Outline. Real- Time Distributed Conges5on Control for Electrical Vehicle Charging

Transcription

1 Real- Time Distributed Conges5on Control for Electrical Vehicle Charging Catherine Rosenberg (University of Waterloo) Joint work with Omid Ardakanian and Srinivasan Keshav Outline ü The electric grid as it is ü The smart grid ü ISS4E: Internet research and the smart grid ü Real- Time Distributed Conges5on Control for Electrical Vehicle Charging 2 1

2 The Electric Grid Major components of the electric grid: Genera5on Electricity produc5on Natural gas, coal, hydro, nuclear, renewables, etc Transmission Network of high voltage power lines Analogous to fiber links in the internet core Distribu5on Lower voltage power lines carrying electricity from transmission lines to consumers Analogous to access networks in the Internet 3 Today s Transmission Grid Transmission Capacity... ü Designed to meet annual 15 mn peak with redundant capacity.. ü Planning/Implementa5on requires several years many projects are commi\ed for construc5on well before they are needed... The transmission system, for the most part, is sophis1cated, reliable, reasonably secure... BUT...It operates at peak capacity for short 1mes each year. 2

3 The Distribu;on Network Essen5ally a radial network the loss of a major feeder line results in customer outages... Technology is mixed (Some is REALLY OLD!!) Some equipment installed more than 75 years ago remains in opera5on. System is generally designed for one way flow to the users... Monitoring of customer service has been limited... The u5lity has limited means of iden5fying local overloads or thed. The System Operator has Real Challenges Electricity is consumed the instant that it is created there is prac5cally no storage on the network for electricity... The u5lity has the task of ensuring that the genera5on meets the demand on a second by second basis.. Large generators in par5cular the newer ones, do not change load easily or quickly... Yet... ü When someone turns on a stove, or even a light... A generator somewhere is adjusted to meet the demand... Adding capacity at peak ;mes is expensive... At off- peak 5mes it is very cheap. 3

4 Issues with the Grid Grid is over- provisioned (sized for the peak), no storage - always need to match demand with supply. Philip Giudice, Commissioner, Massachusetts Department of Energy, Nov. 30, % of the generating capacity in Massachusetts is needed fewer than 88 hours per year 7 Issues with the Grid Reliability Outdated switches, lack of sensors results in poor visibility of the grid 8 4

5 In Summary Power systems operators can control... ü Genera5on ü Transmission ü Distribu5on Loads, for the most part are uncontrolled... ü Demand Response Some control of industrial loads to reduce peak loads ü Exis5ng metering systems do not provide customers with informa5on needed to monitor and avoid peak periods U;li;es control the supply of energy but have very limited control over the demand... The system is sized for the peak Facts If the grid were just 5% more efficient equivalent to permanently elimina5ng the fuel and greenhouse gas emissions from 53 million cars. If every American household replaced just one incandescent bulb with CFL would conserve enough energy to light 3 million homes è Terrific opportuni5es for improvement

6 Outline The electric grid as it is The smart grid ISS4E: Internet research and the smart grid Real- Time Distributed Conges5on Control for Electrical Vehicle Charging 11 The Smart Grid Bi- direc5onal energy flows Renewables and DG - millions - non- tradi5onal - intermi\ent Smart Grid Consumer in the loop (incen5viza5on) New (elas5c) loads: EVs + smart appliances Storage Reliable & fast communica5on + sensors 12 6

7 The Smart Grid Is About Objec5ves: Improve efficiency Reduce total GHG Emissions Increase u5liza5on (defer capital expenses) Maintain or improve reliability and security Upgrading the Distribu5on Network Customers - near real 5me data U5li5es Be\er monitoring of loads and devices Be\er distribu5on protec5on allow remote genera5on (Distributed Genera5on & Micro grids) Reduc5on of thed Controlling demand 13 Rolling out the smart grid will require massive change A rela5vely sta5c, predictable, stable system with inelas5c loads and a few points of control A highly dynamic system with elas5c loads and millions of points of control A paradigm shift 14 7

