Extreme Gradient Boosting for Day-Ahead Photovoltaic Output Power Forecast Refinement

The deployment of Photovoltaic (PV) systems benefits any energy system, provided that the corresponding uncertainties are mitigated by means of forecasting techniques. While nowcasting is gaining interest for the short-horizon PV prediction, long-horizon forecasters ensure reliable and efficient scheduling of the energy system. In this paper, we focus on microgrids, where typically a Day-Ahead (DA) Dispatch Plan (DP) is initially submitted and then adjusted to account for actual realizations. However, DA PV forecasts remain highly sensitive to weather forecast accuracy, which has shown limited improvements, potentially undermining the DP reliability. Our research aims to improve a well-established Physical Hybridized Artificial Neural Network (PHANN) model by means of a data-driven refinement method based on an eXtreme Gradient Boosting (XGBoost) algorithm. We tested this machine-learning enhancement method on a real-world microgrid in Politecnico di Milano, obtaining a 1.4% overall improvement and promising daily results when compared to the original model. Moreover, we lower-bounded this refinement by means of an hourly robust persistence, to identify any potential areas for improvement.