Very short-term solar radiation forecasting with all-sky images and meteorological data

The ability to forecast solar radiation accurately on a very short time scale is essential for optimizing the performance and reliability of solar power systems. This work introduces a deep learning approach based solely on Convolutional Neural Networks (CNNs) to predict solar radiation over a 5-minute horizon. By processing sequences of infrared all-sky images alongside meteorological variables, the proposed model captures intricate spatial patterns and relevant environmental factors influencing solar irradiance. Experimental evaluation reveals that the proposed approach achieves a forecast skill improvement of 8.67% over the persistence model, with a root mean square error (RMSE) of 87.53 W/m² compared to 95.84 W/m², and a mean absolute error (MAE) of 42.48 W/m² versus 44.50 W/m². These results highlight the capability of CNN-based models to enhance the reliability of very short-term solar radiation forecasting and support the advancement of intelligent renewable energy management systems.