Data-Driven Chance-Constrained Regulation Capacity Offering for Distributed Energy Resources

TLDR

The study examines how a strategic aggregator offers regulation capacity for distributed energy resources such as plug‑in electric vehicles in power markets. The goal is to maximize aggregator revenue while limiting penalty risk from poor service delivery. The authors develop data‑driven, risk‑averse day‑ahead two‑stage stochastic and hour‑ahead distributionally robust chance‑constrained models, with a conic safe approximation, to handle uncertainties in DER parameters and sub‑hourly regulation signals, and validate them via numerical experiments. The resulting approach controls regulation quality according to risk aversion and, by learning parameter distributions from data, outperforms robust optimization and traditional chance‑constrained methods in accuracy and robustness cost.

Abstract

This paper studies the behavior of a strategic aggregator offering regulation capacity on behalf of a group of distributed energy resources (DERs, e.g., plug-in electric vehicles) in a power market. Our objective is to maximize the aggregator's revenue while controlling the risk of penalties due to poor service delivery. To achieve this goal, we propose data-driven risk-averse strategies to effectively handle uncertainties in: 1) the DER parameters (e.g., load demands and flexibilities) and 2) subhourly regulation signals (to the accuracy of every few seconds). We design both the day-ahead and the hour-ahead strategies. In the day-ahead model, we develop a two-stage stochastic program to roughly model the above uncertainties, which achieves computational efficiency by leveraging novel aggregate models of both DER parameters and sub-hourly regulation signals. In the hour-ahead model, we formulate a data-driven distributionally robust chance-constrained program to explicitly model the aforementioned uncertainties. This program can effectively control the quality of regulation service based on the aggregator's risk aversion. Furthermore, it learns the distributions of the uncertain parameters from empirical data so that it outperforms existing techniques (e.g., robust optimization or traditional chance-constrained programming) in both modeling accuracy and cost of robustness. Finally, we derive a conic safe approximation for it which can be efficiently solved by commercial solvers. Numerical experiments are conducted to validate the proposed method.