Hybrid regression–machine learning modelling for mapping of monthly average daily diffuse solar radiation in Thailand using satellite-based data

Diffuse solar radiation (DSR) plays a crucial role in solar energy assessment and atmospheric research; however, its direct measurement in Thailand is limited to only a few stations. This study aimed to produce the first high-resolution DSR map for Thailand using satellite-based datasets to address these measuring gaps. Data from 2012 to 2024 were collected from eight ground stations: four providing directly measured DSR and the other four stations providing DSR estimated from global and direct solar radiation for map validation. A diffuse fraction model was developed using the clearness index and atmospheric parameters. Satellite-based datasets from multiple sources were bias-corrected using quantile mapping (QM) to ensure consistency with ground measurements. Significant predictors were identified using t-statistics and p-values, followed by subset selection to optimize the regression structure by establishing a baseline model with significant variables and progressively evaluating additional variables. Multiple linear regression (MLR) models with various parameter combinations were then compared to determine the most accurate configuration. The modelling framework adopted a hybrid approach in which the optimal predictor set from MLR and subset selection was subsequently used to develop cross-validated (CV) and artificial neural network (ANN) models. The comparative analysis showed that Model M6A, derived from the CV method, outperformed all other models (MBD = 2.81%, RMSD = 13.49%, IOA = 0.957). This model was then used to generate high-resolution DSR maps across Thailand, which were validated against an independent dataset from four ground stations (MBD = 2.58%, RMSD = 16.71%, IOA = 0.886), confirming the reliability of the DSR maps for further solar applications. The proposed hybrid framework effectively integrates regression-based predictor selection with machine learning refinement, improving model robustness and spatial representation. The resulting high-resolution DSR climatology provides a valuable dataset to support solar energy planning and climate-related research in Thailand.