Deep Learning for Satellite-Based Solar Irradiance Forecasting
Abstract
This work presents the solar irradiance evaluation of a novel intraday solar forecasting method that applies deep learning to geostationary satellite imagery. The U-Net architecture is used to predict the future cloud position up to 5 hours ahead, which is then converted into a solar irradiance forecast using a semi-empirical satellite-to-irradiance model. The method is implemented using satellite earth reflectance images from the GOES-16 visible channel. It is trained to maximize reflectance prediction accuracy (image prediction) and is evaluated for both satellite reflectance and solar irradiance forecast (Global Horizontal Irradiance, GHI). The solar irradiance prediction is evaluated using high-quality ground measurement for a site located in the Pampa Húmeda region of southeastern South America. The research demonstrates significant improvements over traditional methods such as those based on the cloud motion vector estimation and extends previous research reported in Marchesoni-Acland et al. ( 2023 ).
Keywords
Solar forecasting, deterministic forecast, satellite imagery, GOES16, GHI