Predicting the Electricity Demand Response via Data-driven Inverse Optimization

Transcription

1 Predicting the Electricity Demand Response via Data-driven Inverse Optimization Workshop on Demand Response and Energy Storage Modeling Zagreb, Croatia Juan M. Morales 1 1 Department of Applied Mathematics, University of Málaga, Spain June 19 th, 2018 June 19 th, / 26

2 Outline Motivation Forecasting the price-responsive costumers demand Defining the estimation problem Solving the estimation problem Case studies One-hour ahead prediction: Simulated pool of buildings with HVAC systems One-day head prediction: Real-life experiment Conclusions June 19 th, / 26

3 Marginal utility Forecasting The Estimation Problem Solution Method Solution Method Case Studies Conclusions Future Work References Assumptions A cluster of price-responsive consumers is considered This cluster is expected to consume more at a favorable price We describe the pool of price-responsive consumers as a utility maximizer agent Step-wise marginal utility function Energy consumption June 19 th, / 26

4 Consumers price-response model maximize x b,t, b,t subject to B x b,t (u b,t p t ) b=1 P t B x b,t P t, t (λ t, λ t ) b=1 0 x b,t E b, t (φ b,t, φ b,t ) It is a linear optimization problem (LOP). Unknown variables: Marginal utilities u b,t Power bounds P t, P t We seek values of u b,t, P t, and P t based on observations of x b,t and p t, given E b. We use the estimated utility maximizer problem to predict x t+1. June 19 th, / 26

5 The estimation problem: Optimality condition minimize Ω T t=1 ɛ t subject to P t λ t P t λ t + B E b φ b,t ɛ t = b=1 φ b,t φ b,t + λ t λ t = u b,t p t φ b,t, φ b,t, λ t, λ t, ɛ t 0 } Ω = {ɛ t, P t, P t, u b,t, λ t, λ t, φ b,t, φ b,t Inverse optimization (IOP) is used to determine the parameters of the model to make predictions of the load. B x b,t (u b,t p t ) b=1 June 19 th, / 26

6 Leveraging auxiliary information Model parameters P t, P t and u b,t, might vary over time. We assume a number of time varying regressors Z such that R P t = µ + α r Z r,t (1) P t = µ + Regressors relate to time and weather: Temperature of the air outside Solar irradiance Hour indicator Past price and load r=1 R α r Z r,t (2) r=1 R u b,t = µ u b + αr u Z r,t (3) r=1 June 19 th, / 26

7 Leveraging auxiliary information The bid must make sense for any plausible value of the features, in particular, The minimum consumption limit must be lower than or equal to the maximum consumption limit The minimum consumption limit must be non-negative June 19 th, / 26

8 Leveraging auxiliary information For example, P t = P + α r Z r,t P + α r Z r,t = P t, t T, for all Z r,t r R r R Assume that Z r,t [Z r, Z r ], then P P + which is equivalent to Maximize Z r,t s.t. Z r Z r,t Z r, r R { } (α r α r )Z r,t 0, t T. r R P P + r R(φ r,t Z r φ r,t Z r ) 0 t T φ r,t φ r,t = α r α r r R, t T φ r,t, φ r,t 0 r R, t T. June 19 th, / 26

9 Solving the estimation problem The estimation problem is non-linear and non-convex. We statistically approximate its solution by solving two linear programming problems instead. 1 A feasibility problem (estimation of power bounds). 2 An optimality problem (estimation of marginal utilities). A two-step data driven estimation procedure to achieve optimality and feasibility of x in a statistical sense. June 19 th, / 26

10 Feasibility problem: Estimation of power bounds PP xxx PP PP tt, PP tt, μμ, μμ, αα rr, αα rr June 19 th, / 26

11 Optimality problem: Estimating marginal utilities P x optimal? P u b,t, μƹ u b, α u r June 19 th, / 26

12 Solving the estimation problem In the bound estimation problem, the penalty parameter K is statistically tuned through validation: x, p, Z data set Training set Used for parametter fitting for each posible value of K Validation set Used to tune parameter K Test set Used to assess forecasting performance We choose K so that the out-of-simple prediction error is minimized K as indicator of the price-responsiveness of the load: June 19 th, / 26

13 Case study 1: One-hour ahead prediction We simulate the price response behavior of a pool of 100 buildings equipped with heat pumps (assuming economic MPC is in place). Two classes of buildings are considered, depending on the comfort bands of the indoor temperature. June 19 th, / 26

14 Case study 1: One-hour ahead prediction We conduct a benchmark of the methodology against simple persistence forecasting and autoregressive moving average with exogenous inputs. Simple persistence model: The forecast load at time t is set to be equal to the observed load at t 1. ARMAX: The aggregate load x is a linear combination of the past values of the load, past errors and regressors. x t = µ + ɛ t + P ϕ p x t p + p=1 R γ r Z t r + r=1 Q θ q ɛ t q q=1 Forecasting performance is evaluated according to MAE and 1 NRMSE = 1 T B x x max x min b,t x t T t=1 b=1 T t=1 B b=1 x b,t x t MASE = T x t x t 1 T 1 T t=2 June 19 th, / 26 2

15 Case study 1: One-hour ahead prediction Load (kwh) Price ( /MWh) Load(Flex) Pmax Pmin InvFor ARMAX No Flex Flex MAE NRMSE MASE Persistence ARMAX InvFor Persistence ARMAX InvFor :00 12:00 00:00 12:00 00:00 12:00 Time June 19 th, / 26

