Our previous FES project developed learning-based control algorithms to support large-scale integration and reliable operation of Distributed Energy Resources (DERs). That work framed DER coordination as a decision-making problem under uncertainty and primarily used the predict-then-optimize (PTO) paradigm, where forecasting and optimization are trained separately. While effective, PTO trains forecasts using prediction error rather than the quality of the downstream control decisions, resulting in objective mismatch.

Since completing the FES project, we have advanced this line of work by applying decision-focused learning (DFL), which trains forecasting models directly for decision quality (i.e. control performance). A 2024 benchmarking paper suggests that DFL is not guaranteed to outperform PTO and that the performance of DFL techniques is highly dependent on the problem setting. Using a real water treatment plant with solar panels and battery storage, we benchmarked ten DFL algorithms against PTO and demonstrated reductions of 24.92% in operating costs and 28.75% in carbon emissions. In this project, we use DFL for risk-aware control of distributed energy resources.