Hourly solar radiation estimation and uncertainty quantification using hybrid models

Solar energy, considered to be the most abundant renewable resource, is one of the most effective methods for reducing carbon emissions. The quantification of the uncertainty in the model estimates due to the uncertainty in the input parameters has received very little attention, although models with different computational principles have been developed to estimate surface solar radiation. This study aims to establish and compare four hybrid models by coupling a physical model with machine learning models. Uncertainty in model estimations caused by uncertainty in cloud optical thickness, aerosol optical depth, precipitable water vapor, and total column ozone is quantified. The results of the radiative transfer model reveal a strong dependence on aerosol optical depth, cloud optical thickness, and total column ozone, but not on precipitable water vapor. The average uncertainties in the radiative transfer model estimates caused by the uncertainties in aerosol optical depth, cloud optical thickness, precipitable water vapor, total column ozone, and all of them together reached 37.76, 182.19, 22.76, 3.00, and 219.67 W m−2 at all sites, respectively. Uncertainties in atmospheric parameters greatly limit the performance of hybrid models. RTM-RF has the strongest robustness compared to RTM-XGBoost, RTM-CatBoost, and RTM-LightGBM. The proposed hybrid model can be considered as a pertinent decision-support framework for the estimation of solar radiation components to further support clean energy utilization. Optimization of cloud inversion algorithms to improve the product accuracy of cloud optical properties over land and oceans is central to improving the accuracy of surface solar radiation estimates.