Decision-focused learning integrated with data-driven robust optimization for energy management

Addressing uncertainties in power systems is critical for enhancing renewable energy integration and ensuring overall system reliability. Existing research typically treats forecasting and optimization as separate processes, without effectively integrating the impact of predictive accuracy on the robustness and quality of downstream energy management decisions. To fill this gap, this study proposes a novel framework, Artificial Intelligence for Robust Optimization (AIROpti), which tightly integrates a forecasting model with a subsequent data-driven robust optimization model. Two cases are examined: a distributed energy system comprising twenty households and a market level energy system. Both analyses account for uncertainty in demand and electricity prices. Regarding predictive performance, the day ahead load forecasting model attains a symmetric mean absolute percentage error (SMAPE) of 17.99 % for the aggregated demand across twenty households and 3.82 % at the market level, demonstrating high accuracy. However, our experiments also show that improved forecast accuracy does not necessarily translate into more robust downstream energy management. AIROpti yields a 25.94 % to 44.34 % reduction in regret relative to high accuracy prediction models and a 3.99 % to 75.74 % reduction relative to traditional decision-focused learning baselines. Moreover, it enhances the robustness of operational strategies, thereby reducing worst case performance bounds by 13.77 % and 59.71 % across the two cases. These results highlight AIROpti's ability to produce forecasts that prioritize the robustness of energy management rather than merely achieving higher prediction accuracy.