Site adaptation of global horizontal irradiance from the Copernicus Atmospheric Monitoring Service for radiation using supervised machine learning techniques

Abstract

Satellite and reanalysis-derived solar products have gained great attention due to the inadequate number of radiometric stations worldwide, however, they are associated with considerable uncertainties. This study deals with the ground-based validation of Global Horizontal Irradiance from CAMS radiation service (GHICAMS) and the application of supervised machine learning algorithms (MLAs) to site-adapt GHICAMS. The validation of GHICAMS against measurements shows significant systematic and dispersion errors for all-sky (nMBE = 4.9% and nRMSE = 15.7%) and cloudy conditions (nMBE = 17.6% and nRMSE = 38.8%). Under clear skies, CAMS performs adequately (nMBE <1% and nRMSE <5%). All MLAs lead to reduced errors for the site-adapted irradiances. MBE is improved by more than 50%, accompanied by significant reductions in RMSE for various solar zenith angles and cloud fractions. The best results are revealed for the tree-based MLAs and especially for Random Forests.