Towards Carbon Neutrality: A Novel STIRPAT- XGBoost-SHAP Framework for Provincial Energy Transition Prediction in China

China's commitment to carbon neutrality by 2060 demands accurate and interpretable forecasting tools at the provincial level. This study develops and validates a hybrid framework integrating the STIRPAT model, XGBoost machine learning, and SHAP interpretability analysis to forecast coal and total energy consumption across six representative Chinese provinces (2005 -2021). The STIRPAT model reveals industrial structure as the dominant driver of coal dependence, while SHAP confirms structural consistency and highlights nonlinear effects of urbani sation and income. The XGBoost model achieves competitive forecasting performance (Mean Absolute Percentage Error (MAPE): 5.26% for coal and 3.02% for total energy) and effectively captures regional disparities in coal transition trajectories. These results support differentiated and structurally grounded policy interventions, offering practical guidance for subnational energy planning under China's dual-carbon strategy. The framework offers broad applicability to other structurally diverse and data-constrained contexts.