Solar power nowcasting using satellite images and deep learning advances

Accurate nowcasting of solar power production is crucial for power system operators and energy market participants. However, estimating real-time aggregated solar power production over a wide area is challenging due to spatial distribution, small-to- medium-range installations, and lack of smart metering facilities. This study investigates the application of deep neural architectures, such as U-Net, Vision Transformer, and Yolo, to extract cloud dynamic features from satellite imagery and estimate the aggregated solar power production of the Hellenic interconnected distribution system. The performance of these models is compared with a state-of-the-art solar power forecasting model that uses numerical weather predictions (NWPs) from the European Meteorological Service (ECMWF).