Demand Dispatch with Heterogeneous Intelligent Loads

Transcription

1 Demand Dispatch with Heterogeneous Intelligent Loads HICSS 5, January 6, 217 Sean P. Meyn Florida Institute for Sustainable Energy Joel Mathias, UF & Ana Bušić, Inria Department of Electrical and Computer Engineering University of Florida Special thanks to Dr. Yue Chen and to our sponsors Google, NSF, DOE

2 Demand Dispatch with Heterogeneous Intelligent Loads Outline 1 Celebrating Three Years of HICSS 2 Homogeneity by Design & One Way Communication 3 Conclusions 4 References

3 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: 1 / 19

4 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? (Ancillary Service) 1 / 19

5 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? (Ancillary Service) 35 Fig. 1. Coal-fired generators do not follow regulation signals precisely... Some do better than others Gen 'A' Actual Regulation Gen 'A' Requested Regulation 15 1 Regulation (MW) Regulation (MW) 5 : Gen 'B' Actual Regulation Gen 'B' Requested Regulation : 7: 8: 9: 6: 7: 8: 9: Regulation service from generators is not perfect Frequency Regulation Basics and Trends Brendan J. Kirby, December 24 1 / 19

6 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? Reliable? Will AS be available each day? It may vary with time, but capacity must be predictable. 1 / 19

7 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? Reliable? Cost effective? 1 / 19

8 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? Reliable? Cost effective? Is the incentive to the consumer reliable? 1 / 19

9 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? Reliable? Cost effective? Is the incentive to the consumer reliable? Customer QoS constraints satisfied? Fresh fish, comfy house, clean pool, happy farmers and data centers... 1 / 19

10 Celebrating Three Years of HICSS We want: Responsive Regulation Demand Dispatch the Answer? A partial list of the needs of the grid operator, and the consumer: High quality AS? Reliable? Cost effective? Is the incentive to the consumer reliable? Customer QoS constraints satisfied? Demand dispatch can do all of this (by design) 1 / 19

11 Celebrating Three Years of HICSS Control Architecture Intelligence at the Load distinguishes our work from Mathieu, Hiskens, Callaway, Kizilkale... Step 1: Load-level Feedback Loops Yti (kw) Xti (state) on/off Pacify Aggregate Dynamics ζt (BA/aggregator) 2 / 19

12 Celebrating Three Years of HICSS Control Architecture Intelligence at the Load distinguishes our work from Mathieu, Hiskens, Callaway, Kizilkale... Step 1: Load-level Feedback Loops Yti (kw) Xti (state) on/off Lit (QoS) Pacify Aggregate Dynamics ζt (BA/aggregator) 2 / 19

13 Celebrating Three Years of HICSS Control Architecture Intelligence at the Load distinguishes our work from Mathieu, Hiskens, Callaway, Kizilkale... Step 2: Condition Grid Reference Signal MW A Reg signal (one week) 6 MW8 4 2 = / 19

14 Celebrating Three Years of HICSS Control Architecture Intelligence at the Load distinguishes our work from Mathieu, Hiskens, Callaway, Kizilkale... Step 2: Condition Grid Reference Signal MW A Reg signal (one week) 6 MW8 4 2 = = HVAC + Pool Pumps 3 / 19

15 Celebrating Three Years of HICSS Control Architecture Intelligence at the Load distinguishes our work from Mathieu, Hiskens, Callaway, Kizilkale... Step 3: Actuator Feedback Loop Easily controllable by design r t + G c y t Compensator Bunch of loads MW y t r t Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 If I had one million pools, my problems would be solved! -TB, 215 t 4 / 19

16 Celebrating Three Years of HICSS Control Architecture Two Questions r t + G c y t Compensator Bunch of loads 6 4 y t r t MW Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 If I had one million pools, my problems would be solved! -TB, 215 t Do we need such accurate tracking? If not, does one-way communication suffice? Smart Fridge / Dumb Grid? 5 / 19

17 Celebrating Three Years of HICSS Goals of This Work Goals 1 Address highly heterogeneous loads (recall KP s lecture) 6 / 19

18 Celebrating Three Years of HICSS Goals of This Work Goals 1 Address highly heterogeneous loads (recall KP s lecture) 2 Investigate one-way communication: Traditional Resources Disturbances ω desired + Σ G c Compensation Air Conditioners Res. Water Heaters Comm. Water Heaters Pool Pumps G p GRID ω t Actuation 6 / 19

19 Traditional Resources Disturbances ω desired + Σ G c Compensation Air Conditioners Res. Water Heaters Comm. Water Heaters Pool Pumps G p GRID ω t Actuation Homogeneity by Design & One Way Communication Building a 1 GW Battery

