.4, 0.6, 0.8, 1.0]T 5 5 5 5 5 TABLE 14.4 Comparison of Different Algorithms for Load Shedding Algorithm Loads To Be Shed Dynamic programming Greedy by proit Greedy by weight Greedy by proit density 1(−50) 1(−50) 1(−50) 2(−100) 150 Agent 1 Agent 2 100 Net power Agent 3 50 Agent 4 Agent 5 0 –50 –100 –150 0 2 4 6 8 10 12 Number of iterations 14 16 18 20 FIGURE 14.10 Average total net power discovery process after load 1 is disconnected. 14.3.2.2 Reliability To test the robustness of the proposed algorithm, the three operating conditions analyzed during robustness analysis are investigated, which are loss of distribution line, disconnection of load, and loss of agent during an uninished information discovery process. First, consider the loss of distribution line between Buses 1 and 2, which can be represented as a12 = 0. Within 16 steps, the agents will reach the same consensus 348 Smart Grids as before. Since the coeficients are no longer optimal, the information discovery process takes more iterations to converge. This example proves that the algorithm is robust to loss of distribution lines. The information discovery process for this case is shown in Figure 14.11. Second, assume the load that was originally connected to Agent 2 is disconnected. In response to the change of operating condition, Agent 2 reinitializes local net power to 0 and index vector to 0 s in all cases. The information discovery processes of average net power and indexes of connected loads are illustrated in Figures 14.12 and 14.13, respectively. From Figure 14.12, it can be seen that Agent 2 still works in the same way as before, even though its associated load is disconnected. From Figure 14.13, one can see that all agents except Agent 2 will converge to the same value (i/N) as before. Since i is the predeined index of agent i, the number of agents that are currently 150 Agent 1 Agent 2 100 Net power Agent 3 50 Agent 4 Agent 5 0 –50 –100 –150 0 FIGURE 14.11 2 4 6 8 10 12 Number of iterations 14 16 20 Average total net power discovery process after a distribution line loss. 150 Agent 1 Agent 2 Agent 3 Agent 4 Agent 5 100 Net power 18 50 0 –50 –100 –150 0 2 4 6 8 10 12 Number of iterations 14 16 18 FIGURE 14.12 Average total net power discovery process after load 2 is disconnected. 20 349 A Consensus-Based Fully Distributed Load Management Algorithm 5 1st row 2nd row 3rd row 4th row 5th row Index 4 3 2 1 0 0 2 4 6 8 10 12 Number of iterations 14 16 18 20 FIGURE 14.13 Network structure information discovery process after load 2 is disconnected. connected in the power system can be easily determined as ive. By checking the positions of the zeros, all agents know that the load on Bus 2 is now disconnected. The obtained information on total net power and load connection is suficient for distributed load shedding. Last, consider the loss of Agent 4 during an ongoing information discovery process. Since the working agents do not know that a reinitialization is necessary, they will continue the uninished information discovery process until convergence. By inspecting the converged index vectors, the agents can easily identify the fault and restart the information discovery process with updated information. The above process can be observed in Figures 14.14 and 14.15. It should be noted that, the earlier a fault happens during the information discovery process, the larger is its impact on accuracy. If the fault happens approaching the end of the information discovery process, its impact may be negligible. 150 Agent 1 Agent 2 Agent 3 Agent 4 Agent 5 Net power 100 50 0 –50 –100 –150 0 5 10 15 Number of iterations 20 FIGURE 14.14 Average total net power discovery process of agent loss. 25 30 350 Smart Grids 5 1st row 2nd row 3rd row 4th row 5th row Index 4 3 2 1 0 0 5 10 15 Number of iterations 20 25 30 FIGURE 14.15 Network structure information discovery process of agent loss. 14.3.3 LOAD RESTORATION When a fault occurs in a microgrid system, corresponding circuit breakers will be opened to isolate the fault so as to protect the power equipment. During the process, some unfaulted loads are forced to be out of service. Restoring power supply to these de-energized but unfaulted loads as soon as possible is of great importance. Eficient restoration methods are very important for energy use and service quality. Restoration may require the investigation of massive combinations of switching operations that satisfy multiple objectives under certain constraints. The proposed MAS-based load restoration algorithm [11] is now applied to the 16-bus system [18] shown in Figure 14.16. Data are modiied as shown in Table 14.5. In Table 14.4, PGi is the local generation, PLi is the local load, Xi0 is the local net power, and P1–2 is the priority of the loads, with P1 > P2. 2 3 4 10 1 5 11 6 7 8 9 12 14 15 Connected Disconnected FIGURE 14.16 The 16-bus test power system. 13 16 351 A Consensus-Based Fully Distributed Load Management Algorithm TABLE 14.5 Data of the 16-Bus System Neighbors PGi PLi Xi0 Priority 1 2 3 4 5 6 7 8 9 10 11 12 2, 5 1, 3, 5, 10 2, 10 10, 11 1, 2, 6 5, 7, 9 6, 8, 9 7, 9 6, 7, 8, 11, 15 2, 3, 4 4, 9, 12 11, 13 40 40 50 0 0 0 0 0 0 0 40 0 0 0 0 20 35 20 15 25 5 30 0 45 40 40 50 −25 −35 −20 −15 −25 −5 −30 40 −45 — — — P2 P2 P1 P2 P1 P1 P2 — P2 13 14 15 16 12, 14 13, 16 9 14 0 0 30 50 10 40 0 0 −10 −40 30 50 P2 P2 — — No. 2 3 4 Fault 10 1 5 11 6 7 8 9 12 14 15 Connected 13 16 Disconnected FIGURE 14.17 Postfault coniguration of the 16-bus system. Assume that a sudden fault occurs to the generator on Bus 11. The protection system reacts immediately to isolate the fault by opening ive switches: S2,10, S3,10, S5,6, S9,15, and S12,13 [20]. The postfault system after protective action is illustrated in Figure 14.17. For the postfault system, the generator at Bus 11 cannot be restored immediately and the eight loads attached to Agents 4 and 6–12 should be restored as soon as possible. The initial information matrices of the 16 agents and the converged matrix are shown in Table 14.6. 352 TABLE 14.6 Initial and Converged Information Matrices Initial Matrices M T 0 0 0 ⎡0 ⎢ ⎢0 ⎢⎣0 0 5 −35 0⎤ ⎥ 0⎥ 0 ⎥⎦ ⎡0 ⎢ ⎢0 ⎢⎣0 0 9 −5 ⎡0 ⎢ ⎢0 ⎢⎣0 13 0 −10 0⎤ ⎥ 0⎥ 0 ⎥⎦ T 5 T 0 0 0 2 ⎡0 ⎢ ⎢0 ⎢0 ⎣ 0 6 −20 ⎡0 ⎢ ⎢0 ⎢⎣0 9 T 0⎤ ⎥ 0⎥ 0 ⎥⎦13 T 2 0 40 ⎡0 ⎢ ⎢0 ⎢⎣0 0 10 −30 14 0 −40 0 0 0 0 0 0 T 0⎤ ⎥ 0⎥ 0 ⎥⎦ 6 0⎤ ⎥ 0⎥ 0 ⎥⎦ T ⎡0 ⎢ ⎢0 ⎢0 ⎣ ⎡0 ⎢ ⎢0 ⎢ ⎣0 10 T 0⎤ ⎥ 0⎥ 0 ⎥⎦14 T ⎡0 ⎢ ⎢0 ⎢0 ⎣ 0 0 0 0 0 0 3 3 0 50 T 0⎤ ⎥ 0⎥ 0 ⎥⎦ 7 0 7 −15 T 15 0 30 ⎡0 ⎢ ⎢0 ⎢0 ⎣ ⎡0 ⎢ ⎢0 ⎢0 ⎣ 0⎤ ⎥ 0⎥ 0 ⎥⎦11 0 0 0 T 0⎤ ⎥ 0⎥ 0 ⎥⎦15 T 0 0 0 4 0 4 −25 0 12 −45 ⎡0 ⎢ ⎢0 ⎢0 ⎣ T 0⎤ ⎥ 0⎥ 0 ⎥⎦ 8 0 8 −25 T 0⎤ ⎥ 0⎥ 0 ⎥⎦12 T 16 ⎤ ⎥ 0⎥ 50 ⎥⎦16 ⎡ 1 / 16 ⎢ ⎢ 2 / 16 ⎢ 3 / 16 ⎢ ⎢ 0 ⎢ 5 / 16 ⎢ ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎢ ⎢ 0 ⎢ ⎢ 0 ⎢13 / 16 ⎢ ⎢14 / 16 ⎢15 / 16 ⎢ ⎣⎢16 / 16 0 0 0 4 / 16 0 6 / 16 7 / 16 8 / 16 9 / 16 10 / 16 0 12 / 16 0 0 0 0 40 / 16 ⎤ ⎥ 40 / 16 ⎥ 50 / 16 ⎥ ⎥ −25 / 16 ⎥ −35 / 16 ⎥ ⎥ −220 / 16 ⎥ ⎥ −15 / 16 ⎥ −25 / 16 ⎥ ⎥ −5 / 16 ⎥ −30 / 16 ⎥ ⎥ 0 / 16 ⎥ ⎥ −45 / 16 ⎥ −10 / 16 ⎥ ⎥ −40 / 16 ⎥ 30 / 16 ⎥ ⎥ 50 / 16 ⎥⎦ Smart Grids 0 0 0 1 1 0 40 Converged Matrix M 353 A Consensus-Based Fully Distributed Load Management Algorithm For the eight agents in the working area, that is, Agents 1–3, 5, and 13–16, Mi,1–3 are initialized with the corresponding index, 0, and local net power. For Agent 11, since it is not ready for restoration yet, Mi,1–3 are initialized with 0, 0, and 0. Since Agents 4 and 6–12 are ready for restoration, the Mi,1–3 in the initial matrices are initialized with 0, index, and load demands. By applying the proposed information discovery algorithm to every element of the 16 initial matrices, all 16 information matrices will converge to the same matrix, as shown in Table 14.5. For example, the (1, 1) elements of the 16 matrices are 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0. According to Equation 14.11, the (1, 1) element of the converged matrix will be 1/16. Since load restoration is possible only when excessive generation becomes available, total net power should be checked irst. The average total net power can be calculated by summing up the elements in the third column that have a nonzero element in the irst column. For convenience, the information discovery process of the average total net power in the microgrid is shown in Figure 14.18. From Figure 14.18, it can be seen that all agents’ discovered information converged to the same value of 7.81 = (40 + 40 + 5 – 35 – 10 – 40 + 30 + 50)/16. Since the average total net power is positive, it is possible to restore part or all of the unfaulted but out-of-service loads. To identify loads to restore, it is necessary to ind the indexes and demands of the ready-to-restore loads. The indexes and demands of loads that are ready for restoration are discovered simultaneously, as shown in Figures 14.19 and 14.20. For simpliication, only the (i, 2) and (i, 3) elements of M are displayed for the ith ready-to-restore agent. During the information discovery process, the information matrices are initialized with local information and all converge to the corresponding average values. From Figures 14.18 through 14.20, it can be seen that the information discovery process converges well within 90 iterations. Considering the complexity of the 16-bus microgrid, the speed of the algorithm is more than acceptable. It should be noted that the information discovery process is evaluated based on number of iterations, not time. This is because actual communication time is decided by speciic implementation (hardware and software) and is thus hard to quantify. Since the exchanged information is very simple (index and local net power), it can be predicted that the time used to reach a consensus is very short, especially with the development of computer and communication techniques. Average net power 60 40 20 0 –20 –40 0 10 20 30 40 50 60 Number of iterations 70 80 90 FIGURE 14.18 Average total net power discovery process of the 16-bus system. 100 354 Smart Grids 12 Agent 4 Agent 6 Agent 7 Agent 8 Agent 9 Agent 10 Agent 11 Agent 12 10 Index 8 6 4 2 0 0 10 20 30 40 50 60 Number of iterations 70 80 90 100 FIGURE 14.19 Indexes discovery process of out-of-service loads. Average net power 10 0 Agent 4 Agent 6 Agent 7 Agent 8 Agent 9 Agent 10 Agent 11 Agent 12 –10 –20 –30 –40 0 10 20 30 40 50 60 Number of iterations 70 80 90 100 FIGURE 14.20 Demands discovery process of out-of-service loads. From the converged vector, all required information can be easily calculated. Since the number of agents in the system is ixed, each agent can easily ind the true values from the average values. By checking the nonzero elements in the irst column and the corresponding elements in the third column, the indexes and local net power of working generators and loads can be obtained. Similarly, by checking the nonzero elements in the second column and the corresponding elements in the third column, every agent knows the indexes and demands of loads that are ready for restoration. By checking Tables 14.5 and 14.6, it can be seen that loads connected to Buses 6, 8, and 9 have higher priority compared with loads connected to Buses 4, 7, 10, and 12. For this example, the priority indexes in Equation 14.15 are selected to be 10 and 1 for P1 and P2, respectively. In addition, all agents know the demands of the outof-service loads and the available power, which is 7.81 × 16 = 125. Now, all required information for load restoration has been obtained and the loads can be selected by solving the 0/1 knapsack problem. Here, DP and three greedy algorithms are investigated. Load selection results according to different algorithms are summarized in Table 14.7 [19]. Compared with DP, greedy algorithms may not provide the best energy utilization. However, greedy algorithms usually require less time for calculation, which is a very important property, especially for large or complex systems. For smaller-scale 355 A Consensus-Based Fully Distributed Load Management Algorithm TABLE 14.7 Comparison of Different Algorithms for Load Restoration Loads Selected for Restoration Index (Demand) Algorithm Dynamic programming Greedy by proit Greedy by weight Greedy by proit density Utilization Ratio (%) 6(−20), 8(−25), 9(−5), 10(−30), 12(−45) 8(−25), 6(−20), 9(−5), 12(−45), 10(−30) 12(−45), 10(−30), 4(−25), 8(−25) 9(−5), 6(−20), 8(−25), 7(−15), 4(−25), 10(−30) 100 100 100 96 systems like the example used here, DP is a good choice. For large and complex systems, greedy algorithms will be a better choice. The microgrid after restoration is shown in Figure 14.21, in which the generator and load in circles are disconnected by opening the corresponding circuit breakers. The information discovery process of average total net power is shown in Figure 14.22. Since the converged value is 0, no more load restoration is possible. 2 3 4 10 1 5 11 6 7 8 9 12 14 13 15 16 Connected Disconnected FIGURE 14.21 Coniguration of the 16-bus system after restoration. Average net power 60 40 20 0 –20 –40 0 10 20 30 40 50 60 70 80 Number of iterations FIGURE 14.22 Average total net power discovery process after restoration. 90 100 356 Smart Grids 14.4 CONCLUSION In this chapter, a stable MAS-based load management algorithm is proposed for microgrids based on the average consensus theorem. Even though each agent only communicates with its direct neighbors, global information that is needed for distributed load management can be discovered. Compared with existing research on MAS-based power system applications, the proposed algorithm has the following advantages. First, stability of the algorithm can be rigorously guaranteed. Second, the algorithm can be applied to power systems of any coniguration and size. Third, an improved parameter-setting algorithm, Mean Metropolis, is proposed to adjust the coeficients for exchanged information. Simulation results demonstrate the effectiveness of the proposed algorithm. REFERENCES 1. S. Borg, G. Andersson, D.J. Hill, and I.A. Hiskens, On inluence of load modeling for undervoltage load shedding studies, IEEE Transactions on Power Systems, 13(2), 395–400, 1998. 2. J. Mendoza, R. López, D. Morales, E. López, Ph. Dessante, and R. Moraga, Minimal loss reconiguration using genetic algorithms with restricted population and addressed operators: Real application, IEEE Transactions on Power Systems, 21(2), 59–67, 2006. 3. J.G. Vlachogiannis and N.D. Hatziargyriou, Reinforcement learning for reactive power control, IEEE Transactions on Power Systems, 9(3), 1317–1325, 2004. 4. J.M. Solanki, S. Khushalani, and N.N. Schulz, A multi-agent solution to distribution systems restoration, IEEE Transactions on Power Systems, 22(3), 1026–1034, 2007. 5. K. Huang, D.A. Cartes, and S.K. Srivastava, A multiagent-based algorithm for ringstructured shipboard power system reconiguration, IEEE Transactions on Power Systems, 37(5), 1016–1021, 2007. 6. M. Wooldridge, An Introduction to Multiagent Systems, John Wiley & Sons, New York, 2002. 7. T. Logenthiran, D. Srinivasan, and D. Wong, Multi-agent coordination for DER in MicroGrid, in Proceedings of the IEEE International Conference on Sustainable Energy Technologies, pp. 77–82, 24–27 November, 2008, Singapore. 8. J.M. Solanki and N.N. Schulz, Using intelligent multi-agent systems for shipboard power systems reconiguration, in Proceedings of the 13th International Conference on Intelligent Systems Application to Power Systems, 6–10 November, 2005, Arlington, VA. 9. A. Dimeas and N.D. Hatziargyriou, Operation of a multiagent system for microgrid control, IEEE Transactions on Power Systems, 20(3), 1447–1455, 2005. 10. Y. Xu, W. Liu, and J. Gong, Stable multi-agent based load shedding algorithm for power systems, IEEE Transactions on Power Systems, 26(4), 2006–2014, 2011. 11. Y. Xu and W. Liu, Novel multi-agent based load restoration algorithm for microgrids, IEEE Transactions on Smart Grid, 2(1), 152–161, 2011. 12. F. Xiao, L. Wang, and Y. Jia, Fast information sharing in networks of autonomous agent, in Proceedings of the American Control Conference, pp. 4388–4393, 11–13 June, 2008, Seattle, WA. 13. R. Olfati-Saber, J. Fax, and R. Murray, Consensus and cooperation in networked multiagent systems, Proceedings of the IEEE, 95(1), 215–233, 2007. 