Decoding the dynamics of daily total global solar radiation: Computational learning strategies powered by explainable AI

This study investigates the prediction of daily total global solar radiation using advanced machine learning and hybrid models. The input variables include various meteorological and environmental parameters such as total precipitation, average snow height, total solar intensity, existing snow height, soil temperature, actual pressure, average temperature, cloudiness, minimum cloud height, and relative humidity. Based on SHAP value analysis, 5 cm Soil Temperature (+67.08) and Total Solar Intensity (+34.76) emerged as the most influential features for the models, while parameters like Current Snow Depth (+0.33) and Average Snow Depth (+0.36) contributed minimally. The analysis spans a time period from 2006 to 2022 using daily data. The dataset is divided into 70% for training and 20% for validation, with 10% of the training data allocated for testing. Several models were employed, including AdaBoost, CatBoost, Deep Belief Network (DBN), Light Gradient Boosting Machine (LGBM), Adaptive Neuro-Fuzzy Inference System (ANFIS), Echo State Neural Networks (ESN), and Autoencoder-based regression. According to the analysis results, the Catboost model produces the most accurate predictions with RMSE: 97.25 KW.hr/m 2 , MAE: 73.64 KW.hr/m 2 , AIC: 17075, NSE: 0.65 and KGE: 0.72 values, while the DBN model shows the second best model with RMSE: 98 KW.hr/m 2 , MAE: 74.48 KW.hr/m 2 , AIC: 17103, NSE: 0.64 and KGE: 0.74. The analysis results have the potential to provide information to decision makers and planners, especially in the evaluation of the performance of solar panels, the optimization of agricultural processes, and the development of climate change adaptation strategies.