Decision-Focused Learning for Optimized Participation of an Industrial Consumer in Energy-Only and Reserve Markets

Many decision-making applications in modern power systems involve solving two sequential problems: (i) predicting unknown parameters and (ii) optimizing decisions under prediction uncertainty. Conventionally, the prediction model is learned by minimizing a statistical error (with respect to the true observations), without considering the impact of the predictions on the quality of decisions. In contrast, this paper aims to improve this gap by investigating Decision-Focused Learning wherein the optimization model is integrated into the training pipeline of the prediction model. The goal is to learn a predictor that endogenously identifies the salient features (e.g., timing of peaks, duration of valleys, etc.) that lead to the best downstream decisions. This end-to-end learning approach is applied to the day-ahead scheduling problem of a flexible consumer participating in both energy-only and reserve markets. A machine learning model is used to forecast the day-ahead energy prices, which has been trained with a value-oriented loss function, referred to as regret that evaluates the quality of the operation decisions. Results demonstrate the ability of Decision-Focused Learning to improve decision quality, leading to economic benefits for the consumer. Outcomes also reveal initializing training with a conventional least-squared-error, before using the regret, enhance model performance.