14. R.A. Horn, Doubly stochastic matrices and the diagonal of a rotation matrix, American Journal of Mathematics, 76, 620–630, 1954. A Consensus-Based Fully Distributed Load Management Algorithm 357 15. R.A. Horn and C.R. Johnson, Matrix Analysis, Cambridge University Press, Cambridge, UK, 1987. 16. L. Xiao and S. Boyd, Fast linear iterations for distributed averaging, Systems and Control Letters, 53, 65–78, 2004. 17. R. Olfati-Saber, Algebraic connectivity ratio of Ramanujan graphs, in Proceedings of 2007 American Control Conference, pp. 4619–4624, 11–13 July, 2007, New York. 18. R. Perez-Guerrero, G.T. Heydt, N.J. Jack, B.K. Keel, and A.R. Castelhano, Optimal restoration of distribution systems using dynamic programming, IEEE Transactions on Power Delivery, 23(3), 1589–1596, 2008. 19. T.H. Cormen, C.E. Leiserson, and R.L. Rivest, Greedy Algorithms, pp. 370–404. MIT Press, Cambridge, MA, 2001. 20. T. Nagata, Y. Tao, and H. Fujita, An autonomous agent for power system restoration, IEEE Power Engineering Society General Meeting, 1, 1069–1074, 2004. 15 Expert Systems Application for the Reconiguration of Electric Distribution Systems Horacio Tovar-Hernández and Guillermo Gutierrez-Alcaraz CONTENTS 15.1 Introduction .................................................................................................. 359 15.2 Expert System ............................................................................................... 362 15.2.1 Optimal Power Flow ......................................................................... 363 15.2.2 Volt/VAr Control ..............................................................................364 15.3 Case Study .................................................................................................... 365 15.3.1 Scenario 1 ......................................................................................... 366 15.3.2 Scenario 2 ......................................................................................... 371 15.3.3 Scenario 3 ......................................................................................... 372 15.4 Summary ...................................................................................................... 373 References .............................................................................................................. 373 15.1 INTRODUCTION In the past three decades, intensive effort has been devoted to formulating and solving distribution network optimal planning and operating problems. Nowadays, advances in communication and computer technologies have made feasible the development of more sophisticated methodologies to solve some of the particular problems deriving from those two general challenges. For example, in an operational framework, advances in distribution automation (DA) technology have substantially improved control and management capabilities, so that it is possible to reconigure distribution electric systems directly from an operation control center [1]. Also, time-varying consumer demand, always an issue, can be addressed by very speciic feeder reconiguration to enable load transfers from heavy or overloaded feeders to feeders with available load capacity [2]. 359 360 Smart Grids The concept of the smart grid is based on a distribution management system (DMS) comprising a set of hardware and software applications installed in a control center for monitoring and controlling distribution systems with a high degree of automation. An important objective is to enhance system operation eficiency and reliability. For this purpose, many utilities rely on two-way communications to monitor and control equipment at distribution substations and feeders. A DMS incorporates the measurement data from DA and advanced metering infrastructure to provide more precise load estimations of feeder sections for integrated Volt/VAr and service restoration controls [3]. The DMS, which is able to maintain information about the current and past topologies of the electric distribution system, facilitates and automates the creation of switching orders for planned and unplanned work that operators or ield personnel can execute according to traditional procedures, or that can be executed automatically [1]. The software infrastructure associated with a DMS can be classiied into the sets of basic or conventional applications, and advanced or intelligent applications. The former set includes conventional algorithms such as load lows, short-circuit analysis, state estimation, and load forecasting, while the latter combines these applications with database and operational criteria management to help operators to obtain the optimal operational set points of each network component [4]. The DMS is superior to older technology in terms of lexibility and handling operation planning horizons by acquiring, iltering, and manipulating data for eficient decision making. Within this environment, operators are able to determine the optimal state of the distribution system as much as an hour in advance [2]. Distribution systems are subject to operational changes that can be either smooth or critical, depending on the magnitude of the perturbation experienced. If load variations or less severe failures occur, resulting in some overloads or voltages out of limits, then a transition from the normal to the emergency state occurs, and some corrective actions can be performed to restore the system to a normal operational state. A different situation occurs when a perturbation is a permanent failure; in this case, the distribution system moves from normal to an emergency state, in which fault-clearing actions are carried out and some subsequent restorative actions can be performed to take the distribution system to a new normal operational state. These processes are illustrated in Figure 15.1. In the normal state, relatively small or smooth changes occur, for example, hourly variations of load, and normal-state operational conditions may vary. Therefore, operators can make some decisions to take the distribution system from a nonoptimal Optimal system reconfiguration Normal state Operational changes Emergency state Fault clearing Restorative state Corrective actions Restorative actions FIGURE 15.1 Operational states, transitions between them, and remedial actions in electrical distribution systems. Expert Systems Application for the Reconiguration of Distribution Systems 361 operational condition to a better operational state by using a short-term operation planning process comprising the following strategies [4]: 1. 2. 3. 4. 5. Detection of restriction violations Corrective action planning Maintenance outage programming Loss reduction by reconigurations Volt/VAr control A combination of these strategies allows the development of sophisticated methodologies with the capacity to obtain an optimal operative state. For example, loss reduction can be optimized by reconiguration and Volt/VAr control. Under the smart grid framework, a great number of algorithms and methodologies have been developed to optimize distribution system operations. In particular, solving the problem of loss minimization is a key factor in laying out circuits as well as predicting a desired system coniguration for different normal-operation or contingency cases. For example, many proposals solve the reconiguration problem from a normal-operation condition perspective involving either one or several load scenarios in order to follow load variations more accurately and to obtain minimum distribution losses for each load scenario [5]. Both of these approaches normally change the open/closed status of tie switches in a progressive manner, until they ind a radial coniguration of the distribution system with minimum losses, while keeping nodal voltages within limits and avoiding overloads on each feeder section, including substation transformers. In some cases, such proposals are based on conventional power low solvers combined with topological search algorithms [6]; in others, they are based on evolutionary programming techniques, including power low solvers [7,8] and sometimes topological search methods [9]. Nevertheless, solving the more complex problem of reconiguration with Volt/VAr control has not been addressed with the same interest, even though its solution should result in a greater reduction of distribution losses while still maintaining better nodal voltage proiles. It has been proposed to solve this problem by multiobjective/multistep evolutionary programming techniques [9–11] or optimal power low (OPF) Benders decomposition algorithms [12,13], both including the radial-structure restriction. Furthermore, expert systems (ESs) have been shown to be useful for dealing with distribution system operation. ESs have been developed as decision-making support methodologies for distribution system operators to solve problems of restoration, feeder load balancing, voltage control, and loss minimization in distribution systems [14,15]. These tools make use of computational algorithms such as load low, topological search, sensitivities, and optimization techniques, in order to obtain solutions that are feasible and optimal in some sense to the system operator [16]. This combination of artiicial intelligence tools and numerical algorithms evolves into methodologies that could be more eficient than those based only on very complex numerical algorithms or evolutionary multiobjective methodologies for solving mixed-integer nonlinear programming [17]. In many cases, however, ESs lack solid mathematical foundations, so there is no guarantee that they reach an optimal solution; rather, it may be only a good approximation [16]. Nevertheless, due to the discrete nature of tie switching, transformer and regulator tap adjustment, and capacitor 362 Smart Grids bank commutation, the solutions obtained with ESs and those based on optimization algorithms or evolutionary programming will be very similar in practical terms. In this chapter, we propose an ES for optimal loss reduction that uses a coordinated application for reconiguration and Volt/VAr control in the distribution system. For this purpose, the ES improves a set of solutions proposed by an OPF-based reconiguration algorithm, executing reactive power control actions by capacitor bank commutations and regulator tap adjustments. 15.2 EXPERT SYSTEM As mentioned, intelligent methodologies are needed to support operational decision making in an information-rich environment. In a dynamic electrical distribution system, the system operator must make decisions in real time to maintain the distribution system in an optimal operational state. ESs, or knowledge-based systems, were developed approximately 30 years ago to assist system operators in deining appropriate control actions to keep transmission or distribution systems operating under security restrictions [14,16]. Some ESs have been designed as aids for transmission and distribution system operators for voltage and reactive power control. In general, ES applications in electrical engineering are designed as knowledge-based systems based on heuristics and numerical calculations that give more intelligent solutions to complex problems than those given by conventional algorithms [16]. Our proposed ES is a rule-based system with a forward-chaining reasoning that involves two general stages: (1) inding one set of distribution system reconigurations with voltages within limits and no overloads by means of an OPF, and (2) performing a reactive Volt/VAr control to reduce losses given by the set of reconigurations found in stage (1). We organized the If-Then rules into steps for making orderly and eficient decisions. In our hypothetical real-time environment where the ES must be able to act for the next load scenario, each step represents a level of the following solution process: Step 1. Identify the actual state of switches between feeders, capacitor banks, and nodal loads for the current hour, deined as h−1. Step 2. For the next load demand scenario, D(h), ind all n options for reconiguration and classify the options as feasible reconigurations (FRs). Calculate each FR by the OPF algorithm, setting all capacitor banks in a disconnected state, so that each FR found does not present overloads or voltage deviations out of limits. If n = 0, the ES modiies the voltage limits until the OPF inds at least one FR. Step 3. For each FR, perform a reactive power control by means of capacitor bank commutations to investigate their impact over distribution losses until there is no capacitor bank commutation that reduces losses. Step 4. Find the reconiguration with the minimum distribution losses, and name it as the new operating state, E(h). Step 5. Verify that E(h) ≠ E(h−1). If true, deine the changes needed in switches and capacitor banks to pass from the E(h−1) state to the E(h) state, and convey this information to the system operator. Expert Systems Application for the Reconiguration of Distribution Systems 15.2.1 363 OPTIMAL POWER FLOW The OPF algorithm obtains reconigurations with voltages remaining within certain limits [12,13,18]. However, at this stage the solutions of the OPF are allowed to present slight voltage deviations out of limits, which can be corrected by reactive power/ voltage control by capacitor bank switching and adjusting regulator taps. Interfaces between transmission and distribution are treated as a voltage-controlled node with one generator. These generators will have the same cost production function, such that the power sharing among them is not inluenced by the most economical generation contribution, but by the least generation needed to satisfy all the loads connected to the distribution system. Hence, OPF solves an economical dispatch for each coniguration r ∈ R, where R is the set of all possible radial network reconigurations. Each reconiguration r without overloads and voltages out of limits is included in a subset of FRs. Mathematically, the OPF can be formulated as a nonlinear optimization problem: ng Min C T ( r ) = {Pgi } ∑ ( d + e Pg + f Pg ) i i i i 2 i (15.1) i =1 N S. to Pgk ( r ) − Pdk ( r ) = ∑ V V (G k j kj cos ( θk − θ j ) + Bkj sin ( θk − θ j ) ) ∀k ∈ N j =1 (15.2) N Qgk ( r ) − Qdk ( r ) = ∑ V V (G k j kj sin ( θk − θ j ) − Bkj cos ( θk − θ j ) ) ∀k ∈ N (15.3) j =1 ∑S pk ( r ) ≤ S pMax ∀ p ∈ Nt (15.4) pk ( r ) ≤ S pMin ∀ p ∈ Nts (15.5) k∈K out ∑S k∈K out VkMin Vk ,r S j ,r ≤ S Max j VkMax (15.6) j ∈ Nb (15.7) Nrs = N − ng where: N Nb Nt Kout Spk is the total number of nodes is the total number of branches is the set of all substation transformers is the set of feeders receiving power from transformer p is the power lowing from transformer p to feeder k (15.8) 364 Smart Grids S pMax is the rating of transformer p θk is the voltage angle at node k Pdk is the active load at node k Qdk is the reactive load at node k Pgk is the active power injected at node k Qgk is the reactive power injected at node k |Vk| is the voltage magnitude at node k VkMin , VkMax are the allowed minimum and maximum voltage magnitude at node k Sj is the loading of feeder j S Max is the rating of feeder j j di, ei, and fi are the production cost factors Gkj and Bkj are the real and imaginary parts of the admittance matrix element Ykj, respectively ng is the total number of substations or generators The objective function (Equation 15.1) is the minimization of the production cost of power provided by the utility. Constraints 15.2 and 15.3 represent the nodal power balance equations. Constraints 15.4 and 15.5 represent the maximum and minimum power output limits of the substation or generator, respectively. Inequality constraints 15.6 and 15.7 are the system operative and security constraints (upper and lower bounds), respectively [13]. Restriction 15.8 is related to the radial structure of the distribution system, where Nrs is the number of branches included in the spanning tree representing each one of the possible distribution system reconigurations [2]. In our formulation, the OPF does not perform reactive power control to minimize distribution losses, chiely because this process is carried out by capacitor bank switching and regulator tap adjustments, which are discrete control actions, and the OPF algorithm would have to solve a more complex mixed nonlinear and integer programming problem. Traditionally, reactive power dispatch for transmission systems has been formulated assuming that these devices are continuous variables in the OPF formulation. However, commercial capacitor bank capacities are relatively large (e.g., 300 kVAr modules), so that modeling them as continuous variables in OPF for distribution systems would lead to impractical results. 