20 Homogeneity by Design & One Way Communication Control Architecture Dumb Grid? Only at Low Frequencies Balancing Authority wants High Gain Here 2 Low Gain Here 15 Phase (deg) Magnitude (db) Water Pumping Pool Pumps Chiller Tanks Residential Water Heaters Refrigerators Fans in Commercial Buildings Uncertainty Here Grid Transfer Function Frequency (rad/s) Frequency (rad/s) 7 / 19

21 Homogeneity by Design & One Way Communication Control Architecture Dumb Grid? Only at Low Frequencies Balancing Authority wants High Gain Here 2 Low Gain Here 15 Phase (deg) Magnitude (db) Water Pumping Pool Pumps Chiller Tanks Residential Water Heaters Refrigerators Fans in Commercial Buildings Uncertainty Here Grid Transfer Function Frequency (rad/s) Frequency (rad/s) Primary control is of interest, but not a topic of this lecture. 7 / 19

22 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Randomized Control a Convex Relaxation Nominal model: (continuous time) Controlled rate matrix: A ζ = r[ I + S ζ ] 8 / 19

23 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Randomized Control a Convex Relaxation Nominal model: (continuous time) Controlled rate matrix: A ζ = r[ I + S ζ ] temp(centigrade) on/off state Sampled TCL temp TCL temp Sampled TCL On/Off state TCL On/Off state 6 time(min) / 19

24 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Randomized Control a Convex Relaxation Nominal model: (continuous time) Controlled rate matrix: A ζ = r[ I + S ζ ] temp(centigrade) on/off state Sampled TCL temp TCL temp Sampled TCL On/Off state TCL On/Off state 6 time(min) Myopic design: S ζ (x, x ) = S (x, x ) exp(ζu(x ) Λ ζ (x)) 8 / 19

25 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Mean field model Given a homogeneous collection of N loads {X i }. Empirical distribution at time t : µ N t (x) := 1 N N I{Xt i = x}, x X. i=1 Approximated by the mean-field equations, d dt µ t = µ t A ζt Linearized Dynamics: In all examples, Passive! 9 / 19

26 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Linearized Dynamics Flatten Response Magnitude (db) -2-4 Bode Diagrams: G (jf) M (jf)g (jf) 6 Phase (deg) f Each load knows its dynamics: Inverse filter based on linear model Matlab Robust Control Toolbox 1 / 19

27 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Linearized Dynamics Flatten Response Open-loop tracking with 4, heterogeneous TCLs: Open Loop Tracking (MW) Refrigerators 6 15 Fast Electric Water Heaters Slow Electric Water Heaters A balancing reserves (filtered/scaled) Power state 6 4 Mean-field Model Stochastic Output 1 6 hrs 24 hrs Nominal 24 hrs Demand Dispatch 2 loads of each class, each with its own local filter. 1 / 19

28 Homogeneity by Design & One Way Communication Local Control Design Local Control Design Linearized Dynamics Flatten Response Open-loop tracking with 4, heterogeneous TCLs: Open Loop Tracking (MW) Refrigerators 6 15 Fast Electric Water Heaters Slow Electric Water Heaters A balancing reserves (filtered/scaled) Power state 6 4 Mean-field Model Stochastic Output 1 6 hrs 24 hrs Nominal 24 hrs Demand Dispatch 2 loads of each class, each with its own local filter. Violation of QoS constraints is impossible 1 / 19

29 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Putting Together the Components Traditional Resources Disturbances ω desired + Σ G c Compensation Air Conditioners Res. Water Heaters Comm. Water Heaters Pool Pumps G p GRID ω t Actuation Local Control Design for Each Load Class 11 / 19

30 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Putting Together the Components Simulation Elements ω desired + Σ G c Compensation 1 Actuation: Five load classes, and generation Traditional Resources Air Conditioners Res. Water Heaters Comm. Water Heaters Pool Pumps Actuation Disturbances G p GRID ω t 2 Disturbance: A Balancing Reserves 3 Grid: ERCOT model from Chavez, Baldick, Sharma 212 Example Grid Transfer Function Magnitude (db) Phase (deg) Frequency (rad/s) 11 / 19

31 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Putting Together the Components Simulation Elements ω desired + Σ G c Compensation 1 Actuation: Five load classes, and generation Traditional Resources Air Conditioners Res. Water Heaters Comm. Water Heaters Pool Pumps Actuation Disturbances G p GRID ω t 2 Disturbance: A Balancing Reserves 3 Grid: ERCOT model from Chavez, Baldick, Sharma 212 Example 1 4 BA Control: Integral control 2 15 High Gain Here Low Gain Here Grid Transfer Function Magnitude (db) I Control Phase (deg) Frequency (rad/s) 11 / 19

32 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Closed-Loop Response Grid frequency maintained at 6Hz ±.5 12 / 19

