Comparative Analysis of Solar Radiation Forecasting Models: Unveiling the Potential of Feature Selection Techniques

Solar Radiation (Rs) stands as the quintessential force fueling radiance, warmth, and life on Earth. Yet, its intermittent nature poses challenges to seamless integration into energy systems. This study pioneers a methodology employing feature selection (FS) methods to refine Rs prediction accuracy. Principal Component Analysis (PCA) is synergized with diverse machine learning (ML) methods, including Random Forest (RF), Gradient Boosting Models (GBM), Logistic Regression (LR), Classification and Regression Tree (CART), and Decision Tree (DT). Concurrently, Recursive Feature Elimination (RFE) enhances model performance. Quantitative metrics reveal LR and RF as top performers, showcasing the lowest normalized Mean Absolute Errors (nMAE). GBM slightly surpasses them, while CART records the highest nMAE. Under the RFE approach, strong predictive accuracies are observed for CART, LR, and GBM models. Comparison across different FS methods highlights RF’s consistency in achieving low nMAE under both PCA and RFE. LR demonstrates predictive accuracy under both methods, while GBM exhibits a competitive edge, particularly under PCA. DT and CART models exhibit varying performance, with DT displaying specific metrics and CART registering the highest nMAE under PCA. GBM, DT, and CART models demonstrate adaptability to RFE, each showcasing distinctive predictive accuracies. The stability of these models during repeated cross-validation is underscored by corresponding nMAE standard deviations (std). This detailed examination elucidates the nuanced influence of FS on ML models, offering insights for optimizing solar energy (SE) integration into power grids. Overall, LR and RF emerge as reliable performers, while varying outcomes highlight the significance of model-specific considerations in enhancing SE integration into power grids.