15.2.2 VOLT/VAR CONTROL Traditionally, reactive power control is performed to minimize losses and control voltages through capacitor bank switching and regulator tap adjustments. The resulting operational state of these devices is maintained during a long period of time deined by considerable changes in seasonal load patterns. Capacitor bank switching can be ixed or commutable. In the latter case, commutation is made by local criteria such as temperature changes, low voltage, and so on. Furthermore, regulator taps are adjusted by measuring the voltage on a speciic distribution node. However, under the smart grid paradigm, it is assumed that the control center must monitor and control these devices by considering their beneit to the entire distribution system, that is, minimization of loss and voltage deviation. Expert Systems Application for the Reconiguration of Distribution Systems 365 Capacitor bank commutations and regulator tap adjustments should strictly be modeled as discrete events. To ind the best combination of capacitor bank commutations and regulator adjustments to reduce the distribution losses to a minimum, the ES performs a dynamic programming process. Experimental results show that a good practice is irst to analyze the effect of capacitor banks, and then realize the tap adjustments in regulators, because capacitor banks have a great impact on loss reduction, while regulator taps have a minor effect, but can be used to make inal changes in voltage magnitudes, which also help to diminish losses. In particular, for our proposed computational methodology, we consider that capacitor banks have a modular capacity based on 300 kVAr blocks. Once the OPF has obtained the subset of FRs, the ES performs the Volt/VAr control following the next steps: 1. Select an FR from the subset FR. 2. For the capacitor bank set, select one capacitor block and switch it in or out, depending on its actual state. Observe its impact on distribution losses; if they augment, leave this capacitor bank in its actual state; if they drop, add the capacitor block commutation to a feasible control action (FCA) list and turn back the capacitor block to its original status. To avoid capacitor bank commutations with an almost null impact on loss reduction, it is convenient to establish a threshold value, for example, 5 kW. 3. If there are more capacitor bank blocks to investigate, go back to Step 2. If not, go to next step. 4. If the FCA is empty, go to Step 6. If not, go to the next step. 5. Order the control actions in the FCA from greater to lesser impact. Once the FCA has been ordered, the control action that causes the greatest loss reduction is put into a selected control action (SCA) list. 6. For each regulator, investigate adjustments in its tap, which can reduce the distribution loss value reached in Step 2. Select the adjustment with the greatest loss reduction, and add it to the SCA list. To avoid regulator tap adjustments with an almost null impact on loss reduction, it is convenient to establish a threshold value, for example, 5 kW. 7. Return to Step 1 if there are more FRs to investigate. If not, create the deinitive control action (DCA) list and end the process. The inal DCA list will be the SCA list corresponding to the least-loss reconiguration. The process with OPF and Volt/VAr control, as described above, is designed for a load demand scenario h. If there are more scenarios to be analyzed, then h and h−1 are updated, and the whole process is repeated until all the load scenarios have been analyzed. 15.3 CASE STUDY A practical distribution network with two contiguous substations is used to illustrate the ES functioning. It is considered that the distribution system operates under threephase balanced conditions; therefore, the positive sequence network represents the 366 Smart Grids system network. The complete network data are given in [17], and the corresponding single-line diagram is depicted in Figure 15.2. Both substations operate independently as radial feeders, but they can be interconnected by one of the normally open tie-line switches. Optimal network coniguration problem is solved by brute force combined with graph theory implemented in MATLAB®. For illustrative purposes, only switches connecting tie-lines are considered for load transfer (S84, S85, S87, S88, and S91). However, to preserve a radial structure, when switch S91 operates, switch S21 or switch S83 should operate simultaneously, depending on where the load transfer takes place; switches S5, S7, S11, S12, S55, S64, S72, and S76 are in the same situation. Intra-area switches, for example, switches 89 and 97, are mainly used to achieve load balancing between feeders belonging to a substation, depending on the diversity and amount of each feeder demand. Under this consideration, it is possible to operate 15 switches and generate 32,768 possible reconigurations. A subset of this large number of possible reconigurations will be feasible and the rest will not be feasible due to overloads, out-of-limit voltages, or both. Table 15.1 shows individual switch operation states for 11 possible reconigurations, which were selected randomly, the only purpose being to show how our proposal works. However, in a real-life environment, there should be an algorithm that helps to ind the more promising subset of FRs. Table 15.2 shows the location and capacity of capacitor banks; they can be connected or disconnected by 300 kVAr blocks. In this case study, regulators do not exist in the distribution system, and thus they are not included in the reconiguration process. Also, it is assumed that the ES acts to propose a new reconiguration depending on the load scenario for the next hour, while taking the current scenario into account. For illustration purposes, we analyze three of the hourly load scenarios that could occur during a period of time, namely, a day ahead. To initiate the ES process, we consider that the distribution system is operating under the reconiguration 00 shown in Table 15.1, and the nine capacitor banks of Table 15.2 are disconnected. Additionally, voltage limits are speciied as 0.95 and 1.05 p.u. in all nodes, and we will assume that all feeder sections have the very high overload limit of 100 MVA. 15.3.1 SCENARIO 1 Scenario 1 presents the loads shown in Table 15.3, for all the nodes of the distribution system. The substation voltages are set to 1.01 p.u. The ive general steps are developed below: Step 1. The ES identiies the actual operation state E(0) based on the information presented in Tables 15.1 through 15.3, also considering that the distribution system operates under the 00 reconiguration, and all capacitor banks are disconnected. Step 2. The OPF obtains FRs observing overloads and voltages for the 11 reconigurations. For scenario 1, reconigurations 01, 09, and 10 are classiied as unfeasible, while reconigurations 00, 02, 03, 04, 05, 06, 07, and 08 are classiied as feasible, that is, they are included in the subset FR. Expert Systems Application for the Reconiguration of Distribution Systems 367 Sub 1 B6 B5 B4 B3 B2 B1 Bus 1 43 11 90 12 26 32 92 46 94 28 89 18 29 6 7 21 93 22 38 10 9 8 S84 24 41 39 42 40 S85 S91 95 S88 S87 64 72 55 96 53 83 76 73 77 B11 56 65 B10 B9 47 B8 B7 Sub 2 FIGURE 15.2 Distribution system with two substations and 11 feeders operating in radial structure. 368 TABLE 15.1 Individual Tie-Line Switch States for the 11 Possible Conigurations Reconiguration S5 S7 S11 S12 S21 S55 S64 S72 S76 S83 S84 S85 S87 S88 S91 00 01 02 03 04 05 06 07 08 09 10 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 0 1 1 1 1 1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 1 Smart Grids Expert Systems Application for the Reconiguration of Distribution Systems TABLE 15.2 Capacitor Bank Location and Capacity (kVAr) No. 1 2 3 4 5 6 7 8 9 Node Capacity 6 21 28 37 51 63 71 75 81 2400 2100 2400 1800 2100 1500 2400 2400 2400 TABLE 15.3 Active and Reactive Nodal Demands for Scenario 1 Node MW MVAr Node MW MVAr Node MW MVAr 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 0.08 0.24 0.28 0.176 0.88 0.32 0.24 0.24 0.24 0 0.96 0.64 0.56 0 0.24 0.4 0.56 0.96 0.24 0.32 0.04 0.04 0.04 0.04 0.08 0.4 1.44 0.16 0.04 0.16 0.2 0.08 0.4 −0.224 0.16 0.104 0.208 0 0.32 0.48 0.4 0 0.12 0.28 0.32 0.8 0.24 0.28 0.016 0.016 0.008 0.024 0.048 0.28 1.28 0.096 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 0 1.44 0.16 0.16 0.64 0.08 0.08 0.016 0.016 0.016 0.016 0.16 0.04 0 0.024 0.64 0.16 0 0 0 0.16 0.64 0.4 0.4 0.4 0.16 0 0.024 0 1.28 0.12 0.08 0.48 0.048 0.048 0.008 0.008 0.008 0.008 0.128 0.024 0 0.016 0.56 0.12 0 0 0 0.128 0.48 0.24 0.04 0.24 0.064 0 0.016 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 0.48 0 0.016 0.016 0.16 0.24 0.24 0 0.04 0 0.32 0 0 1.6 0.16 0.64 0.4 1.6 0.72 0 0.32 1.6 0.16 0.4 0.08 0.32 0 0 0.096 0 0.008 0.008 0.024 0.192 0.16 0 0.024 0 0.288 0 0 0.96 0.12 0.16 0.096 1.08 0.304 0 0.288 0.8 0.112 0.288 0.024 0.288 0 0 369 370 Smart Grids Step 3. With the results of Step 2, the ES performs the reactive power control to minimize distribution losses with the connection of 300 kVAr capacitor bank blocks for each of the FRs of scenario 1. Once the ES has found all the capacitor bank blocks that should be connected in order to reduce the distribution losses to a minimal value, it goes to the next step. Step 4. Distribution losses for all FRs and the corresponding order with and without reactive control are shown in Table 15.4. Based on this information, the ES deines 07 as the best reconiguration, because it has the minimum distribution losses. The results from this table show the different order before and after applying the reactive power control; this strategy, performed by the ES to minimize losses, holds promise for helping distribution companies to achieve substantial savings, as is evident from the columns without and with reactive power control. Step 5. The ES inds the control actions to be performed to reach the new operating state E(1). The initial and inal network conigurations are coniguration 00 and coniguration 07, respectively. It can be seen that the only changes needed to pass from the actual reconiguration to the new one are the following: Control action 1: Close switch S88 Control action 2: Open switch S12 Note that the control action to close tie-line switch S88 is executed irst, in order to prevent the isolation of some portion of the electrical network. Since we assumed that capacitor banks were initially in disconnected status, they are connected as speciied by the next control actions: Control action 3: Connect 1200 kVAr at node 6 Control action 4: Connect 1200 kVAr at node 21 TABLE 15.4 Distribution Losses for All FRs of Scenario 1 Without Reactive Power Control Reconiguration 00 02 03 04 05 06 07 08 With Reactive Power Control Losses (MW) Ordering Losses (MW) Ordering 0.3952 0.4010 0.3930 0.4065 0.3956 0.3921 0.3949 0.3984 4 7 2 8 5 1 3 6 0.3005 0.3040 0.2995 0.3081 0.2993 0.2997 0.2943 0.3042 5 6 3 8 2 4 1 7 Expert Systems Application for the Reconiguration of Distribution Systems 371 Control action 5: Connect 1500 kVAr at node 28 Control action 6: Connect 900 kVAr at node 37 Control action 7: Connect 1200 kVAr at node 51 Control action 8: Connect 1200 kVAr at node 63 Control action 9: Connect 1500 kVAr at node 71 Control action 10: Connect 1500 kVAr at node 75 Control action 11: Connect 1500 kVAr at node 81 In this case, the control action order to connect capacitor bank blocks has been established according to Table 15.2, because no capacitor-connection sequence will cause overloads or voltages out of limits, and the inal result of realizing control actions 4–11 will be the same independently of the execution order. Note that there are many control actions for reaching the optimal reconiguration. However, when there are small changes in reconiguration between scenarios, the control actions will be few, mainly because we already have an optimal solution, as will be observed when scenario 2 is analyzed. 15.3.2 SCENARIO 2 The loads of scenario 2 are obtained by multiplying the active and reactive loads of scenario 1 by 1.125. Once again, the substation voltages are set to 1.01 p.u. The ES carries out Step 1 considering scenario 2, actual reconiguration 07, and the corresponding capacitor bank status shown through the control actions 3–11 listed above. The results of Step 2 show that reconigurations 01, 02, 04, 09, and 10 are now unfeasible, while 00, 03, 05, 06, 07, and 08 are feasible. Application of Steps 3 and 4 produces the results shown in Table 15.5, where minimum losses are reached once again for reconiguration 07, and the three capacitor bank control actions. For Step 5, the ES displays the next control actions, which are the only changes with respect to the operational state E(1) for reaching E(2): Control action 1: Connect 300 kVAr at node 28 Control action 2: Connect 300 kVAr at node 37 Control action 3: Connect 300 kVAr at node 81 TABLE 15.5 Distribution Losses for All FRs of Scenario 2 Without Reactive Power Control Reconiguration 00 03 05 06 07 08 With Reactive Power Control Losses (MW) Ordering Losses (MW) Ordering 0.5054 0.5025 0.5060 0.5013 0.5052 0.5096 3 2 5 1 4 6 0.3819 0.3808 0.3811 0.3816 0.3748 0.3874 5 2 3 4 1 6 372 Smart Grids According to these results, capacitor banks should be connected with the capacity listed below: Node 6: 1200 kVAr Node 21: 1200 kVAr Node 28: 1800 kVAr Node 37: 1200 kVAr Node 51: 1200 kVAr Node 63: 1200 kVAr Node 71: 1500 kVAr Node 75: 1500 kVAr Node 81: 1800 kVAr 15.3.3 SCENARIO 3 The loads for scenario 3 are obtained by multiplying the loads of scenario 1 by 1.25 and substituting the loads at nodes 73, 74, 75, and 76 with the values of (0.8,0.2), (0.5,0.12), (2.0,1.35), and (0.9,0.38), MW and MVAr, respectively. In this case, substation voltages are set to 1.02 p.u. For the realization of Step 1, E(2) is taken as the actual operational state, but considering the corresponding loads of scenario 3. The results of Step 2 show that feasible and unfeasible reconigurations are the same as for scenario 2. Table 15.6 shows the solutions obtained by the ES after carrying out Steps 3 and 4. It is clear that now the reconiguration 03 is the new coniguration for scenario 3. For scenario 3, the control actions displayed by the ES to the distribution system operator would be the following: Control action 1: Close switch S12 Control action 2: Close switch S85 Control action 3: Open switch S88 Control action 4: Open switch S7 Control action 5: Disconnect 300 kVAr at node 6 Control action 6: Connect 300 kVAr at node 21 Control action 7: Connect 300 kVAr at node 28 TABLE 15.6 Distribution Losses for All Reconigurations of Scenario 3 Without Reactive Power Control Reconiguration 00 03 05 06 07 08 With Reactive Power Control Losses (MW) Ordering Losses (MW) Ordering 0.6881 0.6847 0.6888 0.6829 0.7137 0.6849 4 2 5 1 9 3 0.5192 0.5181 0.5184 0.5187 0.5291 0.5244 5 1 2 3 4 6 Expert Systems Application for the Reconiguration of Distribution Systems 373 Control action 8: Connect 300 kVAr at node 51 Control action 9: Connect 600 kVAr at node 63 Control action 10: Connect 300 kVAr at node 71 Control action 11: Connect 600 kVAr at node 75 Control actions 1–4 are executed to change from reconiguration 07 to 03. Control action 5 shows that the disconnection of a 300 kVAr capacitor bank block should take place, while control actions 6–11 indicate the connection of 300 or 600 kVAr capacitor bank blocks relative to their status in scenario 2. This result appears to be because the load of scenario 3 is greater than the load of scenarios 1 and 2. The inal connection statuses of the capacitor banks are: Node 6: 900 kVAr Node 21: 1500 kVAr Node 28: 2100 kVAr Node 37: 1200 kVAr Node 51: 1500 kVAr Node 63: 1800 kVAr Node 71: 1800 kVAr Node 75: 1800 kVAr Node 81: 1800 kVAr 15.4 SUMMARY According to the results shown in this chapter, it is clear that considering only reconiguration techniques for loss reduction may result in suboptimal reconigurations. Also, the reactive power/voltage control signiicantly reduces distribution losses obtained with a reconiguration methodology alone. Thus, we believe that this key feature for optimizing smart grid performance will give operators the ability to make “smarter” operational decisions, once the planning for optimal location of capacitors has been done. We suggest that system operators will ind that the proposed approach of the ES integrated by the OPF and Volt/VAr control is highly effective for minimizing losses and achieving superior voltage proiles. Furthermore, it is important to note that the ES-based methodology can be applied in planning operational scenarios for large periods of time, that is, seasonal load conditions, because we are considering three-phase balanced distribution system modeling, speciic load proiles, and normal-state operational conditions for reconiguration, capacitor commutation, and regulator tap adjustments. Also, in the realtime frame, coordinated network reconiguration and Volt/VAr control can handle the time-varying effects of feeder bus loads and distributed generation outputs. REFERENCES 1. F. Borjas, A. Espinosa, A. Quintero, B. Sierra, and R. Torres-Abrego, An architecture for integrating an expert system with NEPLAN in a DMS/EMS operational environment, in Proceedings of the 2009 IEEE International Conference on Systems, Man, and Cybernetics, 11–14 October, San Antonio, TX, 2009. 