33 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Closed-Loop Response Grid frequency maintained at 6Hz ± A Control Input U Actuation U D LP Control Input Pool Actuation Generators 15 1 Power (MW) 5-5 Power (MW) Day 1 Day 2 Day3 Day 4 Day 5 Day 1 Day 2 Day3 Day 4 Day 5 Close-Up on Pools Acting as Virtual Generators / Virtual Batteries 12 / 19

34 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Closed-Loop Response Grid frequency maintained at 6Hz ±.5 Provided dead-band is absent ω(t) Grid Frequency (Hz) millihertz governor dead-band Governor dead-band absent x seconds Thank you, Mani for the inspiration, and thank you Joel for the last minute simulation! 13 / 19

35 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Periodic Capacity: HVAC Gain g(t) = 1 δ sin(f d t) ACs: capacity has 24 hour period; time-varying gain function g(t). Power (MW) Closed loop response: Actuation U A D Power (MW) AC Output: Input to ACs AC Actuation δ = 1 2 Day 1 Day 2 Day3 Day 4 Day 5 Frequency regulation remains perfect with δ =.5 14 / 19

36 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Periodic Capacity: HVAC Gain g(t) = 1 δ sin(f d t) ACs: capacity has 24 hour period; time-varying gain function g(t). Rads/Sec Grid Frequency Power (MW) AC Output: Input to ACs AC Actuation δ = 1 Day 1 Day 2 Day3 Day 4 Day 5 Power (MW) Not bad with δ = 1! 14 / 19

37 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Taming the Duck 25 2 Net Load Curve Low pass 15 GW 1 5 Mid pass -5 12am High pass 3am 6am 9am 12pm 3pm 6pm 9pm 12am Hypothetical CAISO Net-load over one day in 22, and its frequency decomposition. 15 / 19

38 Homogeneity by Design & One Way Communication Putting Together the Components Macro Control Design Taming the Duck Low Freq. Generation L r t Residual load Actuator output GW am Load Response Pool s-wh f-wh AC 3am 6am 9am 12pm 3pm 6pm 9pm 12am Residual Load = Net Load Low Pass : tracked nearly perfectly 15 / 19

39 Gas Turbine Coal Batteries Water Pump Control Σ C Power Grid LOAD CA Actuator feedback loop Conclusions H Voltage Frequency Phase

40 Conclusions Conclusions The virtual storage capacity from demand dispatch is enormous With appropriate filtering and local control, DD can provide excellent ancillary service, even without two-way communication. 4 G r = G 1 + G 2 + G 3 3 G r (t) GW 2 1 G 1 G 2 G 3 = Real Time Market DD can replace the RTM Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 16 / 19

41 Conclusions Conclusions Questions Engineering Better science for QoS constraints Fraction of loads without opt-out.5 with opt-out Cleaning QoS Cycling QoS 17 / 19

42 Conclusions Conclusions Questions Engineering Better science for QoS constraints Better understanding of time-varying capacity Network issues Conflict or harmony between transmission and distribution? 17 / 19

43 Conclusions Conclusions Questions Engineering Better science for QoS constraints Better understanding of time-varying capacity Network issues Conflict or harmony between transmission and distribution? Policy and Economics Virtual energy storage is surely cheaper than batteries. Challenge: economic theory for a zero marginal cost market. Solutions: Contracts for services: FP&L s On Call program Mileage model: FERC Order 755 Will new FERC orders streamline innovation in 217? 17 / 19

44 References Pre-publication version for on-line viewing. Monograph available for purchase at your favorite retailer More information available at August 28 Pre-publication version for on-line viewing. Monograph to appear Februrary 29 Control Techniques Markov Chains and Stochastic Stability FOR Complex Networks P n (x, ) π f π(f ) < C sup Ex [SτC (f )] < V (x) f (x) + bic (x) Sean Meyn S. P. Meyn and R. L. Tweedie References 18 / 19

45 References Selected References More at J. Mathias, A. Bušić, and S. Meyn. Demand dispatch with heterogeneous intelligent loads. In Proc. 5th Annual Hawaii International Conference on System Sciences, Jan 217. A. Bušić and S. Meyn. Distributed randomized control for demand dispatch. In IEEE Conference on Decision and Control, pages , Dec 216. S. Meyn, P. Barooah, A. Bušić, Y. Chen, and J. Ehren. Ancillary service to the grid using intelligent deferrable loads. IEEE Trans. Automat. Control, 6(11): , Nov 215. Y. Chen, A. Bušić, and S. Meyn. Individual risk in mean field control with application to demand dispatch. IEEE Trans. on Smart Grid, 215 (under revision prelim. version IEEE CDC) Y. Chen, A. Bušić, and S. Meyn. State estimation for the individual and the population in mean field control with application to demand dispatch. CoRR and to appear, IEEE Transactions on Auto. Control, 217. (prelim. version IEEE CDC) 19 / 19