Comparative Evaluation of Machine Learning Models for Global Horizontal Irradiance Estimation in an Arid Coastal Climate

Abstract

Accurate estimation of global horizontal irradiance (GHI) is relevant for characterizing solar resources in regions with limited measurement infrastructure. This study compared six machine learning models Polynomial Ridge Regression, Decision Tree, Random Forest, XGBoost, Artificial Neural Network, and K–Nearest Neighbors for the contemporaneous estimation of GHI at a five-minute resolution in an arid coastal climate. After quality control and restriction to daytime periods, 45,024 observations were analyzed using five external chronological blocks with an expanding-window scheme, generating 31,517 out-of-sample estimates. XGBoost achieved the highest R2(0.657±0.192) and the lowest RMSE (108.01 ± 13.67 W m−2), whereas Random Forest yielded lower MAE, WMAPE, and MASE values. DM–HAC sensitivity analysis favored XGBoost under squared-error loss for six of the seven evaluated bandwidths, whereas no significant difference between XGBoost and Random Forest was found under absolute-error loss. None of the models achieved the nominal conformal coverage level of 90%; XGBoost showed the highest empirical coverage and the narrowest prediction intervals (PICP = 0.791; PINAW = 0.312). Predictor-set reduction improved the performance of four of the six algorithms. Overall, XGBoost and Random Forest exhibited complementary performance profiles, indicating that model selection should jointly consider predictive accuracy, temporal stability, predictor sensitivity, and uncertainty.