374 Smart Grids 2. E. López, H. Opazo, L. García, and P. Bastard, Online reconiguration considering variability demand: Applications to real networks, IEEE Trans. Power Syst., 19(1), 549–553, 2004. 3. S.-Y. Su, C. Lu, R.-F. Chang, and G. Gutiérrez-Alcaraz, Distributed generation interconnection planning: A wind power case study, IEEE Trans. Smart Grid, 2(1), 181–189, 2011. 4. E. Lakervi and E. J. Holmes, Electric Distribution Network Design, 2nd edn., IEE Power Engineering Series 21, Peter Peregrinus Ltd, Exeter, 1995. 5. S.-A. Yin and C.-N. Lu, Distribution feeder scheduling considering variable load proile and outage costs, IEEE Trans. Power Syst., 24(2), 652–660, 2009. 6. S. Civanlar, J. J. Grainger, and S. S. H. Le, Distribution feeder reconiguration for loss reduction, IEEE Trans. Power Deliver., 3(3), 1217–1223, 1988. 7. C.-C. Liu, S. J. Lee, and K. Vu, Loss minimization of distribution feeders: Optimality and algorithms, IEEE Trans. Power Deliver., 4(2), 1281–1289, 1989. 8. D. Jiang and R. Baldick, Optimal electric distribution system switch reconiguration and capacitor control, IEEE Trans. Power Syst., 11(2), 890–897, 1996. 9. D. Zhang, Z. Fu, and L. Zhang, Joint optimization for power loss reduction in distribution systems, IEEE Trans. Power Syst., 23(1), 161–169, 2008. 10. C.-F. Chang, Reconiguration and capacitor placement for loss reduction of distribution systems by ant colony search algorithm, IEEE Trans. Power Syst., 23(4), 1747–1755, 2008. 11. C. T. Su and C. S. Lee, Feeder reconiguration and capacitor setting for loss reduction of distribution systems, Electr. Power Syst. Res., 58(2), 97–102, 2001. 12. H. M. Khodr, J. Martínez-Crespo, M. A. Matos, and J. Pereira, Distribution systems reconiguration based on OPF using Benders decomposition, IEEE Trans. Power Deliver., 24(4), 2166–2176, 2009. 13. H. M. Khodr, J. Martínez-Crespo, Z. A. Vale, and C. Ramos, Optimal methodology for distribution systems reconiguration based on OPF and solved by decomposition technique, Euro. Trans. Electr. Power, 20(6), 730–746, 2009. 14. M. M. Salama and A. Y. Chikhani, An expert system for reactive power control of a distribution system part 1: System coniguration, IEEE Trans. Power Deliver., 7(3), 940–945, 1992. 15. J. R. P.-R. Laframboise, G. Ferland, A. Y. Chikhani, and M. M. A. Salama, An expert system for reactive power control of a distribution system part 2: System implementation, IEEE Trans. Power Syst., 10(3), 1433–1441, 1995. 16. T. S. Dillon and M. A. Laughton, Expert System Applications in Power Systems, Prentice Hall, New York, 1990. 17. C. Wang and H. Z. Cheng, Optimization of network coniguration in large distribution systems using plant growth simulation algorithm, IEEE Trans. Power Syst., 23(1), 119–126, 2008. 18. C. J. Dent, L. F. Ochoa, and G. P. Harrison, Network distributed generation capacity analysis using OPF with voltage step constraints, IEEE Trans. Power Syst., 25(1), 296–304, 2010. 16 Load Data Cleansing and Bus Load Coincidence Factors* Wenyuan Li, Ke Wang, and Wijarn Wangdee CONTENTS 16.1 Introduction .................................................................................................. 375 16.2 Load Curve Data Cleansing ......................................................................... 377 16.2.1 Basic Concepts .................................................................................. 377 16.2.2 Method for Cleansing Outliers ......................................................... 378 16.2.2.1 Kernel Function Representation of Load Curve ................ 378 16.2.2.2 Conidence Interval for Identifying y-Outliers .................. 379 16.2.2.3 Candidates for x-Outliers ................................................... 380 16.2.2.4 Identifying x-Outliers ........................................................ 382 16.2.3 Repairing Outliers ............................................................................ 384 16.2.4 Selection of Parameters .................................................................... 384 16.3 Bus Load Coincidence Factors ..................................................................... 385 16.3.1 Basic Concept of BLCF .................................................................... 386 16.3.2 Procedure for Calculating BLCF ...................................................... 387 16.4 Application of BLCF to System Analysis..................................................... 389 16.4.1 Studied System and Load Curve Data .............................................. 389 16.4.2 Bus Load Coincidence Factors at the Three Substations ................. 390 16.4.3 Power Flow Analysis Using the Bus Load Coincidence Factors ...... 393 16.5 Conclusions ................................................................................................... 397 References .............................................................................................................. 397 16.1 INTRODUCTION Load curve data refer to power consumptions recorded by meters at certain time intervals at buses of individual substations. Load curve data are one of the most important datasets collected and retained by utilities. The analysis of load curve data would greatly improve day-to-day operations, system analysis in smart grids, system visualization, system performance reliability, energy saving, and accuracy in system planning [1–4]. * This work is partly supported by a collaborative research and development grant from the Natural Sciences and Engineering Research Council of Canada. 375 376 Smart Grids The load forecast generally provides annual peak values for the whole system, each region, and individual substations. Loads at substations (buses) never reach their peaks at the same point in time. This phenomenon is called noncoincidence. A system analysis, such as a power low study, is often conducted at the system peak, or a regional peak, or a percentage level of the annual system peak, which may be a seasonal peak. It is a dificult task to accurately determine the load at each bus that corresponds to the annual system or regional peak or the percentage level of the peak load. Because hourly load curve data at all individual buses are generally not available, the majority of utilities have used a so-called worst-case approach by ignoring the noncoincidence among bus loads, such that the annual peaks of all bus loads are assumed to appear at the same time as the annual system peak in power low studies. Such an assumption leads to overestimation of loading levels on the transmission network, which in turn most likely results in overinvestment in system planning or overacting in system operation. For a system with its annual peak in winter, the system summer peak is usually assumed to be a percentage (such as 60%–80%) of the system winter peak, and the same percentage for the system is applied to all buses. In other words, the bus loads corresponding to the system summer peak are obtained by proportionally scaling down the bus loads at the system winter peak. If a system has its annual peak in summer, a similar approach is used to calculate the bus loads at the system winter peak. This proportional approach to power lows at a seasonal system peak may lead to underestimation of loading levels on some network branches and overestimation of loading levels on others. An underestimation of loading levels on branches results in a system risk in planning and operation. Some utilities may consider the coincidence among several regional loads if they have collected chronological regional load curve data. However, this is not suficient, as the coincidence among substations has not been captured. In recent years, many utilities have implemented smart metering projects. Smart meters accurately record time-varying loads at end users or on feeders at intervals of seconds or minutes. This enables chronological load curve data to be automatically acquired at all buses (substations). With load curve data for system or region and at all buses, the bus load coincidence factors (BLCFs) at all bus loads can be easily estimated to accurately calculate bus loads corresponding to any reference point in the system or regional load curve. The concept of BLCFs will be explained in detail in Section 16.3. Even for smart meters, however, invalid load data in the process of information collection and transfer are still unavoidable. Missing or corrupted data can be caused by various factors, including meter problems, communication failures, equipment outages, lost data, and unknown factors. Also, unexpected interruption or shutdown of power due to strikes, unscheduled maintenance, and temporary closure of production lines can produce signiicant, irregular deviations from the usual load pattern, resulting in the related load data records being unrepresentative of actual usage patterns. Poor quality of load curve data due to invalid information can cause errors in data analysis for capturing the coincidence relationship among bus loads. It is important to identify and repair any invalid data. Currently, most utilities handle invalid data manually in an ad hoc manner. In practice, this approach does not work, particularly for a considerable amount of time-varying information from smart meters, as it is impossible to handle a huge data pool using a manual process. 377 Load Data Cleansing and Bus Load Coincidence Factors This chapter addresses two issues: (1) cleansing of load curve data; and (2) calculation of BLCFs. Load curve data cleansing is the prerequisite for calculating accurate BLCFs, while the BLCF is an application of chronological cleansed load curve data. Cleansed load curve data also have many other applications in data analytics of smart grids. Accurate BLCFs are the essential input data for system analyses of smart grids, including power low studies, stability assessment, network loss calculation, and system reliability evaluation. An actual example is given to demonstrate the application of BLCFs in system power low studies. 16.2 LOAD CURVE DATA CLEANSING We can calculate accurate BLCFs by using cleansed load curves. This section discusses the basic concepts and method for cleansing outliers in load curves. 16.2.1 BasIc concePts We call invalid data in a load curve “outliers,” including missing, corrupted, and unrepresentative data due to various different causes. A load curve is a time series with the y-axis representing power consumption and the x-axis representing time. Theoretically, outlier detection in a time series has been a topic in data mining [5,6] and statistics [7]. Outliers can be classiied into two categories: y-outliers and x-outliers. y-Outliers refer to invalid y-axis values compared with the behaviors in a local neighborhood [8]. Figure 16.1 shows two examples of y-outliers, one being a localized spike and the other being a localized dip. Load curve data exhibit some loose form of periodicity (e.g., daily, weekly, monthly, seasonal, and yearly). The term loose means that the actual data values can be different, but the trend of the data repeats itself regularly at some interval. x-Outliers refer to abnormal or unrepresentative data that may occur as a deviation from such periodicities [9]. x-Outliers are caused by random events such as a malfunction of data metering or transfer systems, outages, unexpected full or partial shutdown of production lines, unscheduled strikes, temporary weather changes, and so on. Such events are unlikely to occur again in other periods, and thus are not representative of regular patterns of load curves. Figure 16.2 shows a load curve with weekly periodicity (high weekdays and low weekends), in which part of the data in the irst week is an x outlier to this periodicity, although the values in this part of the data are still within the range of the normal pattern. Load (kW) 25 20 15 10 0 FIGURE 16.1 15 30 Time (h) Examples of y-outliers in a load curve. 45 378 Smart Grids 6 Load (MW) 5 4 3 2 1 0 9 Nov 14 Nov 19 Nov 24 Nov 29 Nov 04 Dec 09 Dec Time (date) FIGURE 16.2 Examples of x-outliers in a load curve. 16.2.2 method For cleansIng outlIers The method for cleansing outliers includes the following steps: (1) load curve data are represented using a smoothing curve, which captures the general trend of the data; (2) a conidence interval on the smoothing curve is built to identify localized y-outliers; (3) the smoothing curve is modeled by a sequence of valleys and peaks, called ∪ shapes and ∩ shapes, which are the potential places where x-outliers could occur; (4) x-outliers are identiied as the valleys, peaks, or both that do not repeat; and (5) the detected outliers are approximately repaired [8,9,10]. 