8 Outline The electric grid as it is The smart grid ISS4E: Internet research and the smart grid Real- Time Distributed Conges5on Control for Electrical Vehicle Charging 15 Our Research Use Internet concepts to smarten the grid 16 8

9 Can Internet Research Help? The Internet resembles the smart grid Cri5cal infrastructure Large- scale Heterogeneous Hierarchical Matches geographically distributed demands with distributed genera5on Distributed highly variable sources Balances centraliza5on and decentraliza5on Simple API 17 Differences Primarily one- way vs. primarily two- way flows Grid has prac5cally no storage Consumers are used to see their electrical bill reflect what they really use Policy makers are very proac5ve Many loads at home are determinis5c, most loads are predictable Packets are addressed, electrons are not 18 9

10 ISS4E Vision To apply our exper,se in Informa1on Systems and Sciences to find innova1ve solu1ons to problems in energy systems. We work within Waterloo Ins1tute for Sustainable Energy (WISE) in collabora,on with ü researchers in related disciplines ü partners in industry 19 ISS4E and WISE! l The Waterloo Ins5tute for Sustainable Energy (WISE) was established at the University of Waterloo in Focal point at UW for research in energy studies l More than 70 faculty members with graduate students and postdoctoral fellows working as mul5- disciplinary research teams across Engineering, Science and Environment. l Research areas: Renewable Energy Storage & Transport Conversion Technologies Emission Management Power System Op5miza5on Sustainable Energy Policy Conserva5on, Demand Mgmt, Energy Efficiency Green Auto Powertrain ISS4E 20 10

11 Lab facili;es! l Sensors for building monitoring" l Smart power strips for home monitoring and control" l ENVI systems for home energy data collection" l Custom-built wireless sensors for solar panel monitoring" " ISS4E is committed to system building and data collection and analysis" Ongoing Projects Architecture DR EV integration Pricing Data analysis Application/tool design Storage 11

12 Outline The electric grid as it is The smart grid ISS4E: Internet research and the smart grid Real- Time Distributed Conges;on Control for Electrical Vehicle Charging 23 Challenges posed by EV integra5on ü The large- scale introduc5on of electric vehicles (EVs) will greatly affect the electrical grid's distribu5on system. ü Each EV can impose a significant load on the distribu5on network: especially with L2 charging. ü Using lower- level (i.e., L1) charging does reduce the impact on the grid but greatly increases the dura5on of the charging process. ü There is an inherent trade- off between charging level, charging dura5on, and impact on the grid. ü EV mobility has the addi5onal impact that EV load may appear at different parts of the distribu5on network at different 5mes. Need smart charging schemes to protect the grid while allowing fast charging when possible 24 12

13 How to control EV charging? ü Exis5ng approaches to control EV charging: use a central controller to compute a charging schedule (using power flow analysis) that does not congest any part of the distribu5on network. It requires an accurate model of the distribu5on network (typically not available or not up- to- date). cast the control algorithm in the form of a distributed op5miza5on. ü Both approaches need to predict the future demand from non- EV loads, the number of charging EVs, and their ini5al SoC so as to compute a schedule ahead of ;me. ü The safety margin built in to hedge against predic5on errors makes both approaches overly conserva5ve. 25 Our Proposal ü The future smart grid is likely to have a large number of measurement and control devices that are interconnected by a ubiquitous communica5on network. ü We propose to use fast- ;mescale measurements and communica5on to control EV charging in real ;me, mo5vated by techniques for conges5on control in the Internet

14 More specifically ü The propaga5on delay between any EV charger and its connected substa5on is less than 1ms. ü Hence, it is feasible to design and implement a control algorithm that changes the EV charging rate in response to the conges5on state of the distribu5on system (a func5on of the uncontrollable loads) every few milliseconds (same order of magnitude as one cycle of AC power (16.6ms)). ü With our proposed approach, if an EV is charging at a rate that affects the reliability of the grid (overhea5ng a transformer or overloading a feeder) its rate can be decreased in a few cycles, aver5ng damage and the invoca5on of grid self- protec5on. ü This fundamental insight changes the approach to EV charging from a slow centralized or decentralized op5miza5on approach to a fast dynamic approach. 27 The 5me scales 28 14

