SHAP-Based Interpretable Gradient Boosting for Cross-Hemispheric Renewable Energy Green Score Generalization

Abstract

Renewable energy has emerged as a critical solution to global energy sustainability challenges. Accurate prediction of renewable energy favourability across diverse geographic locations is essential for effective energy planning and policy formulation. This paper presents a comprehensive machine learning-based framework for predicting a composite renewable energy green score derived from real-time meteorological and geographic data across ten global cities. The dataset, comprising 1,920 hourly records of shortwave solar radiation, wind speed at 100 meters, geographic coordinates, and temporal features spanning January 2026, was subjected to rigorous exploratory data analysis and feature engineering. Six machine learning models were implemented and evaluated: Random Forest, XGBoost, LightGBM, Gradient Boosting, Support Vector Regression, and Multi-Layer Perceptron Neural Network. Models were assessed using R², Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), complemented by 5-fold cross-validation to ensure result robustness. SHAP (Shapley Additive explanations) analysis was further employed to enhance model interpretability. Experimental results demonstrate that Gradient Boosting achieved the highest predictive performance with R² = 0.9984, MAE = 0.0120, and RMSE = 0.0205, outperforming all competing models. Learning curve analysis confirmed its superior generalization capability with a train-validation gap of 0.0053, while residual analysis verified near-zero prediction bias (mean residual = 0.0015). SHAP analysis identified wind speed at 100m and shortwave solar radiation as the two dominant predictive features, collectively contributing over 99.9% of the model's predictive power.