16.2.2.1 Kernel Function Representation of Load Curve The irst step in detecting outliers is to model the intrinsic patterns of load curve data by a smoothing curve. This step can be performed by standard smoothing techniques. Given a time series of load curve data, T = {(ti, yi)} (i = 1, …, n) where yi is the load value at the time point ti, the regression relationship of the data can be modeled by a continuous function [10]: yi = m ( ti ) + εi ( i = 1,..., n ) (16.1) This function is composed of the regression function m and the observation error εi. The estimated load value in the regression function at time t is modeled in a form of nonparametric regression by 1 ⌢ m (t ) = n n ∑ w (t ) y i i (16.2) i =1 where wi(t) (i = 1, …, n) denotes a sequence of weights. Equation 16.2 is called the smoothing curve, in which wi(t) is the weight of the point (ti, yi) at time t in the estimation of the smoothing value. The estimation is neighborhood based in the sense Load Data Cleansing and Bus Load Coincidence Factors 379 that the further ti is from t, the less weight is given to the point (ti, yi). In the kernel smoothing, wi(t) is given by Kernh ( t − ti ) ⌢ fh ( t ) wi ( t ) = (16.3) where 1 t Kern h h Kernh ( t ) = (16.4) is the kernel with the scale factor h. By using the Rosenblatt–Parzen kernel density estimator [11] of the density of t ⌢ fh ( t ) = n −1 n ∑ Kern ( t − t ) h i (16.5) i =1 the Nadaraya–Watson estimator [11] for Equation 16.2 is obtained as follows: ∑ ∑ n ⌢ m (t ) = i =1 n Kernh ( t − ti ) yi i =1 Kernh ( t − ti ) (16.6) The shape of the kernel weights is determined by the function Kern, whereas the size of weights is parameterized by h, which is called bandwidth and referred to below as smoothing parameter. A larger h corresponds to a smoother curve. The kernel function Kern can be chosen to be the normal probability density function [11], that is, Kern ( t ) = 1 −(1/ 2 )t 2 e 2π (16.7) 16.2.2.2 Conidence Interval for Identifying y-Outliers With a smoothing curve expressed by kernel functions, a localized y-axis outlier can be easily identiied by building a conidence interval for normal load data on the smoothing curve. An observation within the conidence interval is considered normal and an observation outside the conidence interval is considered invalid data. The error term ɛi in Equation 16.1 is assumed to be normally and independently distributed with the mean of zero and constant variance σ2. Under this assumption, the conidence interval of a load data point can be computed by its estimated value plus or minus a multiple of the error. The estimated error si(est) is given by [12] 380 Smart Grids si ( est ) = MSE + si2 ( yˆi ) (16.8) where MSE is the error mean square: 1 MSE = n−d n ∑ ( y − yˆ ) i 2 i (16.9) i =1 and si2 ( y^ i ) is the sampling variance of load at time point ti and is given by si2 ( yˆi ) = WW T × MSE (16.10) where W is called the hat matrix and is calculated by [13] W1 ( t1 ) W2 ( t1 ) 1 W1 ( t2 ) W2 ( t2 ) W= ⋮ n ⋮ W1 ( t n ) W2 ( t n ) ⋯ Wn ( t1 ) ⋯ Wn ( t2 ) ⋮ ⋯ ⋯ Wn ( t n ) (16.11) where Wi(tj) corresponds to the weight of observation at time ti for estimating the load value at time tj. Note that d in Equation 16.9 is the number of degrees of freedom of the itted load curve data and can be computed by the trace of the hat matrix W. For a given signiicance level α, the conidence interval at time ti is estimated by yˆi − z1− α / 2 × si ( est ) , yˆi + z1− α / 2 × si ( est ) (16.12) where z1–α/2 is the 100(1 – α/2) percentile of the standard normal distribution. For example, if the 0.05 signiicance level (α = 0.05) is chosen, the corresponding z1–α/2 is 1.96. This means that a normal observation would fall in the conidence interval given by Equation 16.12 with a probability of 95%, or, equivalently, that the probability that a data point located outside the interval is invalid is 95%. 16.2.2.3 Candidates for x-Outliers An x outlier has an unusual trend against the periodicity. With the trend of the load curve being modeled by the smoothing curve, an x outlier tends to occur at a “valley” or a “peak” of the smoothing curve, where the smoothing curve has local minimal or maximal values. To formally deine the location of such valleys and peaks, we irst introduce some terminologies. The slope at time ti for a smoothing curve {(ti , m^ i )}in=1 is deined by Δmi/Δti, where ∆mi = mˆ i − mˆ i −1 , Δti = ti − ti–1, for 2 ≤ i ≤ n. A time t in an interval is called a steep time point if the absolute value of the slope at time t is the maximum in the interval. In other words, at a steep time point, the load increases or decreases at the maximum rate in the interval. Note that there could be more than one steep time point 381 Load Data Cleansing and Bus Load Coincidence Factors in an interval. An interval [a, b] is maximal-decreasing if the slope at every time point in [a, b] is ≤0 and any interval containing [a, b] has at least one point with a positive slope. For two time points c and cʹ in a maximal-decreasing interval [a, b], [c, b] is convex-decreasing if c is the rightmost steep time point in [a, b], and [a, cʹ] is concave-decreasing if cʹ is the leftmost steep time point in [a, b]. In Figure 16.3, the numbers on the smoothing curve are the slopes of the smoothing curve, and the horizontal axis is time; [t1, t5] is a maximal-decreasing interval; and t3 and t4 are two steep time points in [t1, t5]. Since t4 is the rightmost steep time point in [t1, t5], [t4, t5] is a convex-decreasing interval, but [t3, t5] is not, while [t1, t3] is a concave-decreasing interval. Similarly, we can deine maximal-increasing intervals, concave-increasing intervals, and convex-increasing intervals. In Figure 16.3, [t6, t10] is a maximal-increasing interval, t8 and t9 are steep time points, [t6, t8] is a convex-increasing interval because t8 is the leftmost steep time point in [t6, t10], and [t9, t10] is a concave-increasing interval. The following deinitions formalize the notions of valleys and peaks. Deinition of ∪ shape: For a smoothing curve, a ∪ shape is a subcurve T∪ = {(t p , m^ p )}pj = i such that, for some k with i ≤ k ≤ j, [i, k] is a convex-decreasing interval and [k + 1, j] is a convex-increasing interval. Deinition of ∩ shape: For a smoothing curve, a ∩ shape is a subcurve T∩ = {(t p , m^ p )}pj = i such that, for some k with i ≤ k ≤ j, [i, k] is a concave-increasing interval and [k + 1, j] is a concave-decreasing interval. In Figure 16.3, the curve [t4, t8] is a ∪ shape formed by a convex-decreasing interval [t4, t5] and a convex-increasing interval [t6, t8]. The curve [t9, t12] is a ∩ shape formed by a concave-increasing interval [t9, t10] and a concave-decreasing interval [t11, t12]. Notice that the adjacent ∪ shape and ∩ shape overlap at one point at most. Let us see the adjacent ∪ shape [t4, t8] and ∩ shape [t9, t12] in Figure 16.3. The ∪ shape [t4, t8] must end at the leftmost steep time point t8 and the ∩ shape [t9, t12] must start at the rightmost steep time point t9. Intuitively, ∪ shapes and ∩ shapes capture the regions on the smoothing curve where the raw load curve has large drops and large rises. Such regions are the potential places where x-outliers may occur. Therefore, ∪ shapes and ∩ shapes are candidate regions for outliers. All candidate regions can be found by computing ∪ shapes and ∩ shapes following their deinitions. m̂ –0.1 1 –1 3 –3 –2 FIGURE 16.3 t2 t3 –1 3 –3 t1 –0.1 t4 t5 0.1 t6 2 t7 t8 t9 t10 t11 t12 t Explanation of the terminologies in smoothing curve [t1, t12]. 382 Smart Grids 16.2.2.4 Identifying x-Outliers Candidate regions with ∪ shapes and ∩ shapes are extracted based on local neighborhood information (i.e., valleys and peaks). Whether a candidate region contains a real outlier depends on whether the data in the region deviates from the periodicity. A candidate is not a real outlier if similar load data occur regularly in other periods according to the periodicity. The similarity should take into account background noise and time shifting of periodicity. To identify all outliers, we consider every candidate region r found in the previous step. Let C* denote the portion of the raw load curve data contained in r. We want to emphasize that C* contains raw data in the load curve, not the data in the smoothing curve. If the data in C* occur approximately in the corresponding regions in different periods, C* is not a real outlier but a part of the periodicity. To differentiate this, we extract all the subload curves C1, C2, …, Ck, where each Ci is the portion of the raw load curve in the corresponding region of C* in the ith period. If C* is “similar” to the majority of C1, C2, …, Ck, C* is considered normal; otherwise, C* is considered a real outlier. The remaining question is how to measure the similarity between two subcurves C* and Ci with the same length. There are two considerations in choosing the similarity measure. First, the similarity measure should be less sensitive to background noise. For example, the two load curves in Figure 16.4 should be considered similar despite some variability at the irst peak due to background noise. Second, the similarity measure should be less sensitive to time shifting and stretching commonly observed in load curve data. For example, load curves TA and TB in Figure 16.5 should be considered similar despite minor time shifting and stretching. The Euclidean distance that is commonly used is not suitable for our purpose because it is sensitive to time shifting and stretching. If the Euclidean distance were used, the two curves TA and TB in Figure 16.5 would be recognized as dissimilar. The dynamic time warping distance for two sequences [14] is not suitable either, since it would pair up all points of two sequences in comparison, making it impossible to skip noisy points. What we need is a similarity measure that will examine a small neighborhood in search of matching points and skip noisy points. For this purpose, the longest common subsequence (LCSS) [15] concept can be adopted. LCSS is a coarse-grained similarity measure in the sense that it measures similarity in terms of “trends” y t FIGURE 16.4 Two similar load curves with noise. 383 Load Data Cleansing and Bus Load Coincidence Factors y TA TB t FIGURE 16.5 Two similar load curves with time shifting and stretching. instead of exact points. Below, we describe how LCSS is extended to measure the similarity of load curves. Given two subload curves A = 〈a1, a2, …, am〉 and B = 〈b1, b2, …, bn〉, which correspond to C* and Ci, we want to ind the LCSS common to both A and B. The idea is as follows. To allow time shifting and stretching, ai and bj that are within some time proximity are examined for matching. If these load points are similar, they are considered as a match and are kept. Dissimilar values in one or both load curves are dropped. Mathematically, given an integer δ and a real value ɛ, the cumulative similarity Si,j(A, B) or Si,j is deined as follows: 0, Si, j = 1 + Si −1, j −1 max ( Si, j −1, Si −1, j ) if i = 0 ∨ j = 0 if ai − b j ≤ ε ∧ i − j ≤ δ (16.13) otherwise In Equation 16.13, the irst if statement does the initialization for the shortest preix. The second if statement builds the similarity recursively: if |ai − bj| ≤ ɛ and if ai and bj are close enough in time, that is, |i − j| ≤ δ, ai and bj are matched and the similarity is incremented. Note that ɛ represents a tolerance of noise in load value and δ represents a tolerance of time shifting and stretching. Let |A| and |B| be the length of A and B, respectively. The LCSS similarity of A and B is given by γ ( δ, ε, A, B ) = S A, B ( min A , B ) (16.14) where S|A|,|B| is the length of the common subsequence to both A and B, which is cumulatively computed by Equation 16.13. For a user-speciied threshold θ, we say that A and B are similar if γ ( δ, ε, A, B ) ≥ θ (16.15) 384 16.2.3 Smart Grids rePaIrIng outlIers Detected outliers should be repaired before the load curve can be used for calculating BLCFs or other applications. Suppose that we want to repair an outlier C*. Some valid data must be selected to replace the load data of C*. The replacement data can be derived from the normal data in the corresponding time interval in other periods with an adjustment for the increasing or decreasing long-term trends over time. This is expressed as the following multiplicative model [16]: Y ( ti ) = T ( ti ) × R ( ti ) (16.16) where Y(ti) represents the value that will replace the abnormal value at a time ti belonging to an outlier, T(ti) represents the load value in the long-term trend, and R(ti) represents the periodic index, that is, how much the load curve deviates from the long-term trend at time ti. It is possible to estimate T(ti) by the smoothing curve deined in Equation 16.2 with an appropriate smoothness level (bandwidth). However, in the presence of outliers, the smoothing curve may contain invalid data around the outliers to be replaced. To address this problem, all outliers in the load curve are replaced, irst, by the average of the data at the corresponding time in the previous and following periods. If load points at these times are also outliers themselves, the corresponding data from earlier and later periods will be examined until normal data are obtained. After illing in such average values for all outliers, a new smoothing curve is regenerated using Equation 16.2 and is used to estimate T(ti). The periodic index R(ti) for a time ti belonging to an outlier is estimated by the average of the periodic indexes at the corresponding time of its previous and following periods, that is, R ( ti ) = 1 2 ( R ( t − l ) + R ( t + l )) i i (16.17) where l is the length of the periodicity. If the data at the previous and next periods are outliers, earlier and later periods are examined until normal data are obtained. Note that, for a time ti not belonging to an outlier, the periodic index at ti is computed by its deinition: R ( ti ) = yi T ( ti ) (16.18) where yi is the load at time ti. After T(ti) and R(ti) for an outlier are obtained, Equation 16.16 is used to produce the replacement value Y(ti). 16.2.4 selectIon oF Parameters The method described above uses several parameters: the smoothing parameter h (see Equation 16.4), the load stretching threshold ɛ and the time stretching threshold δ (see 385 Load Data Cleansing and Bus Load Coincidence Factors Smoothing curve Detected outlier 30 Load (MW) 25 20 15 10 5 0 2005 FIGURE 16.6 2006 2007 2008 Time (year) 2009 2010 An actual load curve, the smoothing curve, and identiied x-outliers. Equation 16.13), and the LCSS similarity threshold θ (see Equation 16.15). How should the values of these parameters be set? One approach is to use some statistically “optimal” setting, such as the optimal smoothing parameter [17]. In applications where the user has background knowledge, the user often desires to have control over a small number of settings. A practical approach is to provide a mechanism that helps the user to identify a proper setting of parameters in such a case. We use the smoothing parameter h as an example below, and a similar approach can be applied to other parameters. A larger h produces a smoother smoothing curve, and thus models the load data in less detail. In practice, we do not have to make a choice in advance. A software tool with a user-friendly interface has been developed, which allows the user to slide a bar for the smoothing parameter and displays the smoothing curve and the identiied outliers to the user interactively. Based on visual inspection of the raw data, smoothing curve, and identiied outliers, the user can either accept the results or slide the bar again for a different choice of h and get a display of the new smoothing curve and outliers based on the new choice of h. This is analogous to sliding the bar at the lower right of the Word window for setting the font size of a document. With such sliding bars, a user often quickly converges to what he or she considers to be the best setting. As an experiment, the method was applied to an actual industrial load curve in the system of BC Hydro, Canada. This load curve includes hourly MW loads for the 6 years from October 2004 to October 2010, with 24 × 365 × 6 = 52,560 observed records. Figure 16.6 shows the raw load records, the smoothing curve (by the bold line) obtained using a speciic smoothing parameter h [8], and the x-outliers (marked by each rectangle) detected by the proposed method. It can be seen that the smoothing curve models the load curve properly and there is no false-positive or falsenegative error in detecting the x-outliers. 16.3 BUS LOAD COINCIDENCE FACTORS Once outliers in historical load curves are cleansed, calculating BLCFs using load curve data is straightforward. This section illustrates the basic concept and procedure 386 Smart Grids of calculating BLCF, and Section 16.4 will present an example to demonstrate the application of BLCF to power system analysis. 