Feature Selection in Solar Radiation Forecasting

Abstract

Solar radiation (Rs), indispensable for life, fluctuates due to factors such atmospheric conditions. With equipment scarcity, diverse models estimate Rs, with Machine Learning gaining traction for forecasting efficiency. This study evaluates the impact of feature selection on Rs forecasting using time series lag values. Employing Simulated Annealing with Gradient Boosting Models (GBM), Random Forest (RF), and Classification and Regression Trees (CART), we enhance Rs prediction. Results unveil significant differences in model performance. CART exhibits a high normalized mean Absolute error (nMAE) (-0.718), indicating substantial deviation from target values. RF and GBM models show lower nMAEs (0.055, 0.043), implying superior accuracy. CART struggles with data variance, evident from its negative R-squared (-0.744), while RF and GBM models have positive R2 values (0.055, 0.043), suggesting better explanatory power. These findings underscore the importance of feature selection for accurate prediction, with RF and GBM models offering stable performance and promising reliable predictions.