Interpretable Data‐Driven Modeling for Solar Radiation Forecasting

Accurately forecasting solar irradiance is of paramount importance for the effective management of renewable energy resources and the optimization of solar power infrastructure. This study introduces an array of advanced predictive architectures driven by 13 meteorological and temporal variables: month, day, hour, minute, air temperature, dew point, relative humidity, wind velocity, gust, atmospheric pressure, precipitation rate, precipitation accumulation, and ultraviolet (UV) exposure. To rigorously train and validate these data‐centric models, a massive collection of 32,765 observational data points was utilized. The investigation explored 13 distinct machine learning algorithms, specifically decision tree, adaptive boosting (AdaBoost), random forest, K‐nearest neighbors (KNN), ensemble learning, convolutional neural networks (CNN), support vector regression (SVR), multilayer perceptron–artificial neural network (MLP–ANN), radial basis function–artificial neural network (RBF–ANN), alongside linear, ridge, lasso, and elastic net regressions. Assessments of statistical dependencies highlighted that UV exposure possessed the most prominent positive correlation with the target variable (0.89), while relative humidity showed the strongest inverse relationship (−0.71). Moreover, SHAP interpretations confirmed that UV intensity and the specific hour of the day are the primary drivers governing solar radiation variability, followed sequentially by relative humidity and ambient air temperature. Among the evaluated models, random forest achieved the strongest held‐out test performance in terms of coefficient of determination ( R ² ) and mean squared error (MSE), with a test R ² of 0.9047 and a test MSE of 3862.96, whereas AdaBoost achieved the highest overall R ² (0.9910) and the lowest overall average absolute relative error (AARE, 2.83%). Accordingly, random forest was selected as the primary model for subsequent interpretation because model selection emphasized predictive generalization on the held‐out test set rather than aggregate performance across the training and test samples. Notably, it accomplished this precision while successfully averting the pronounced overfitting defects that plagued the decision tree architecture. These results indicate that Random Forest effectively captures nonlinear relationships between meteorological predictors and solar radiation. Ultimately, the developed computational framework functions as an efficient, low‐cost solution for solar radiation prediction, delivering vital advantages for sustainable power generation, energy grid integration, and solar engineering projects.