16.3.1 BasIc concePt oF BlcF Once load curves at all considered buses are cleansed, it is straightforward to calculate the BLCFs. The BLCF is an important concept in power system analyses, particularly in power low studies [4]. The load forecast generally provides annual peak values for the whole system, each region, and individual substations. As mentioned in Section 16.1, loads at substations never reach their peaks at the same point in time. To perform a power low study at the annual system or regional peak, or at a percentage level of the annual peak, which may be a seasonal peak or valley, we must calculate bus loads corresponding to the peak or the level of the peak. The BLCFs at the buses are used for this purpose. The concept of BLCF is illustrated using Figure 16.7. For simplicity, only two substation load curves are considered. The total load curve, which is the sum of the two substation load curves, can be viewed as the system (or regional) load curve. Points A, E, and D are the peaks of the system and two substation load curves, respectively. Apparently, they do not occur at the same time. Points C and B represent the load points of the two substations corresponding to the system (or regional) peak. The BLCFs of the two substations are deined as BLCF1 = LC LE (16.19) BLCF2 = LB LD (16.20) where LC and LE are the loads at Points C and E in the load curve 1 for Substation 1, and L B and L D are the loads at Points B and D in the load curve 2 for Substation 2. 90 80 Load 1 Load 2 Total A Load (MW) 70 60 50 40 B D C 30 E 20 10 0 Time (h) FIGURE 16.7 Concept of BLCF. Load Data Cleansing and Bus Load Coincidence Factors 387 Obviously, cleansed load curve data are important to accurately calculate BLCF. For instance, if the load around Point B, C, D, or E is corrupted, we will get a distorted BLCF value. Even if the load at Point B or C is valid data, a reduced L D or LE will overestimate the BLCF value and an elevated L D or LE will underestimate the BLCF value. Using the example shown in Figure 16.7, we deined the BLCF corresponding to the system peak. For a given load point in a system or regional load curve, the corresponding time point on the x-axis is called the reference point. In this sense, BLCFs can be calculated for any different reference points on a system or regional load curve (such as the annual peak, winter or summer peak, winter or summer valley). BLCFs can be also calculated for different load categories (such as industrial, residential, or commercial customers or their combinations) using a similar concept. Apparently, BLCFs vary dynamically from year to year, as the load curve data change every year. BLCFs should be estimated using multiple years of load data and updated each year. There are uncertainties around load data points. The system or regional peak is different in value and moves around in different years, and various periodicities in a load curve may be shifted. To tackle the uncertainties, instead of using the load value at each single point (such as A, B, C, D, or E), the average of a certain number of loads around each point should be used. Once the BLCFs are estimated, the bus loads corresponding to any reference point on a system or regional load curve can be calculated using the BLCF values for that reference point and the forecasted peaks of substation loads. 16.3.2 Procedure For calculatIng BlcF In the procedure described below, the annual system peak in a 1-year period is used as a reference point to indicate how to estimate the BLCF values and calculate bus loads corresponding to the reference point. The procedure for estimating the BLCFs corresponding to other reference points (a percentage level of the annual system peak or the maximum/minimum load value in a speciied period, such as winter or summer) in a system or region is similar. The procedure includes the following steps [18]: 1. The invalid data for yearly load curves at all considered substations are cleansed using the method given in Section 16.2, including identiication and repairing of invalid data. 2. All the yearly substation load curves are summed up to obtain the yearly system load curve. The maximum hourly load value (peak) in the yearly system load curve is determined and is assumed to be X MW. The corresponding time point on the x-axis is called the reference point. 3. N hourly load points on the system load curve that are closest to the X MW in the load values are determined, where N (such as 24 or 48) is a proper number for handling the uncertainty of the system peak. In the case where the annual system peak is selected as the reference point, the N hourly load values include the system peak, itself being X MW, with others being smaller than or equal to X MW. In other cases in which a nonannual system 388 Smart Grids peak is selected as the reference point, some of the N selected hourly load values may be larger than X MW but others may be smaller than X MW. 4. M annual largest hourly load values on the yearly load curve of each individual substation are determined, where M is a proper number for handling the uncertainty of the substation load peak and may or may not be equal to N, depending on a judgment on the degree of uncertainty of the substation load peak. Let Lki(P) denote the MW load at the kth point of the M hourly load points in the ith substation load curve. 5. We locate the N points on each yearly substation load curve that correspond to the same time points as the N load points on the yearly system load curve, which have been determined in Step 3. Let Lji denote the MW load at the jth point of the N selected hourly load points in the ith substation load curve. The BLCF corresponding to the selected reference point for the ith substation is estimated by BLCFi = 1 N N ∑ j =1 L ji ∑ k=M k =1 Lki ( P ) M (16.21) 6. There are always some differences in the yearly load curves of a substation in different years. In many cases, the differences may be relatively small, as load customers supplied by the substation usually maintain their electricity consumption behaviors. In other cases, the differences may be relatively large because the electricity consumption behaviors of some customers change over the years. To deal with the possible changes, it is better to repeat Steps 1–5 for more than 1 year’s load curve data. On the other hand, we should not use too many years’ data, since cumulative changes may be so signiicant that data from previous years are no longer representative of the most recent pattern. We suggest considering a time frame of the past 3 years for obtaining three BLCF values for each substation. There are three possibilities, as follows: a. Only 1 year’s load curve data are available. In this case, the BLCF for this year is used. b. Two years’ load curve data are available. In this case, the difference between the two BLCF values is calculated as a percentage with respect to the larger value. i. If the difference is smaller than or equal to a threshold (such as 5%–10%), the average of the two BLCF values is used as the inal BLCF. ii. If the difference is larger than the threshold, the BLCF from the most recent year’s load curve is used as the inal BLCF. c. Three years’ load curve data are available. In this case, the differences between each pair of the three BLCF values are calculated as a percentage with respect to the largest value. i. If all the three differences are smaller than the threshold, the average of the three BLCF values is used as the inal BLCF. Load Data Cleansing and Bus Load Coincidence Factors 389 ii. If two of the three differences are smaller than the threshold, the average of the larger two BLCF values is used as the inal BLCF. iii. If only one of the three differences is smaller than the threshold, the average of the two BLCF values corresponding to this difference is used as the inal BLCF. iv. If all the three differences are larger than the threshold, the BLCF from the most recent year’s load curve is used as the inal BLCF. 7. The MW load of each substation corresponding to the selected reference point (such as the annual system peak, or a percentage level of the annual system peak, or the maximum or minimum load value in a selected period) can be calculated using the BLCF value of each substation and the forecast peak load for each substation in the considered year, by Li ( MW ) = Li ( F ) × BLCFi (16.22) where Li(MW) is the MW load at the ith substation (bus) corresponding to the reference point and Li(F) is the forecasted peak load for the ith substation. It should be pointed out that the word “system” in the above description can refer to a whole system, a region or area, a subregion or subarea, or a substation group. 16.4 APPLICATION OF BLCF TO SYSTEM ANALYSIS A real-life case is used in this section to demonstrate the application of BLCF to system analysis. 16.4.1 studIed system and load curve data Figure 16.8 shows a transmission network supplying power to a subarea of BC Hydro in Canada. The subarea includes three substations named YVR, SEA, and RIM. The power is supplied from the upstream substation KI2 through two 60 kV transmission circuits named 60L43 and 60L44. Each circuit includes three sections marked by A, B, and C. Sections A and B are overhead lines and Section C is an underground cable. The MVA ratings of the three sections of each circuit in winter and summer are listed in Table 16.1. The cleansed hourly load curves at the three substations in 2006/2007 are shown in Figure 16.9. By summing up the loads of the three substations at every hour, the total load curve of the subarea is created, as shown in Figure 16.10, where the summer and winter peak periods are framed by the two dashed line rectangles. This subarea has its annual peak in winter, which appears on November 29, at 5:00 pm, for this particular year. The maximum loading levels on the two circuits (60L43 and 60L44) occur at the same time point as the subarea peak. The power lows at the maximum loading levels on November 29, 2006 at 5:00 p.m. are displayed in Figure 16.11. It can be seen that the actual maximum loading levels in the normal state are as follows: In the normal state, 60L43 loading = 56.5/156 = 36.2% In the normal state, 60L44 loading = 48.2/156 = 30.9% In N−1 contingency states, the maximum loading level can be calculated by contingency analysis. It is 105.9/156 = 67.9%. 390 Smart Grids KI2 60L43C 60L43A YVR 60L44A SEA 60L44C 60L43B 60L44B RIM FIGURE 16.8 The subarea power supply network. TABLE 16.1 Rating of the Two Circuits Circuit Section 60L43A and 60L44A 60L43B and 60L44B 60L43C and 60L44C 16.4.2 Winter Rating (MVA) (10°C) 156 100 68.5 Summer Rating (MVA) (30°C) 121 78 63 Bus load coIncIdence Factors at the three suBstatIons The BLCFs are calculated using the method in Section 16.3 and the historical load curve data at the three substations. The BLCF values obtained using the subarea winter peak (which is the annual peak in this example) and summer peak as the reference points are given in Tables 16.2 and 16.3, together with the forecast annual (winter) peak loads in 2017/2018 and the calculated coincidence bus loads in 2017/2018. It is important to appreciate that a coincident load for the subarea summer peak as a reference point should be, and is still, calculated using the forecasted annual peak (i.e., the winter peak in this case) multiplied by its summer BLCF, since a BLCF is always deined as a load at the reference point divided by its annual peak. To make a comparison, the winter and summer peak loads of the three substations in 2006/2007 are 391 07/15/06 08/05/06 08/26/06 10/07/06 10/28/06 12/09/06 12/30/06 01/20/07 02/10/07 03/03/07 06/24/06 07/15/06 08/05/06 08/26/06 09/16/06 10/07/06 10/28/06 11/18/06 12/09/06 12/30/06 01/20/07 02/10/07 03/03/07 03/24/07 06/24/06 07/15/06 08/05/06 08/26/06 09/16/06 10/07/06 10/28/06 11/18/06 12/09/06 12/30/06 01/20/07 02/10/07 03/03/07 03/24/07 03/24/07 06/03/06 06/03/06 06/03/06 11/18/06 05/13/06 05/13/06 05/13/06 09/16/06 04/22/06 04/22/06 04/22/06 06/24/06 04/01/06 04/01/06 16 14 12 10 8 6 4 2 0 04/01/06 RIM load (MW) 90 80 70 60 50 40 30 20 10 0 SEA load (MW) Load Data Cleansing and Bus Load Coincidence Factors 14 YVR load (MW) 12 10 8 6 4 2 0 Time stamp (h) FIGURE 16.9 The cleansed hourly load curves at the three substations. also given in the two tables. Note that the load at the RIM substation includes two components: distribution and industrial customers. The industrial customer at RIM will not be in service until 2017/2018, and therefore its historical load curve data are not available for calculating the BLCF. In this case, the BLCF can be assumed to be 1.0. Such an assumption creates a relatively high loading level, which can lead to a more secure decision in system operation and planning. A small part of the distribution load at RIM will be transferred to another substation by 2017/2018. 392 Smart Grids Total load (MW) 120 100 80 60 40 20 03/24/07 03/03/07 02/10/07 01/20/07 12/30/06 12/09/06 11/18/06 10/28/06 10/07/06 09/16/06 08/26/06 08/05/06 07/15/06 06/24/06 06/03/06 05/13/06 04/22/06 04/01/06 0 Time stamp (h) FIGURE 16.10 The load curve of the subarea. 1.061 63.7 T1 9.8 892 SEA 12 0.0 T2 0.0 6.0 15.8 15.8 56.5 T1 1.033 13.0 56.5 6.8 T3 6.8 6.8 48.1 692 SEA 60B1 634 1.061 63.7 RIM 60B3 1.054 63.2 48.2 633 RIM 60B1 1.053 63.2 1.064 63.9 T1 SW 0.0 81.4 16.6 T3 833 RIM 12 FIGURE 16.11 T2 1.061 63.7 40.5 40.5 12.8 2.4 693 SEA 60B2 41.1 41.1 698 YVR 60 9.8 9.8 0.0 1.050 13.2 The power lows at the actual annual maximum loading levels in 2006/2007. It can be observed that the total load in the subarea is increased from 105.8 MW in 2006/2007 to 123 MW in 2017/2018. This is translated into an average annual growth rate of 1.5% (1.56 MW/year). The difference between the forecasted (noncoincidental) total load and the coincidental total load is 123−119 = 4.0 MW. This is equivalent to the load increase percentage in about 2 years. In other words, if the coincidental loads are used for a planning purpose, this may lead to approximately 2 years’ delay of the system reinforcement required by the load growth in this subarea. This observation can be veriied by the power low studies given in the next subsection. 393 Load Data Cleansing and Bus Load Coincidence Factors TABLE 16.2 BLCFs and Coincidence Loads for the Winter Peak at the Three Substations Substation RIM—distribution load RIM—industrial load SEA YVR Total a Actual Winter Peak Load in 2006/2007 (MVA) Forecast Annual Peak in 2017/2018 (MVA) Winter BLCF Coincidental Load in 2017/2018 (MVA) 83.0 78.2 0.9882 77.3 7.0 N/A (1.0000) 7.0 —a 13.0 9.8 105.8 23.3 14.5 123.0 0.9361 0.8863 — 21.8 12.9 119.0 —, The industrial customer is not yet in service and the load proile is therefore not available. TABLE 16.3 BLCFs and Coincidence Loads for the Summer Peak at the Three Substations Substation RIM—distribution load RIM—industrial load SEA YVR Total a 16.4.3 Actual Summer Peak Load in 2006/2007 (MVA) Forecast Annual Peak in 2017/2018 (MVA) Summer BLCF Coincidental Load in 2017/2018 (MVA) 65.4 78.2 0.8331 65.1 —a 7.0 N/A (1.0000) 7.0 12.6 11.6 89.6 23.3 14.5 123.0 0.9359 0.8344 — 21.8 12.1 106.0 —, The industrial customer is not yet in service and the load proile is therefore not available. PoWer FloW analysIs usIng the Bus load coIncIdence Factors The power low studies for the subarea network are performed using the noncoincidental and coincidental substation loads in 2017/2018. The following four cases are considered: Case 1: Noncoincidental winter peak loads, which refer to the forecasted annual peaks at the given three substations. Case 2: Noncoincidental summer peak loads. For utilities having the annual peak in winter, although load forecast departments at the utilities may also 394 Smart Grids provide the forecasted summer peak for the whole system (or areas) in addition to the forecasted winter peaks for the whole system (areas) and all substations, they generally do not provide the forecasted summer peaks for individual substations. In the existing practice of utilities, a common approach is to calculate substation loads at the system summer peak point by proportionally scaling down from the annual substation peaks using the same ratio of the system (area) winter peak with regard to the system (area) summer peak. In this case, the noncoincidental summer peak loads are obtained using this approach. Case 3: Coincident winter peak loads. In this case, the substation loads are calculated using their BLCFs corresponding to the winter (annual) subarea peak. Case 4: Coincident summer peak loads. In this case, the substation loads are calculated using their BLCFs corresponding to the summer subarea peak. The loading levels in the four cases are summarized in Table 16.4. The power lows for the four cases in the normal states (no component outage) are shown in Figures 16.12 through 16.15, respectively. The following observations can be made: • The difference in the maximum loading level for N−1 contingency states between noncoincidental and coincidental winter peak cases is 2.7% (78.7%−76%). This is equivalent to a loading level increase in about 2 years, as the annual average loading level increase is approximately (118.6−105.9)/11/105.9 = 1.1%. In other words, using the coincidental loads in the power low analysis can delay the upgrade of the two circuits by approximately 2 years. This loading level analysis is consistent with the direct comparison between the forecasted and coincidental total loads, although their percentages are slightly different. TABLE 16.4 Loading Levels in the Four Cases Noncoincidence Loads 60L43 loading (normal state) 60L44 loading (normal state) Maximum loading (N−1 states) Coincidence Loads (Using BLCF) Case 1: Winter (Annual) Peak Case 2: Summer Peak Case 3: Winter (Annual) Peak Case 4: Summer Peak 66.5/156 = 42.6% 53.8/121 = 44.5% 63.8/156 = 40.9% 57.0/121 = 47.1% 54.2/156 = 34.7% 43.9/121 = 36.3% 52.9/156 = 33.9% 46.8/121 = 38.7% 122.7/156 = 78.7% 98.2/121 = 81.2% 118.6/156 = 76.0% 105.1/121 = 86.9% 395 Load Data Cleansing and Bus Load Coincidence Factors 1.059 63.5 T1 14.5 892 SEA 12 0.0 0.0 24.8 66.4 11.1 25.0 T2 66.5 1.059 63.5 T1 12.6 12.6 T3 12.1 54.1 692 SEA 60B1 634 1.060 RIM 60B3 63.6 1.022 12.9 1.057 63.4 54.2 38.1 41.6 21.5 9.2 T2 693 SEA 60B2 38.9 42.4 698 YVR 60 14.5 14.5 0.0 76.7 15.6 T1 SW 22.2 T3 833 RIM 12 1.062 633 63.7 RIM 60B1 1.057 63.4 1.059 13.3 FIGURE 16.12 The power lows in Case 1 (using the noncoincidental winter peak loads in 2017/2018). 