Decision-Focused Learning-based Prediction Interval for Renewable Energy Curtailment Strategy

Abstract

With the increasing penetration of renewable energy sources, curtailment strategies have become indispensable for maintaining both system reliability and economic efficiency. Nevertheless, the uncertainty associated with renewable generation and electricity demand frequently undermines the effectiveness of such strategies. Prediction intervals provide a promising means of quantifying forecast uncertainty, yet traditional methods do not guarantee decision quality, as even small forecast errors can result in suboptimal curtailment outcomes. This paper proposes a decision-focused learning framework for renewable energy curtailment that directly integrates prediction intervals into operational decision-making. The framework employs a mixture density network to produce distributional forecasts from which prediction intervals are derived. A weighted negative log-likelihood surrogate loss is introduced to align predictive uncertainty with downstream decision performance, enabling end-to-end training even in non-differentiable optimization environments. Simulation results on a microgrid with photovoltaic generation and conventional units show that the proposed approach consistently outperforms existing methods.