Exploring the Limits of Machine Learning in the Prediction of Solar Radiation

Predicting solar radiation at diverse time horizons is crucial for optimizing solar energy integration, ensuring grid stability, and regulating energy markets. Two main levels of time granularity are usually recognized as requiring different treatment: solar nowcasting for predictions up to 6 h, and solar forecasting for predictions beyond 6 h. Solar nowcasting typically relies on machine learning methods, while Numerical Weather Prediction (NWP) models are considered better suited for solar forecasting. The goal of this study was to explore the limits of machine learning in solar forecasting. Our results show that machine learning methods can be profitably used for predicting solar radiation beyond 6 h, with comparable performances to NWP models for day-ahead solar forecasting.