1.1 63.8 T1 892 SEA 12 17.2 T2 11.6 T2 8.0 19.6 19.7 53.7 53.8 1.1 63.8 9.2 1.1 63.8 43.9 9.2 43.9 634 RIM 60B3 1.1 63.7 833 RIM 12 30.4 33.9 692 SEA 60B1 633 RIM 60B1 1.1 63.7 1.1 63.9 1.1 13.3 SW 22.0 1.0 13.0 9.2 34.6 T1 T3 31.1 7.4 0.0 693 SEA 60B2 61.4 12.5 11.6 11.6 0.0 698 YVR 60 FIGURE 16.13 The power lows in Case 2 (using the noncoincidental summer peak loads in 2017/2018). 396 Smart Grids 1.060 63.6 T1 892 SEA 12 T2 0.0 22.6 63.7 0.0 T1 11.8 11.8 692 SEA 60B1 1.023 12.9 1.061 63.6 T3 11.4 52.8 634 RIM 60B3 1.058 63.5 52.9 1.060 13.4 SW 22.2 75.8 15.6 1.063 633 RIM 60B1 63.8 1.058 63.5 T1 T3 833 RIM 12 FIGURE 16.14 2017/2018). T2 63.8 1.060 63.6 37.7 41.2 20.1 9.2 693 SEA 60B2 10.5 22.8 38.4 41.9 698 YVR 60 12.9 12.9 12.9 0.0 The power lows in Case 3 (using the coincidental winter peak loads in 1.063 63.8 T1 12.1 892 SEA 12 0.0 0.0 21.9 56.9 T1 1.027 12.9 11.8 11.8 692 SEA 60B1 1.064 63.8 T3 11.4 46.7 634 RIM 60B3 1.062 63.7 46.8 1.066 633 RIM 60B1 64.0 1.062 63.7 T1 SW 22.4 63.9 15.6 T3 833 RIM 12 FIGURE 16.15 2017/2018). T2 57.0 1.063 63.8 31.7 35.2 20.1 9.2 T2 693 SEA 60B2 10.5 22.1 32.4 35.9 698 YVR 60 12.1 12.1 0.0 1.066 13.4 The power lows in Case 4 (using the coincidental summer peak loads in Load Data Cleansing and Bus Load Coincidence Factors 397 • It is noticed that the difference between noncoincidence and coincidence loads for the winter peak is not signiicant in this particular example. However, such a difference may be quite large in other system cases. • It should be emphasized that the noncoincidental summer peak using a traditional load-scaling method could result in an overoptimistic case because the circuit loadings based on the coincidental summer peak case are higher. This observation is opposite to that in the winter peak. As a result, the future system reinforcement in this subarea is driven by summer constraints, and therefore a coincidental summer peak case should be developed and used to capture this concern. • By comparing the results in Cases 3 and 4, it is observed that the percentage loading level in the summer peak is higher than that in the winter peak in this subarea. This implies that the subarea is mainly constrained by the summer rating. This phenomenon can be intuitively judged from the high BLCFs in the summer peak and the high ratio of the summer peak to the winter peak (106/119 = 89.1%) (see Tables 16.2 and 16.3). 16.5 CONCLUSIONS Load curve data at substations play an important role in data analytics of smart grids. Implementation of smart meters makes it possible to record time-varying loads at end users or on feeders at intervals of seconds or minutes and to acquire chronological load curve data at all substations. Unfortunately, invalid load data cannot be avoided in the data collection and transfer process, even for smart meters. Therefore, load curve data cleansing is an essential task. This chapter presents a method for load curve data cleansing based on techniques used in data mining and statistics. The method not only effectively identiies both y-outliers and x-outliers but it also repairs the outliers. Cleansed load curve data are extremely useful information for system analyses and visualization in day-to-day operations and long-term planning. One straightforward application is the calculation of BLCFs at substations. The BLCFs produce an accurate coincidental bus load model in power low studies. The use of noncoincidental loads in the system analysis may lead to overestimation or underestimation of loading levels in a transmission system and in turn result in overinvestment or system insecurity risk. This chapter proposes an approach and a procedure for estimating BLCFs in power grids. An actual utility subarea system is used to demonstrate the application and beneits of using BLCFs. REFERENCES 1. IEEE Standard 399-1990, IEEE Recommended Practice for Power Systems Analysis, IEEE Press, New York, 1990. 2. L. L. Grigsby (editor-in-chief), The Electric Power Engineering Handbook, CRC Press and IEEE Press, Boca Raton, FL, 2001. 3. W. Li, Risk Assessment of Power Systems: Models, Methods, and Applications, IEEE Press and Wiley, New York, 2005. 398 Smart Grids 4. W. Li, Probabilistic Transmission System Planning, IEEE Press and Wiley, Hoboken, NJ, 2011. 5. E. Keogh, J. Lin, S. H. Lee, and H. V. Herle, Finding the most unusual time series subsequence: Algorithms and applications, Knowledge and Information Systems, 11(1), 1–27, 2006. 6. V. J. Hodge and J. Austin, A survey of outlier detection methodologies, Artiicial Intelligence Review, 22(2), 5–126, 2004. 7. V. Barnett and T. Lewis, Outliers in Statistical Data, 3rd edn., Wiley, Chichester, 1994. 8. J. Chen, W. Li, A. Lau, J. Cao, and K. Wang, Automated load curve data cleansing in power systems, IEEE PES Transactions on Smart Grid, 1(2), 213–221, 2010. 9. Z. Guo, W. Li, A. Lau, T. Inga-Rojas, and K. Wang, Detecting X-outliers in load curve data in power systems, IEEE Transactions on Power Systems, 27(2), 875–884, 2012. 10. Z. Guo, W. Li, A. Lau, T. Inga-Roja, and K. Wang, Trend-based periodicity detection for load curve data, in IEEE Power and Energy Society General Meeting, July 21–25, Vancouver, BC, 2013. 11. W. Hardle, Applied Nonparametric Regression, Cambridge University Press, New York, 1990. 12. M. H. Kutner, C. J. Nachtsheim, J. Neter, and W. Li, Applied Linear Statistical Models, 5th edn., McGraw-Hill/Irwin, New York, 2004. 13. J. O. Ramsay and B. W. Silverman, Functional Data Analysis, 2nd edn., Springer, New York, 2005. 14. B. Yi, H. Jagadish, and C. Faloutsos, Eficient retrieval of similar time sequences under time warping, in Proceedings of the 14th International Conference on Data Engineering, (ICDE), pp. 201–208, February 23–27, Orlando, FL, 1998. 15. M. Vlachos, G. Kollios, and D. Gunopulos, Discovering similar multi-dimensional trajectories, in Proceedings of the 18th International Conference on Data Engineering, (ICDE), pp. 673–684, 26 February–1 March, San Jose, CA, 2002. 16. D. M. Bourg, Excel Scientiic and Engineering Cookbook, O’Reilly, Sebastopol CA, 2006. 17. W. Hardle, M. Muller, S. Sperlich, and A. Werwatz, Nonparametric and Semi-parametric Models, Springer, Berlin, 2004. 18. W. Li, Architecture design and calculation method of load coincidence factor application, BCTC report, BCTC-SPPA-R012, April 1, 2007. 17 Smart Metering and Infrastructure Wenpeng Luan and Wenyuan Li CONTENTS 17.1 Introduction ..................................................................................................400 17.2 Smart Metering and Infrastructure Overview .............................................. 401 17.3 Major System Components of SMI .............................................................. 401 17.3.1 Smart Meters ....................................................................................402 17.3.2 Communication System ....................................................................405 17.3.2.1 Meshed Network Radio Frequency....................................405 17.3.2.2 Point-to-Point Radio Frequency.........................................406 17.3.2.3 Power Line Carrier/Broadband Power Line ......................406 17.3.2.4 Wireless Broadband ...........................................................407 17.3.2.5 Cellular Technologies ........................................................407 17.3.2.6 Satellite Technologies ........................................................408 17.3.2.7 Optical Fiber Technologies ................................................408 17.3.2.8 Summary............................................................................408 17.3.3 Meter Data Management System ......................................................408 17.3.4 Home Area Networks ....................................................................... 410 17.4 SMI Beneits ................................................................................................. 411 17.5 SMI and Smart Grid ..................................................................................... 411 17.6 SMI in Supporting System Analysis, Planning, and Operations.................. 412 17.6.1 Load Forecasting .............................................................................. 413 17.6.2 Distribution System State Estimation and Analysis ......................... 413 17.6.3 Outage Location and Response ........................................................ 414 17.6.4 Voltage and Var Optimization .......................................................... 414 17.6.5 Distribution Planning and Distribution Operation ........................... 415 17.6.6 Asset Management............................................................................ 416 17.7 Smart Meter Data Analytics ......................................................................... 416 17.7.1 Basic Concepts .................................................................................. 416 17.7.2 An Example of Smart Meter Data Analytics ................................... 417 17.8 Conclusions ................................................................................................... 419 References .............................................................................................................. 419 399 400 17.1 Smart Grids INTRODUCTION Billing is the main purpose of electricity metering. Utilities install meters to record the energy consumption (kWh) of their customers and bill them accordingly. The meter reading is normally taken by a meter reader who visits customer sites monthly and manually records the customers’ electricity usage. To take advantage of the advancement of communication technologies and to reduce the cost of meter reading, some utilities have implemented automated meter reading (AMR) technologies, which enable them to collect monthly billing data without physically reading the meter. These are often realized by a walk-by or driveby approach using one-way communication technologies such as mobile, radio, and phone-based systems. In recent years, smart metering and infrastructure (SMI), which is also often called advanced metering infrastructure (AMI) [1], has become one of the hottest topics within electric power utilities. SMI is not a tool to simply capture customer energy consumption every month or two, but an integrated hardware and software architecture that is capable of capturing near real-time or real-time consumption, demand, voltage, current, and other information. In other words, the system functionality of SMI goes far beyond just obtaining a monthly meter reading. Combining two-way communication and smart metering technologies that can record customers’ load proiles, SMI provides utilities with system-wide sensing and measuring capability. It not only enables utilities to apply time-of-use (TOU) rates for customers in order to achieve peak shaving to reduce capital investment, but it also enables many advanced functionalities for utilities, such as customer service, outage response, load forecast, demand-side management, and system planning. SMI has been considered a fundamental component of smart grid technology. The world’s largest smart meter deployment was undertaken by Enel SpA, Italy. It deployed smart meters to its entire customer base of over 27 million customers over a 5-year period from 2000 to 2005. It is worth mentioning that regulators have played an important role in driving SMI technology development and implementation. The Ontario Energy Board in Ontario, Canada, legislated in November 2005 that Ontario needed to install smart meters for each of its 5 million customers by 2010. Since the US Energy Policy Act of 2005 requires utilities to provide TOU service to any customer who requests it, many utilities in the United States have been actively pursuing SMI (or AMI) implementation. In 2006 and 2007, the California Public Utility Commission approved the SMI plans for its three main utilities PG&E, SCE, and SDG&E, which would make smart meters available to all electricity and gas customers in its jurisdiction by 2012. An injection of government funding in 2009 resulted in a rapid increase in the deployment of smart meters in North America. As stimulus funds, the American Recovery and Reinvestment Act of 2009 allocated $3.4 billion to smart grid projects and the Canadian government allocated $795 million to smart grid projects in 2009 through its Clean Energy Fund. It is estimated that smart meter deployments had reached about one in three households in the United States in 2012, and about 25 utilities in the United States have completed their system-wide integrations. North American utilities are expected to install 81 million smart meters by 2020. That Smart Metering and Infrastructure 401 would bring projected smart meter penetration to about 55% in North America [2]. Similar levels of SMI deployment can be seen in many other countries and regions, including Australia [3] and Europe. This chapter provides an overview of the SMI system and its components, and discusses its beneits and relation to the smart grid. It also summarizes the various applications of SMI in support of distribution system planning, analysis, and operations, and briely introduces the emerging smart meter data analytics with a real implementation example. 17.2 SMART METERING AND INFRASTRUCTURE OVERVIEW SMI is the totality of the systems and networks that are used to measure, collect, store, analyze, and use energy usage data. In other words, SMI includes smart meters and all other infrastructure components—hardware, software, and communication networks—that are needed to offer advanced capabilities. SMI covers the infrastructure not only from meters to the utility, but also from meters to customers, which enables every customer to analyze and use the energy metering data. SMI also makes energy usage data available to parties other than the utility in supporting the provision of demand response solutions. A typical SMI network employs a two-way communication system and smart metering technology. Instead of a monthly accumulated energy consumption recording, a smart meter records the customer’s consumption at preset intervals on a continuous basis. It communicates the customer’s load proile data to a central location, where the data are sorted and analyzed for a variety of purposes, such as customer billing, outage response, and demand-side management. SMI also uses the same system equipment to send information back through the network to meters to capture additional data, control the meters, or update the meters’ irmware. Figure 17.1 shows an overview of the SMI solution [4]. There is often confusion between SMI and the traditional AMR solution. The major characteristics that distinguish SMI from AMR are summarized in Table 17.1. 