Hybrid model improves the ability to separate the diffuse component of minute-scale global solar radiation

The diffuse component (R_d) of solar radiation (R_s) has a fertilizing impact on vegetation photosynthesis and therefore influences terrestrial carbon sinks and the water cycle. Measurements of R_d are often lacking and the separation of R_d from R_s is a frequently requested task. Recent studies suggest that the hybrid models have a robust predictive ability among the data-driven approaches and have been applied successfully in estimating R_s. However, few have been used to estimate R_d. Therefore, in this study, a hybrid model constructed from the outputs of empirical models and an Artificial Neural Network (ANN) was proposed to estimate minute-scale R_d. This study systematically parameterized five candidate empirical models using minute-scale data, which indicated that the empirical Starke3 model performed best. The ANN models were constructed using the same inputs as the five parameterized models. The results showed that the best-performing ANN4 model, which uses the same inputs as the empirical Starke3 model, could only achieve similar performance. However, the hybrid model outperformed both the best empirical model and the best ANN model, regardless of location and climate, achieving the highest R² of 0.8 and the lowest rRMSE of 30.32% on the test dataset. Furthermore, this study demonstrated that the hybrid model can effectively reduce the residuals. Taken together, we foresee that this hybrid method has the potential to be extensively utilized for estimating near real-time R_d in the future.