REAL-TIME PRICE FORECAST WITH BIG DATA

Transcription

1 REAL-TIME PRICE FORECAST WITH BIG DATA A STATE SPACE APPROACH Lang Tong (PI), Robert J. Thomas, Yuting Ji, and Jinsub Kim School of Electrical and Computer Engineering, Cornell University Jie Mei, Georgia Institute of Technology August 7, 2013 CERTS Review

2 DATA QUALITY AND ITS EFFECTS ON MARKET OPERATIONS DENY OF SERVICE ATTACK ON REAL-TIME ELECTRICITY MARKET COVER UP PROTECTION AGAINST TOPOLOGY ATTACK Lang Tong (PI), Robert J. Thomas, and Jinsub Kim Cornell University August 8, 2013 CERTS Review

3 Project overview Objectives Accurate short-term probabilistic forecasting of real-time LMP. Incorporate real-time measurements (e.g. SCADA/PMU). Scalable computation techniques. Summary of results A real-time LMP model with forecasting and measurement uncertainties. A Markov chain abstraction of real-time LMP computation. Monte Carlo sampling techniques. Preliminary simulations.

4 Outline Motivation Load and real-time LMP as random processes Benchmark techniques Comments on the state of the art. Probabilistic forecasting of real-time LMP Simulation studies Summary and future work

5 Sample paths of load and LMP

6 Day ahead vs. real-time

7 Benchmark techniques Time series ARMA, ARIMA, ARMAX, GARCH Machine learning Neural networks, support vector machines (SVM) Hybrid techniques Jump (switching) models

8 Load forecasting Forecasting Method ANN [1] ANN [2] MAPE 1.46% 1.24% SVM [3] 2.09% [1] P. Mandal, T. Senjyu, N. Urasaki, and T. Funabashi, A neural network based several-hour-ahead electric load forecasting using similar days approach, Int. Journal. of Electric Power and Energy System, vol. 28, no. 6, pp , July [2] P. Mandal, T. Senjyu, N. Urasaki, and T. Funabashi, Short-Term Price Forecasting for Competitive Electricity Market, Power Symposium, NAPS th North. [3] S. Fan, C. Mao and L. Chen, Next-day electricity-price forecasting using a hybrid network, IET Generation, Transmission & Distribution, Volume 1, Issue 1, January, 2007.

9 Day ahead LMP forecasting Forecasting Method ANN [1] ANN [2] ANN [4] MAPE 15.96% 12.92% 8.67% SVM [3] 11.31% GARCH [5] 11.55% ARMAX [6] 9.01% [4] P. Mandal, T. Senjyu, N. Urasaki, and T. Funabashi, A neural network based several-hour-ahead electric load forecasting using similar days approach, Int. Journal. of Electric Power and Energy System, vol. 28, no. 6, pp , July [5] R. Garcia, J. Contreras, A GARCH Forecasting Model to Predict Day-Ahead Electricity Prices, Int. Journal of Electric Power and Energy System, vol. 20, no. 2, MAY [6] J. Zhang, C. Yu, G. Hou, Application of Chaotic Particle Swarm Optimization in the Short-term Electricity Price Forecasting, Int. Journal of Electric Power and Energy System, vol. 28, no. 6, pp.

10 Real-time LMP forecasting Actual and forecast six-hours-ahead electricity price for Victorian electricity market (Nov 1 st to 7 th, 2003). Actual and forecast six-hours-ahead electricity price for PJM market (Jan 8 th to 14 th, 2006). [4] F. Sarafraz, H. Ghasemi, H. Monsef, Locational Marginal Price Forecasting by Locally

11 Summary of the state-of-the-art Extensive literature: Mostly black box techniques Primarily providing point forecasting Rarely deal with on LMP and network effects Extremely accurate load forecasting (1-3%) Relatively poor price forecasting (10-20%)

