Data-Driven Performance Forecasting of Photovoltaic Systems from Macroscopic Data

This paper presents a data-driven approach to forecasting the performance of PhotoVoltaic (PV) systems using macroscopic, aggregated data. Several Machine Learning models, from linear baselines to ensemble methods, were evaluated and compared in terms of monthly PV energy production forecast. The most accurate and robust results were obtained using an eXtreme Gradient Boosting (XGBoost) model trained on the full historical dataset with domain-informed feature engineering. The models were trained and tested on data from hundreds of PV systems distributed across the Italian peninsula, each providing varying amounts of historical monthly production data. Finally, the XGBoost has been compared to a deterministic formula to highlight the improvement obtained by using data-driven methods.