Accurate quantification and robust predictive modeling of diffuse solar irradiance ratios are essential for optimizing photovoltaic (PV) system efficiency, particularly in regions characterized by pronounced climatic heterogeneity. This study investigates the spatial variability of diffuse horizontal irradiance (\({I}_{d}\)) and short-term forecasts of the diffuse fraction (\(d\)) across six representative Saudi Arabian stations, encompassing coastal, desert, and high-altitude environments. A hybrid stacking ensemble model was developed using a multi-algorithmic framework integrating Random Forest (RF), Light Gradient Boosting Machine (LGBM), and Gradient Boosting Regressor (GBR) algorithms within a meta-learning architecture. Model optimization was achieved through recursive feature selection and gradient-boosted meta-learning, to effectively capture complex nonlinear dependencies among atmospheric predictors under diverse sky conditions. Results revealed pronounced geographic variability in diffuse solar irradiance, with coastal stations exhibiting higher \({I}_{d}\) values due to maritime aerosol scattering; Jeddah recorded the highest mean \({I}_{d}\) (283.7 W/m2) and maximum diffuse fraction (\(d\) = 0.95). In contrast, high-elevation sites such as Abha exhibited the lowest mean irradiance (142.7 W/m2). The proposed ensemble model demonstrated superior predictive performance relative to individual learners, achieving up to a 62.48% reduction in \(RMSE\) and maintaining high correlation coefficients (R = 0.94–0.97) across stations. Forecast-horizon evaluation confirmed robust short-term accuracy, with \(RMSE\) increasing moderately from 0.0292 W/m2 for a 12-h horizon to 0.0478 W/m2 for a 48-h horizon, while seasonal assessments revealed optimal performance during spring and summer, with normalized \(RMSE\) values as low as 11.64%. These findings highlight the ensemble model’s robustness and adaptability, providing a scalable, high-fidelity approach for operational solar forecasting and energy management.
