Solar radiation (Rs) is a major renewable energy source and also a crucial factor in designing solar panels, determining water requirement, and irrigation scheduling. In this study, meteorological parameters (air temperature, average air temperature, and relative humidity; Scenario 1), satellite image-based indices (normalized difference vegetation index: NDVI and land surface temperature: LST; Scenario 2), and their combination (Scenario 3) were used as predictors of Rs simulator models in Mashhad watershed (2005-2015). To this end, three different transfer function algorithms of the multi-layer perceptron (MLP), namely Levenberg-Marquardt backpropagation (MLP-LVM), gradient descend backpropagation (MLP-GDB), and batch training with weight and bias learning rules (MLP-BTWB), as well as two other machine learning models, M5 Tree and XGB (eXtreme Gradient Boosting), were employed. In addition, hybrid models obtained through coupling the best MLP transfer function with genetic algorithm (MLP-LVM-GA) and coupling Variational Mode Decomposition with M5 Tree and XGB (VMD-M5 Tree and VMD-XGB, respectively) were also investigated. In the case of MLP-based models, the third scenario proved to be more efficient, with RMSEs equal to 404.2, 424.5, and 423.8J.cm-2.day-1 for the aforementioned transfer functions, respectively. The results showed that GA coupled with the best transfer function algorithm of MLP improved Rs estimates, decreasing estimation error by 14.7, 30.65, and 32.67% in the three defined scenarios, respectively. In training set, VMD-M5 Tree and VMD-XGB hybrid models outperformed corresponding base models under all three scenarios; but in testing set, estimation error was decreased only in the third scenario (by 37 and 42%, respectively). Overall, all models (base and hybrid) had the least Rs estimation errors in the third scenario (application of both meteorological parameters and satellite image-based data).
Read full abstract