Data-Driven Prediction of Provincial Energy Consumption in China and Its Implications for Sustainable Development Policy

Reliable prediction of provincial energy consumption is essential for translating national energy-saving and carbon-neutrality objectives into differentiated provincial action. National estimates can obscure regional heterogeneity, whereas prediction studies that focus only on accuracy provide limited guidance on which industrial, electricity, and emissions conditions require intervention. Using 1003 observations from 31 provincial-level regions for 1990–2025, this study develops an interpretable artificial-intelligence-assisted framework combining six tree-ensemble models, Optuna optimisation, Shapley additive explanations (SHAP), and partial dependence analysis. On the original random test split, the gradient boosting decision tree (GBDT) achieved a coefficient of determination (R2) = 0.9890 and root mean square error (RMSE) = 1011.59. Across 30 repeated random seeds using Optuna-optimised hyperparameters, categorical boosting (CatBoost) and GBDT obtained mean R2 values of 0.9871 and 0.9865, respectively. In a blocked temporal holdout using 1990–2017 for training and 2018–2025 for testing, adaptive boosting (AdaBoost) and GBDT retained R2 values of 0.9202 and 0.9131. SHAP and partial dependence results identify industrial value added, electricity generation, carbon emissions, industrial structure, population, and regional context as the principal sources of differentiated energy consumption pressure; the highest predictions occur when industrial output, electricity supply, and carbon emissions are jointly elevated. These results support a result-linked policy framework: provinces dominated by industrial activity should strengthen sector-specific efficiency benchmarks and technological retrofits; provinces with high industrial and electricity contributions should coordinate industrial load management with green electricity use and power-sector decarbonisation; and provinces with strong carbon emission contributions should implement joint energy–carbon monitoring and fossil fuel structure reviews. Regional and population signals further support zoned energy budgets and infrastructure planning. The framework therefore provides interpretable quantitative evidence for differentiated energy-saving and carbon-reduction policy under China’s low-carbon transition.