Day-Ahead Solar Power Forecasting Using a Hybrid Model Combining Regression and Physical Model Chain

Solar power forecasting is essential for integrating PV plants into power grids, ensuring stability and aiding system operators (SOs) in decision-making. However, existing day-ahead models struggle with rapid weather changes, while deep learning models require extensive historical data, making them impractical for new PV plants.This study proposes a hybrid approach combining the XGBoost algorithm for hourly solar irradiance prediction using Numerical Weather Prediction (NWP) data and a physical model to convert irradiance into power. The XGBoost model is periodically retrained via a sliding window mechanism to adapt to dynamic weather conditions.A case study using two years of 271 kWp PV data from NIST (US) and historical NWP data from ECMWF ENS for GHI forecasting, alongside ECMWF HRES for power conversion, demonstrated the method’s effectiveness. Using just one week of historical data for initial training, the model achieved an nRMSE of 13.35%–13.53%, nMAE of 6.9%–7.03%, and nMBE of -2.03% to -0.29%. The proposed approach improves PV forecasting reliability for new plants with limited data, serving as an intermediary solution until sufficient historical data is available for deep learning models.