A Multi-Source Deep Learning Framework for Very Short-Term Solar Irradiance Nowcasting with Satellite data and All-Sky Images

Accurate very short-term solar irradiance forecasting is critical for the stability and efficiency of photovoltaic (PV) power generation and grid management. This study introduces a multi-source deep learning framework that integrates satellite cloud coverage data and all-sky images to predict Global Horizontal Irradiance (GHI) with a 15-minute forecast horizon. The proposed methodology employs a dual-input Convolutional Neural Network (CNN) to process satellite-derived cloud top height and cloud type images alongside high-resolution all-sky images and meteorological data. The framework is validated using datasets from two locations, including the Politecnico di Milano campus, spanning 17 months of observations. Preliminary results demonstrate the model's ability to outperform baseline persistence methods, particularly under clear-sky conditions, achieving a forecast skill of up to 27%. However, challenges remain in handling complex weather scenarios, such as overcast or rainy conditions, where performance declines. This research highlights the potential of combining multi-source data and advanced deep learning techniques for improving solar irradiance nowcasting, paving the way for more reliable renewable energy integration.