16 Case study 2: One-day (24-h) ahead prediction Data of price-responsive households from Olympic Peninsula project from May 2006 to March 2007 Decisions made by the home-automation system based on occupancy modes, comfort settings, and price The price was sent out every 15 minutes to 27 households June 19 th, / 26

17 Case study 2: One-day (24-h) ahead prediction Load, price, temperature and dew point during December Load in kw Price in USD per MWh Dec 04 Dec 05 Dec 06 Dec 07 Dec 08 Dec 09 Dec 10 Dec 11 Dec 12 Dec 13 Dec 14 Dec 15 Dec 16 Dec Load in kw Temperature Outside in C Load in kw Dew Point in C June 19 th, / 26

18 Benchmark models ARX: Auto-Regressive model with exogenous inputs [Dorini et al., 2013, Corradi et al., 2013] x t = ϑ x X t n + ϑ zz t + ɛ t, with ɛ t N(0,σ 2 ) and σ 2 is the variance. Z t : outside temperature, solar irradiance, wind speed, humidity, dew point (up to 36 hours in the past), plus binary indicators for the hour of the day and the day of the week. Simple Inv: Only the marginal utilities are estimated (12 blocks) as in Step 2, the rest of bid parameters to historical maximum/minimum values observed in the last seven days. Inspired from Keshavarz et al. [2011], Chan et al. [2014]. Inv Few: Our inverse optimization scheme only with the outside temperature and hourly indicator variables as features. Inv All: The same as Inv Few, but including all features. June 19 th, / 26

19 Case study 2: One-day (24-h) ahead prediction Prediction capabilities of different benchmarked methods kw :00 10th Dec 12:00 10th Dec 00:00 11th Dec 12:00 11th Dec 00:00 12th Dec 12:00 12th Dec Actual load ARX Simple Inv Inv Few Inv All 00:00 13th Dec MAE RMSE MAPE ARX Simple Inv Inv Few Inv All June 19 th, / 26

20 Case study 2: One-day (24-h) ahead prediction Estimated marginal utility for the pool of price-responsive consumers Estimated mean utility in $/kwh Price in $/kwh Total load in kw Total load in kw Hour of the day Hour of the day Day June 19 th, / 26

21 Case study 2: One-day (24-h) ahead prediction September March MAE RMSE MAPE MAE RMSE MAPE ARX Simple Inv Inv Few Inv All The prediction performance of the proposed machinery is only slightly lower than that of the state-of-the-art prediction tool developed in Hosking et al. [2013] on the same dataset. However, our methodology produces a market bid! June 19 th, / 26

22 Alternative application: Market bidding June 19 th, / 26

23 Alternative application: Market bidding Day-ahead market Price Energy Balancing market June 19 th, / 26

24 Simple bid BID = {u b,t, t, b; E b, b; P t, P t, t} Maximize x b,t, b,t subject to B x b,t (u b,t p t ) b=1 P t B x b,t P t, t b=1 0 x b,t E b, t June 19 th, / 26

25 Complex bid BID = {u b,t, t, b; E b, b; P t, P t, r u t, r d t, t} Total consumption: P t + b B x b,t ( ) Maximize u b,t x b,t p t x b,t x b,t subject to P t + b B P t 1 + b B t T b B x b,t P t 1 b B x b,t 1 P t b B b B x b,t 1 r u t x b,t r d t t T 1 t T 1 0 x b,t E b b B, t T June 19 th, / 26

26 Conclusions A new method to forecast price-responsive electricity consumption one-step/multiple-steps ahead. The method can be exploited to produce market bids for flexible electricity consumers A two-step algorithm to statistically approximate the exact inverse-optimization solution. A validation scheme to minimize the out of sample prediction error. The proposed methodology has been evaluated on: A synthetic data set corresponding to a cluster of price-responsive buildings equipped with a heat pump and MPC. A data set from a real-world experiment involving electricity consumers able to react to the electricity price. The non-linearity between price and aggregate load is well described by our methodology. June 19 th, / 26

27 Future Work Dealing with errors in the measurements and with bounded rationality (suboptimality). Examining more flexible functional forms between model parameters and regressors. Investigating statistically consistent set-valued functions (feasibility set as a function of regressors) Testing the methodology on other data sets. June 19 th, / 26

28 Contacts Any questions? Juan Miguel Morales OASYS Webpage: oasys.uma.es Full papers Short-term forecasting of price-responsive loads using inverse optimization and A data-driven bidding model for a cluster of price-responsive consumers of electricity are available online at IEEExplore June 19 th, / 26

29 T. C. Y. Chan, T. Craig, T. Lee, and M. B. Sharpe. Generalized inverse multiobjective optimization with application to cancer therapy. Operations Research, 62(3): , O. Corradi, H. Ochsenfeld, H. Madsen, and P. Pinson. Controlling electricity consumption by forecasting its response to varying prices. Power Systems, IEEE Transactions on, 28(1): , Feb ISSN doi: /TPWRS G. Dorini, P. Pinson, and H. Madsen. Chance-constrained optimization of demand response to price signals. Smart Grid, IEEE Transactions on, 4(4): , Dec ISSN doi: /TSG J. R. M. Hosking et al. Short-term forecasting of the daily load curve for residential electricity usage in the smart grid. Applied Stochastic Models in Business and Industry, 29(6): , A. Keshavarz, Y. Wang, and S. Boyd. Imputing a convex objective function. In IEEE International Symposium on Intelligent Control (ISIC), pages IEEE, June 19 th, / 26