Can Deep Learning Replace Cloud Motion Vectors?

Satellite-based (SAT) methods are widely used to forecast surface solar irradiance up to several hours ahead. This study applies the established Heliosat method to derive irradiance from Meteosat Second Generation images. As a key parameter the cloud index (CI) is derived from these images, quantifying the impact of clouds on surface solar irradiance. Conventional SAT-methods utilize cloud motion vectors (CMVs) from consecutive CI-images to predict future cloud conditions and subsequently retrieve irradiance. In this study, we introduce HelioNet, a UNet-like neural network designed to predict future CI situations directly from sequences of preceding CI-images. We benchmark forecasts of two HelioNet configurations against CMV and persistence over a full year (2022), with lead times (LT) up to 4 hours. HelioNet15min recursively generates forecasts at 15-minute resolution. HelioNethybrid begins with forecasts at 15-minute resolution for LT≤45 min, then uses a 45 minute-resolved model to forecast all remaining LT-steps. HelioNet15min achieves root mean square error (RMSE) improvements of up to 15% over the CMV model within the first hour on image level. HelioNethybrid shows superior performance for all LT across all metrics considered, with an average RMSE improvement of >11% on image and >6% on irradiance level.