IrradianceNet: Spatiotemporal deep learning model for satellite-derived solar irradiance short-term forecasting

Abstract

The presence of clouds is widely identified as the primary uncertainty in current surface solar global horizontal irradiance (GHI) forecasts. Despite a wealth of historical satellite-derived irradiance observations, only limited research has investigated this problem from a purely data-driven perspective, something that has seen tremendous success in related domains such as radar- and satellite-based precipitation short-term forecasting. This paper presents IrradianceNet, a novel satellite-based neural network for spatiotemporal forecasting of surface solar irradiance up to 4 h in the future over Europe. Our method is fully data-driven and needs no post-processing or calibration based on sparse ground-based measurements of irradiance. We demonstrate superior forecasting performance compared to several persistence models, the TV-L1 algorithm, and ERA5 reanalysis data for satellite-derived solar irradiance using the European SARAH-2.1 dataset. We also validate these results using ground-based pyranometer observations from the Baseline Surface Radiation Network. Our conclusions remain unchanged when we account for hourly and monthly seasonality. Finally, applying a simple cloud mask scheme, we demonstrate that our performance improvement arises due to a considerable reduction in cloudy pixel errors. This is initial evidence that purely data-driven methods might better approximate and infer future cloud dynamics and their impact on surface solar irradiance.