Determinants of energy poverty among Chinese households: Risk prediction model using machine learning algorithms

Energy poverty poses a significant obstacle to sustainable development worldwide, especially in the context of China’s rapid urbanization and economic transformation. This study aims to develop a robust predictive model to identify the key determinants of energy poverty among Chinese households. Utilizing multi-source data from the China Family Panel Studies (CFPS) spanning 2012 to 2022 and encompassing 18,678 households, combined with prefecture-level socioeconomic indicators and nighttime light remote sensing data, we applied advanced machine learning algorithms to analyze energy poverty risk factors. Five-fold cross-validation was used to assess the performance of various classifiers including LightGBM, Logistic Regression, Support Vector Machine, Random Forest and XGBoost under different resampling techniques such as SMOTE, undersampling, and the combination of SMOTE with Tomek Links. Model interpretability was enhanced through SHAP (SHapley Additive exPlanations) analysis. Subgroup analyses examined urban-rural disparities, while sensitivity analyses confirmed the robustness of our findings across alternative definitions of energy poverty. The results show the following. First, the Random Forest model outperformed other algorithms, achieving an average validation AUC of 0.9021. Sampling methods such as SMOTE and its combination with Tomek Links significantly improved the prediction of the minority class affected by energy poverty. Incorporating prefecture-level socioeconomic indicators and remote sensing data further increased model accuracy. Second, SHAP analysis identified household income, internet usage, and educational attainment as the primary predictors of energy poverty. Additionally, households engaged in agriculture within urban areas were more vulnerable to energy poverty. Third, sensitivity analyses demonstrated that the model’s performance remained robust under different definitions and calculations of energy poverty. This study highlights the effectiveness of machine learning techniques in diagnosing energy poverty risks and informs targeted policies such as expanding rural internet infrastructure, providing tailored urban energy subsidies for agricultural households, and investing in education and vocational training to reduce urban-rural disparities in China.