Solar irradiance separation with deep learning: An interpretable multi-task and physically constrained model based on individual–interactive features

As an essential part of solar forecasting and resource assessment, separation modeling has received widespread attention over the past half a century. Despite the numerous proposals thus far, most models are semi-empirical in nature, with limited accuracy. The other option, namely, machine-learning models, does not show a definitive advantage and usually lacks comparisons with the latest quasi-universal model. This study proposes an interpretable multi-task and physically constrained separation model based on individual–interactive features (IIF-IMCSM). The model has three blocks: (1) an informative predictor identification block, (2) an individual–interactive feature extraction block, and (3) a physically constrained irradiance component estimation block, each carrying some modeling innovations. Differing from other separation models, IIF-IMCSM simultaneously produces estimates for both the beam and diffuse components that satisfy the closure equation, and it overcomes the common drawback of lacking interpretability of machine-learning models. Based on five comprehensive datasets covering diverse radiation regimes of the globe, it is found that the overall normalized root mean square errors of IIF-IMCSM for beam normal irradiance and diffuse horizontal irradiance are 12.51% and 24.50%, as compared to 16.32%, 34.86%, and 13.47%, 26.56% for the top-performing semi-empirical and machine-learning models.