17.3 MAJOR SYSTEM COMPONENTS OF SMI A SMI system is comprised of a number of technologies and applications that have been integrated into one solution. The four major SMI components are: • • • • Smart meters Communication system Meter data management systems (MDMS) Home area networks (HAN) In addition, many SMI interfaces are needed in order for other system-side applications to use information collected by SMI, including load forecast, outage response, customer support, and system operations. Some of these system-side applications are discussed in the following sections. 402 Smart Grids FAN Multidwelling units Feeder and transformer meters Collectors DMS SAP Other enterprise applications Enterprise service bus OMS Smart grid devices WAN ADCS Residential and commercial meters MDMS Data center Portal SMNO IHD HAN Internet PHEV FIGURE 17.1 Overview of SMI solution. TABLE 17.1 Differences Between SMI and AMR AMR Collection of registered readings from a distant location One-way communication Monthly collection of data Coverage limited to a small area or a portion of a system No additional function 17.3.1 SMI An infrastructure to collect, store, analyze, and utilize remotely interval meter data Two-way communication Variable data granularity supporting TOU pricing Whole-system coverage Additional functionalities, such as remote irmware upgrades, HAN supports smart meters Conventional electromechanical meters have been used by utilities for residential customer billing for many years. These meters simply record the total energy consumed over time with an incremental energy counter. Although digital meters have been used for billing in the last two decades, they are also mainly used to record accumulated energy (there may be some limited records of megawatt [MW] demand). The measurements from both electromechanical meters and nonsmart digital meters are collected manually by physical site visits and, thus, record only the readings at the 403 Smart Metering and Infrastructure time of the visit. Smart meters are intelligent, solid-state, programmable devices that can perform many functions beyond energy consumption recordings. By using builtin memories, smart meters can record and store readings at preset intervals (e.g., 15 min, 30 min, or hourly) and prescheduled times. With built-in communication modules, they can connect to a two-way communication system, not only to send readings from meters to the data center but also to deliver information or control orders from the data center to meters. The two-way communication functionality supports on-demand reading, which enables veriication of customer energy consumption in time, remote connection and disconnection, and detection of tampering or out-of-range voltage conditions. These functions also make TOU rate or real-time pricing and demand management programs possible. Most of the smart meters available on the market can also send out a “last gasp” message when loss of power is detected and a “irst breath” message when power service is restored. Such information provides signiicant beneits for outage location and response. An example of a smart meter [4] is shown in Figure 17.2. The measurements provided by smart meters may be different depending on utilities’ requirements and practices. The main measurement attributes provided by a residential smart meter installed by BC Hydro, a Canadian utility, are listed in Table 17.2 [5]. Disconnect/ reconnect switch LAN antenna Hard connect/disconnect LCD display Infrared port for communication LAN communication card HAN communication card [optional] Tamper Operating system Supercapacitor Solid-state “last gasp” memory detection instruction set (remotely upgradable) RFID (radio- Metrology frequency ID) components FIGURE 17.2 A demonstrative smart meter. 404 Smart Grids TABLE 17.2 Main Measurements Provided by a Residential Smart Meter Present Register Instantaneous voltage phase A Self-Read Register Continuous accumulated demand (watts) delivered Instantaneous watts delivered Accumulated demand (watts) aggregate delivered Instantaneous watts delivered phase A Maximum demand (watts) delivered Instantaneous watts delivered phase B Varh delivered Instantaneous watts delivered phase C Varh quadrant 1 Instantaneous watts received phase A Varh quadrant 4 Number of demand resets Varh received Number of inversion tampers Wh delivered Number of minutes on battery Wh received carryover Number of power outages Number of removal tampers Number of times programmed Interval Data Maximum voltage phase A Maximum voltage phase B Maximum voltage phase C Minimum voltage phase A Minimum voltage phase B Minimum voltage phase C Varh quadrant 1 Vh phase A Vh phase B Vh phase C Wh delivered A smart meter can also work as a gateway for utilities to communicate with the customer’s HAN, which enables customers to view their near real-time consumption information and receive price signals from utilities. Depending on customers’ willingness, smart meters can relay load control commands from the utility to customers’ appliances for emergency load control or demand response programs. Typical smart meter functionalities include the following: • Record interval (daily, hourly, or subhourly) energy consumption and demand data • Support demand read capability • Provide bidirectional metering, which will accommodate distributed generations at customer sites • Provide notiication on loss of power and service restoration • Provide tamper alarms and enable theft detection • Provide voltage measurement, voltage alarms, and power quality monitoring • Be remotely programmed and irmware upgraded over the air • Support remote time synchronizing • Enable TOU rate billing • Enable remote connection and disconnection service • Limit loads for purposes of demand response • Communicate and interact with intelligent appliances or devices in a customer’s HAN • Protect meter data security Smart Metering and Infrastructure 405 Since smart meters enable demand response programs, this leads to a reduction in emissions and carbon dioxide pollution. The detailed consumption interval data accessible by customers will encourage consumers to reduce their unnecessary electricity usage and shift their consumption to nonpeak time. 17.3.2 communIcatIon system The backbone of the SMI system is the communication network. Employing communication networks, the SMI system is capable of reading a smart meter many times each day, and can deliver meter data, including outage and tamper alarms, in near real time from the meter to the utility data center. Utilities have different geographical conditions, load densities, and legacy communication facilities, and, therefore, the communication system to be implemented for SMI is different in each case. The communication network generally uses one of several different network architectures, which are comprised of meter interfaces, data collectors, or concentrator towers, and a central data collection device located in the data center. A SMI communication system usually has hierarchical architectures and hybrid communication systems with various technologies. Data collectors or concentrators are used and installed on poles, in substations, or on other facilities to form the link between meters and network management application software. The National Institute of Standards and Technology (NIST) in the United States has deined the communication network in two layers [6]: the SMI wide area network (WAN) communication system serves as a backhaul communication from meter data collectors in the ield to the head end (operation center), whereas the ield area network communication system covers the last-mile communication from smart meters to meter data collectors. It is often seen in existing designs that one or two major lastmile communication technologies are used for the communication architecture and a variety of technologies are used for the data backhaul system. We summarize the major communication technologies in SMI implementation as follows. 17.3.2.1 Meshed Network Radio Frequency Wireless mesh networks were originally developed for military applications. In the mesh network, each communication node not only processes its own data, but it also serves as a relay for other nodes. In other words, it propagates the data in a network. This architecture relies on a local area network (LAN) utilizing the unlicensed industrial, scientiic, and medical (ISM) band (902–928 MHz) with either a proprietary protocol or an open standard such as IPv6. Although the transmission power is limited to 1 W, a 1–5 mile range from meter to meter data collector is achieved by using other meters as repeaters. The number of meter-to-meter hops varies from system to system and determines the latency in the network. Although the hops can theoretically reach a large number (such as 16–20), the number has been minimized in practical applications to achieve an acceptable network performance. The meter number in one mesh network varies from several hundred to several thousand. This meshed network radio frequency (MNRF) technology is associated with a limited service range, resulting in an increased number of regional collectors. However, the lower cost for a regional collector offsets the negative effect due to 406 Smart Grids the number of regional collectors in the system. These systems generally rely on a cellular interface at the regional collector, although telephone or other proprietary communication links can be an option. Mesh networks are effective systems for the deployment of reliable large-scale SMI networks and are widely implemented in North America due to their scalability and reliability. However, the low-power radio signals used in SMI systems are deployed in the unlicensed spectrum and are susceptible to interference and link blockage that can affect the performance. Also, mesh networks have to be properly dimensioned to operate within the constraints of the mesh routing protocol and system capacity limits. Proper network planning and analysis techniques are valuable tools that can mitigate these issues in advance. 17.3.2.2 Point-to-Point Radio Frequency Point-to-point radio frequency (PPRF) wireless networks are also called star networks. In this type of network, end devices are directly connected to the collectors. A solution called “buddy mode” enables any end device to relay another end device, and further connect to the collector, eventually leading to network coverage. The PPRF is designed to have all end devices connected to a collector through no more than two hops. Most PPRF systems use licensed frequencies with higher power than unlicensed bands. Through the higher power capacity and enhanced design in receiver sensitivity, the current generation of PPRF systems can cover a service range of 5–15 miles from a meter to a regional collector. Compared with the MNRF network, the PPRF star network has a much bigger coverage with one regional collector in terms of end device numbers and geographical areas. This leads to a signiicant beneit in minimizing cost, as the ratio of meters to regional collectors can be increased. Since end devices can be connected to the collector directly or within two hops, the PPRF star network often introduces much lower latency in communication and, thus, has a better network performance. However, PPRF networks are usually proprietary and are limited in interoperability. The availability of added functionality is affected by the vendor’s development capabilities and schedule. 17.3.2.3 Power Line Carrier/Broadband Power Line Power line communication technologies use existing power lines as the communication vehicle. Power line carrier (PLC) provides a narrowband capacity, whereas broadband power line (BPL) offers a higher-capacity communication channel. In either case, additional interface equipment is needed at substations if a relatively long communication route is required. With the PLC architecture, the communication from substation to meter is achievable. With BPL, however, additional equipment to boost the signal is required along a medium-voltage line at about 1-mile intervals. In most cases, bridges are also required around distribution transformers. PLCs are attractive to utilities, the owners of the assets, and make it easy for them to manage, operate, and maintain both lines and communication devices on the lines. As a matter of fact, PLC is the dominant smart metering communication technology in Europe. In this case, data collectors, which are installed at the secondary side of distribution transformers, talk to end customer meters through PLC, and Smart Metering and Infrastructure 407 are usually connected to the data center through other public cellular or WAN technologies. There are three advantages to the PLC communication in Europe. First, PLC provides reasonable communication performances in terms of data throughput and reliability in a short distance range from a distribution transformer to the end customers. Second, data collectors are installed at the same voltage level as that of the distribution system with end customers and, thus, no additional bridge is needed for PLC signals to cross the distribution transformer. Third, the number of customers connected to one distribution transformer can be relatively high (often over 40 customers), which reduces the number of data collectors required and makes the connection between the WAN and data collectors cost effective. However, the situation is different in North America. The number of customers served by one distribution transformer is usually small (fewer than 20 usually), and, thus, it is not cost effective to have data collectors installed at the secondary side of the distribution transformer. On the other hand, if bridges are used to allow PLC to bypass distribution transformers and enable the data to be transmitted through the medium-voltage lines, the cost is still high and the communication performance is low, as both the data transmitting rate and additional functionalities are limited. Although BPL offers a broader range of additional functionalities, the current cost per home serviced is also relatively high. In short, although the PLC and BPL technologies are attractive to utilities for obvious reasons, unfortunately, the prohibitive capital requirements and increasing data volumes have limited their widespread adoption in North America. Many utilities have chosen wireless technologies to accomplish their goals of developing reliable two-way communication for SMI networks. 17.3.2.4 Wireless Broadband Wireless broadband utilizes the emerging Institute of Electrical and Electronics Engineers (IEEE) 802.16 standard (Wi-MAX) for direct communication to meters equipped with an 802.16 chip set. This architecture has the potential to provide a direct communication link to meters and other devices within a 50-mile radius with a proprietary communication network. A broadband wireless architecture provides the beneit of wide geographic coverage, with complete control over the communication infrastructure and no ongoing airtime charges while maintaining open industry standards. Any device equipped with an 802.16-compliant chip set can be addressed through the network. A secondary beneit would be the potential to leverage the communication network for remote workforce management, to implement additional distribution automation, and to provide broadband connectivity to communities not currently connected. Although this approach provides an effective control over potential future increases in airtime cost, utilities will bear the capital and ongoing operation, maintenance, and administration cost of the network. 17.3.2.5 Cellular Technologies Cellular technologies are among the most popular WAN communication technologies in the SMI world, mainly due to their immediate