Direct solar radiation prediction using multi-model evaluation with trignometric cyclic process

Accurate prediction of renewable sources in general and solar radiation is critical for optimal integration of solar energy systems. The study explores eight Machine Learning models namely Linear Regression Model (LRM), Random Forest Regressor (RFR), Gradient Boosting Regressor (GBR), Gaussian Process Regression (GPR), Artificial Neural Network (ANN), k-Nearest Neighbors (NN), Support Vector Regression (SVR), and Deep Learning (DL) for predicting the direct solar radiation at climatically distinct six sites in Saudi Arabia. Models are evaluated using eight statistical metrics along with time series and absolute error analyses. The present work introduced the Trigonometric Cyclical Encoding (TCE), which significantly improved the temporal learning. Comparative SHAP-based analysis revealed that TCE enhanced the explanatory power of temporal features by 49.26% and 53.40% for monthly and daily cycles. Results show that DL achieved the lowest Root Mean Square Error (RMSE) and highest coefficient of determination, while ANN consistently indicated high accuracy at all the sites. Error and time series analyses denoted stable predictions by ANN and DL; whereas LR, RFR, and k-Nearest Neighbors (NN) showed larger fluctuations. The proposed TCE technique additionally improved the model output by maintaining the overall fitness of the models between 81.79% and 94.36% in all scenarios. This study reinforces the effective planning of solar energy integration in different climatic conditions.