15 Contribu5ons ü We show that the conges5on control problem in the context of a distribu5on system is similar in many aspects to the conges5on control problem in the Internet. ü We propose a measurement and signaling architecture to provide real- 5me explicit feedback to EV chargers. ü We present three real- 5me distributed conges5on control mechanisms for charging EVs. Our focus is on establishing a vision and proposing a high- level architecture, rather than valida5on and analysis, which we defer to further study. 29 The System Today Tomorrow 30 15

16 Conges5on Control: Internet vs. Grid ü Defini;on of conges;on: a user sees conges5on if In CN: buffer overflow on a path In DN: the current passing through at least one feeder on the path persistently exceeds its current limit or the winding hot spot temperature of at least one transformer exceeds a threshold ü Topology: The Internet is a general mesh network consis5ng of many sources and des5na5ons that are connected by communica5on links and routers. If a source congests a link by sending a burst of data, all other sources that send data through this link see its impact on their QoS even if their packets are going to different des5na5ons. A distribu5on system has a tree topology in which every node has a parent which supplies its demand and the root of the tree supplies the demand of all loads in this tree. If a few loads congest a feeder by consuming high power, only downstream loads that are supplied by this feeder are affected. Other loads located in this tree will not be affected. 31 Conges5on Control: Internet vs. Grid ü Infrastructure for sending measurement and control signals: In the Internet, data packets carry control informa5on and therefore the same infrastructure is used for transmivng both data and control signals Power lines deliver electricity to customers and conges5on signalling is done separately (using an auxiliary comm. network). ü Conges;on no;fica;on: There are two types of conges5on feedback in the Internet: explicit and implicit. Intermediate routers can explicitly report conges5on to end- nodes. End- nodes can also infer conges5on by measuring packet loss or es5ma5ng the round- trip delay; this is known as implicit conges5on no5fica5on. In the grid it is difficult to infer conges5on implicitly

17 Conges5on Control: Internet vs. Grid ü Self- protec;on: Both systems protect themselves against conges5on. Internet routers are configured to drop packets to avoid conges5on. In a distribu5on network, the protec5on system consists of re- lays and circuit breakers that trip and disconnect the load in case of conges5on. The protec5on mechanisms differ in that the packet dropping schemes do not interrupt service to clients (though it may impact the QoS); however, when a protec5ve relay trips all downstream loads are disconnected, leading to a service disrup5on. 33 Conges5on Control: Internet vs. Grid ü Uncontrolled loads: Both systems are designed to deal with uncontrollable demands. Specifically, UDP traffic is uncontrolled in the Internet and conges5on control mechanisms do not deal with this type of traffic (although UDP packets can easily be filtered and discarded if necessary). Similarly, there are uncontrollable loads in the grid which do not respond to control signals. The main difference is that the current infrastructure does not permit the segrega5on of the uncontrolled loads from the controlled ones

18 Our Vision ü Our vision requires a joint measurement and signaling infrastructure to detect the outset of conges5on very quickly and to inform the chargers that are in the congested region (others should not need to decrease their rates). ü When an EV needs charging, it starts charging at a low rate and then increases it slowly up to L2 as long as it does not receive a signal from the grid that indicates (pre)- conges5on. ü This (pre)- conges5on might be due to the chargers themselves or to an increase in the uncontrollable loads. ü Thanks to the efficient communica5on and control infrastructure, the charger can react nearly immediately to conges5on signals, aver5ng the use of grid protec5on ac5ons from circuit breakers. ü Note that no predic5on of the EV SoC or their mobility is required: charging happens for the EVs present in the system at any point in 5me, and their charging rates are controlled every few milliseconds. )35 Design Goals ü Maintaining grid reliability: a crucial goal in the design of a control mechanism for charging EVs is to maintain the same level of reliability and to ensure that no addi5onal power outage is introduced due to their charging. ü High u;liza;on: without overshoo5ng via a margin (λ) used to hedge against the risk of system over- loading due to transient system behaviour. ü Minimize oscilla;ons: Oscilla5ons are usually inefficient and could affect the life5me of the ba\eries. ü Fairness: Alloca5on of charging rates to EV chargers must be done according to a fairness criteria. ü Robustness: a fail- safe mechanism is needed in case of failure of the comm. infrastructure 36 18

