Machine Learning Based Diagnostic Comparative Modeling of Region-Specific Solar Irradiance in Nigeria

Abstract

This study evaluates and compares the predictive performance of Multiple Linear Regression (MLR), Support Vector Regression (SVR), and Random Forest (RF) models for estimating solar irradiance across two climatically distinct regions of Nigeria, namely the North and the South. The models were adopted using important meteorological predictors, including minimum and maximum temperatures (tmin, tmax), relative humidity (rehu), wind speed (wins), evaporation pitch (evpi), sunshine hours (sunh), and solar radiation (radi). The analytical framework integrates Pearson correlation analysis, Variance Inflation Factor (VIF) assessment, regression diagnostics, residual distribution analysis, variable importance measures, and performance evaluation using Root Mean Square Error (RMSE) and the coefficient of determination (R²). Correlation analysis revealed weak to moderate relationships among the predictors, while all VIF values were below 2, indicating the absence of multi-collinearity and supporting the inclusion of all variables in the multivariate models. Model adequacy was further confirmed through diagnostic and residual analyses, which also highlighted clear regional variations in model performance. In the northern region, the RF model demonstrated superior predictive capability, achieving the lowest RMSE (0.0179) and the highest R² (0.93), whereas SVR and MLR exhibited substantially weaker performance with R² values of 0.35 and 0.23, respectively. In contrast, in the southern region, SVR emerged as the most accurate model (RMSE = 0.0078, R² = 0.89), marginally outperforming RF (RMSE = 0.0099), while MLR remained the least effective (RMSE = 0.0132, R² = 0.62). These findings were supported by residual and scatter plot analyses, which showed closer agreement between observed and predicted values for SVR and RF. Variable importance analysis consistently identified maximum temperature (tmax) as the most influential predictor in both regions. Overall, the results emphasize the importance of region-specific model selection and rigorous diagnostic validation for producing reliable and interpretable solar irradiance forecasts.