Decision-Focused Retraining of Forecast Models for Optimization Problems in Smart Energy Systems

In order to enable the energy transition, a higher share of renewable energy sources is required in the electricity grid. However, the volatile nature of these renewable sources can lead to stability issues. Therefore, countermeasures must be integrated into modern electricity grids to maintain stability. However, many countermeasures rely on optimization problems on multiple grid levels to be successfully integrated. Furthermore, these optimization problems often require forecasts that are tailored to deliver value for the considered optimization problem. Nevertheless, existing applications of decision-focused learning to provide this value scale poorly for energy system optimization problems. Therefore, we propose a novel method called Decision-Focused Retraining that combines prediction-focused learning and decision-focused learning. In this method, an existing forecasting model is retrained to generate forecasts delivering increased value for the optimization problem. First, a prediction-focused learning approach with a suitable base loss is used to pre-train the forecasting model. Afterward, the model is fine-tuned by combining a global instance-independent surrogate NN with the prediction-focused base loss to optimize the forecasting model. We evaluate our approach on an exemplary optimization problem, the dispatchable feeder optimization problem, considering over 199 buildings, which leads to an improvement of at least 7.29%.