19 Measurement ü The distribu5on network is equipped with measurement devices con5nuously measuring the current in the feeder and the winding hot spot temperature of the transformer. ü They compute an average of the current and the temperature every t M ms. ü Hence, we can compute the difference between the current limit of the feeder (with a margin) and the current passing through it and the difference between the maximum winding temperature (with a margin) and the measured temperature. ü If these differences are posi5ve it means that the feeder/ transformer is not over- loaded. Otherwise, it is nearly overloaded (depending on the value of the margins) and the protec5ve relay will, most probably, trip if this condi5on persists or worsens. 37 The Logical Infrastructure 38 19

20 Decision Making (Rate Alloca5on) ü The space of possible rate alloca5on algorithms is large. ü We outline three distributed conges5on control schemes that illustrate different points in the design space: Intelligent Endpoint Approach Local Scheduling Approach Distributed Explicit Rate Control ü Our chosen algorithms differ in the en55es that makes decisions about charging rates of EVs and the degree of communica5on overhead. 39 Intelligent Endpoint Approach ü In this approach, each EV charger independently decides on its charging rate, much like a TCP endpoint. ü Decision making is distributed because every EV charger sets its rate without direct knowledge about the rates of other chargers. ü MCC node ac;ons: Every MCC node con5nuously measures and checks if its corresponding feeder/transformer is congested. Every t M ms, the root MCC node broadcasts a packet that contains a conges5on flag. This packet is routed hop- by- hop by intermediate MCC nodes un5l it reaches the EV chargers. Each congested intermediate MCC node can modify the packet that it receives from its parent by sevng the conges5on flag. ü EV charger ac;ons: Every EV charger examines the conges5on flag upon receiving a packet from its parent and uses an AIMD algorithm to set its rate

21 Local Scheduling Approach ü The EV charger is slaved to its parent MCC, which makes local scheduling decisions on behalf of the EVs a\ached to it. ü Decision making is distributed in this strategy similar to the previous scheme; however, it is done by leaf MCC nodes instead of EV chargers. This permits to discriminate amongst and schedule their downstream EV chargers. 41 Distributed Explicit Rate Control ü In this approach, all MCCs coordinate to select a charging rate for their subtree, in an a\empt to minimize oscilla5ons (loosely draws on XCP). ü EV charger ac;ons: Every EV charger sends a packet toward the root every t C ms to nego5ate its charging rate for this control interval. This packet contains the current charging rate along with the requested next charging rate. When this packet returns to the EV charger, it adjusts its charging rate to the charging rate encapsulated in the packet. ü MCC node ac;ons: When an MCC node receives a rate requested packet of an EV charger, it may reduce the request rate if its corresponding feeder is congested and this rate is higher than the fair share of this EV charger. Then, it forwards all packets that it has received to its parent. ü When a rate request packet arrives at the root MCC node, the root sends it back to the EV charger along the same path

22 Comparison ü Who has control? : In the first scheme, control is distributed among EV chargers. However, the second and the third schemes cede control to the u5lity ü Oscilla;ons: The third scheme tries to minimize oscilla5ons by accurate and con5nuous computa5on of the remaining capacity and doing rate alloca5on on this basis. ü Communica;on overhead: The third scheme has a higher overhead because control packets travel bi- direc5onally rather than unidirec5onally, as they do in the first two schemes. 43 Conclusions ü A paradigm shid! ü Need to transform this vision into prac5cal solu5ons that are efficient and robust! ISS4E.ca 44 22