A geographic multi-scale machine learning framework for predicting solar irradiation on tilted surfaces

Abstract

This study addresses the challenge of improving Global Tilted Irradiation (GTI) predictions, with Jordan serving as the case study. The novelty of the work lies in developing machine learning models that predict GTI at national, regional (3 regions), and city-specific (12 cities) levels, a previously unexplored approach in the literature. The research examines the comparative efficiency of using a single model for an entire country versus tailored models for individual regions and cities, shedding light on the trade-offs in model evaluation. Various regression models, including Neural Networks (NNs), Linear Regression (LR), Regression Trees (RTs), Ensemble of Regression Trees (ERTs), Support Vector Machine (SVM), and Kernel Approximation, were evaluated using performance metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and coefficient of determination (R²). NNs consistently performed best, achieving the lowest RMSE (1.5787 kWh/m²) and highest R² (99.8600 %) at the regional level. Sensitivity analysis further explored the impact of different time resolutions, revealing that monthly data outperformed daily data in terms of accuracy and computational efficiency. Ultimately, we conclude that region-specific models and monthly data resolution are optimal for practical GTI prediction.