A feature optimized machine learning framework for solar radiation prediction in the high-altitude Western Himalayan region

Accurate prediction of solar potential is essential for efficient integrating photovoltaic system particularly in geographically complex regions where meteorological variability poses challenges to grid stability. This study focuses on improving the solar potential estimation in arid and semi-arid Himalayan regions through feature selection and machine learning approaches. The study evaluates seven feature selection methods such as Pearson Correlation, Mutual Information, XG Boost Importance, LASSO, Random Forest Importance, Permutation Importance and SHAP to identify most influential parameters for solar radiation. The results showed that temporal variable is most important followed by temperature in prediction accuracy of solar radiation. K-based feature selection results show that the selected five feature subset provides an effective accuracy–complexity trade-off, reducing RMSE by approximately 17.5% (from 75.17 to 62.01 W/m²) while maintaining competitive predictive performance compared with the full-feature models. Out of all the models tested, MLP and XGBoost yield optimal results with all features used (RMSE is about 59.86 W/m² and R² is around 0.953). Interestingly, even after using just five selected features, XGBoost and LSTM still provide accurate predictions (RMSE is about 62.46 W/m² and R² is roughly 0.949). The results depict that the model framework has managed to provide highly accurate predictions (R² > 0.95) while reducing input dimensionality, showing the potential of the approach for practical solar energy prediction scenarios.