• Decision-focused learning methods are proposed for flexible building operations. • Two physics-informed heuristic losses are defined for end-to-end optimal controls. • Data experiments have been conducted based on 102 datasets from 9 buildings. • Proposed methods consistently outperform two-stage counterparts by 6.7% to 28.7%. • The methods can provide implicit regularization to ensure better generalization. Conventional control optimization methods for building energy systems typically adopt a two-stage decision-making paradigm, where predictive modeling and optimization are conducted in an independent and sequential manner. Such a paradigm assumes that the prediction accuracy would directly translate into decision quality, which may not always hold in practical engineering scenarios. To enhance the performance of model predictive controls in building energy management, this study proposes a novel decision-focused end-to-end learning method that transforms domain-specific knowledge into heuristic surrogate losses, thereby embedding downstream optimization objectives into the prediction model training process. Two physics-informed surrogate losses are designed for regulating building energy storage system operations, one weights prediction errors by their marginal contribution to operation costs considering price dynamics and grid interactions, while the other quantifies deviations in energy storage system operations from perfect foresight in terms of energy storage state trajectories and potential arbitrage opportunities. Comprehensive data experiments have been conducted using 102 40-day datasets collected from 9 buildings. The results demonstrate that the proposed methods can consistently outperform the conventional two-stage approach, achieving evident reductions in normalized regret by 6.7% to 28.7%. A counterintuitive finding reveals that the decision quality improves despite higher prediction errors, and the heuristic losses proposed can provide implicit regularization to ensure the generalization performance under unseen operating conditions. This study provides a novel perspective on developing computationally efficient, easy-to-implement, and interpretable control methods for enhanced building energy flexibility.
Fan et al. (Fri,) studied this question.