12 Outline Motivation Probabilistic forecasting of real-time LMP A stylized real-time ex post LMP model Information structure LMP states and Markov chain representations Probabilistic forecast Preliminary simulations Summary and future work

13 A stylized real-time ex post LMP model

14 State estimation

15 Ex ante dispatch

16 Ex post eligible generators

17 Ex post LMP

18 Ex post LMP via IDCOPF

19 Information structure and Markov chain

20 Probabilistic forecasting

21 Outline Motivation Probabilistic forecasting of real-time LMP Preliminary simulations IEEE 16 bus with NYISO sample load profiles Varying load errors and correlations Monte Carlo techniques to estimate transition matrices Summary and future work

22 Load and price distributions 1 LMP distribution at time Probability Markov State (LMP) index Scaled average actual load profile (upper/lower bounds for predicted load profile) from NYISO June 1 st actual load profile upper bound lower bound 32 1 LMP distribution at time 42 MW Probability Markov State (LMP) index Probability LMP distribution at time Markov State (LMP) index Probability LMP distribution at time Markov State (LMP) index time index (5 mins interval in 24 hours)

23 LMP trajectories $/MW load bus1 bus2 bus3 bus4 bus5 bus6 bus7 bus8 bus9 bus10 bus11 bus12 bus13 bus14 LMP trajectories time index (5 mins interval in 24 hours)

24 Time inhomogeneous Markov Chain

25 Transition Matrices Starting Markov State (LMP) index Transition matrix at time Ending Markov State (LMP) index Starting Markov State (LMP) index Transition matrix at time Ending Markov State (LMP) index Starting Markov State (LMP) inde Transition matrix at time Ending Markov State (LMP) index Starting Markov State (LMP) index Transition matrix at time Ending Markov State (LMP) index

26 6 hours ahead prediction error Probability of incorrect prediction Absolute Percentage Error (MAPE Prediction error for best and second best prediction Best prediction 2 nd best prediction time index (5 mins interval in 12 hours) Starting Markov State (LMP) index Prediction transition matrix at time Ending Markov State (LMP) index MAPE of best (most likely) prediction time index (5 mins interval in 12 hours) Starting Markov State (LMP) index Prediction transition matrix at time Ending Markov State (LMP) index

27 LMP prediction for different time horizon $/MW Actual Best prediction 2 nd best LMP prediction for 5 mins ahead $/MW LMP prediction for 1 hour ahead Actual Best prediction 2 nd best $/MW Bus index Actual Best prediction 2 nd best LMP prediction for 3 hours ahead $/MW Bus index LMP prediction for 6 hours ahead Actual Best prediction 2 nd best Bus index Bus index

28 LMP distribution for different prediction horizons LMP distribution for 5 mins ahead LMP distribution for 1 hour ahead Probability (Log) 10-2 Probability (Log) Markov State (LMP) index LMP distribution for 3 hours ahead Markov State (LMP) index LMP distribution for 6 hours ahead Probability (Log) 10-2 Probability (Log) Markov State (LMP) index Markov State (LMP) index

29 Planned research Forecasting techniques Incorporating load models More sophisticated Monte Carlo techniques (important sampling, MCMC) Generator behavior models Forecasting with renewable sources Demand response Extensive simulation studies and comparisons

30 DATA QUALITY AND ITS EFFECTS ON MARKET OPERATIONS DENY OF SERVICE ATTACK ON REAL-TIME ELECTRICITY MARKET Lang Tong (PI), Robert J. Thomas, and Jinsub Kim Cornell University Presented at CERTS Review August 7-8, 2013

31 Observability

32 A graph theoretic condition (KCD 80)

33 Undetectable attack (KJTT 11)

34 Only half of the meters are needed!

35 Bad data identification and removal

36 Framing attack via QCQP

37 DoS attack on real-time operations

38 Simulation examples (AC)

39 Conclusion Attacks on cyber physical systems (power systems) are real: State attacks Topology attacks Dispatch attacks A key step toward protection is deploy authentication mechanisms.