Predicting Strategic Energy Storage Behaviors

TL;DR

This work addresses predicting strategic energy storage behaviors in electricity markets by modeling storage operators as price-responsive agents and learning their private costs and constraints from data. It introduces a gradient-based inverse-optimization framework that differentiates through the storage decision problem, with a specialized treatment for convex quadratic costs and a sequential-convex-programming extension for generic, non-quadratic costs, including an ICNN-based approach to preserve convexity. The authors provide convergence guarantees for the quadratic case and demonstrate accurate parameter identification and behavior forecasting on synthetic data and a real Queensland dataset, outperforming black-box baselines and standard optimization tools in many settings. The method supports market monitoring and tariff design by enabling accurate forecasting of storage arbitrage actions, offering a practical path toward mitigating market power and improving dispatch and reliability in future power systems.

Abstract

Energy storage are strategic participants in electricity markets to arbitrage price differences. Future power system operators must understand and predict strategic storage arbitrage behaviors for market power monitoring and capacity adequacy planning. This paper proposes a novel data-driven approach that incorporates prior model knowledge for predicting the strategic behaviors of price-taker energy storage systems. We propose a gradient-descent method to find the storage model parameters given the historical price signals and observations. We prove that the identified model parameters will converge to the true user parameters under a class of quadratic objective and linear equality-constrained storage models. We demonstrate the effectiveness of our approach through numerical experiments with synthetic and real-world storage behavior data. The proposed approach significantly improves the accuracy of storage model identification and behavior forecasting compared to previous blackbox data-driven approaches.

Figures (7)

• Figure 1: Diagram of the proposed energy storage agent model identification and forecasting framework. Prior knowledge of the energy storage agent is modeled as an optimization problem, in which the objective is to minimize the energy cost and degradation cost, subject to storage physical constraints. Parameters in the energy storage models are unknown to the system operator. We use a gradient-based method to update and identify the parameters in the energy storage model, to minimize the difference between the predicted storage response and the actual response.
• Figure 2: Comparison of test loss (MSE) versus iterations using our method and baselines, the solid line represents the median test loss of 10 experiments and shadow represents the 80%/20% quantile test loss of 10 experiments. Our method use 20 training samples and baselines use 200 training samples.
• Figure 3: Comparison of energy storage dispatch using our methods and baselines (MLP and RNN). Our methods take 20 samples for training, while MLP and RNN take 200 samples for training. Figure shows a test example for one day. Ours (quadratic) method identifies the true model and can predict the behavior exactly, thus the green line overlaps with the true response (blue).
• Figure 4: (Left) Cost objective during the training process. (Right) Convergence of the sequential convex programming reaches for 20 iterations.
• Figure 5: Comparison of energy storage dispatch using our methods and baselines (MLP and RNN). Figure shows a test example for one day. Positive values represent charging and negative values represent discharging.

Theorems & Definitions (2)

• Theorem 1
• Proof 1: Proof of Theorem \ref{['convergent_for_eq']}