Extreme-Event-Informed Post-Processing of Numerical Weather Predictions for Enhanced Photovoltaic Power Forecasting

Abstract

Numerical weather prediction models tend to exhibit larger forecast errors during extreme weather events, such as heat waves and cold spells. In this context, the systematic biases of the Global Forecast System (GFS) in irradiance and temperature propagate nonlinearly through the thermal derating of photovoltaic modules, compromising solar generation estimates. To address this problem, we propose a machine learning post-processing framework that incorporates explicit extreme-event indicators, evaluating five algorithms (Gradient Boosting, Random Forest, Ridge, XGBoost, and LightGBM) using strict temporal cross-validation with five splits. The framework is assessed on a case study comprising 2.5 years of measurements from the solarimetric station of the Federal University of ABC (UFABC), Santo André, Brazil, where Random Forest and LightGBM consistently outperformed linear models, reducing irradiance MAE to 51.98 W/m2 and temperature MAE to 1.01 °C, while Ridge showed the poorest performance due to its inability to capture nonlinear relationships. The meteorological bias correction reduced the overall AC power MAE from 36.2 kW to 20.8 kW, increased the coefficient of determination from 0.633 to 0.869, and brought the daily energy error below 1.1% across all weather regimes, indicating that explicit extreme-event encoding effectively mitigates the nonlinear propagation of meteorological bias into photovoltaic power estimates.