Analysis of Influencing Factors and Prediction of Provincial Energy Poverty in China Based on Explainable Deep Learning

Energy poverty remains an important challenge for sustainable development in China, with pronounced regional disparities and evolving temporal dynamics that require accurate and interpretable prediction tools. This study develops a provincial panel-based framework that combines Energy Poverty Index (EPI) construction, SSA-LSTM prediction, SHAP-based model interpretation, and two-way fixed effects (TWFE) regression analysis. Using provincial data for China (2003–2022), we first construct a composite EPI with the entropy weight method, then apply a Sparrow Search Algorithm (SSA) to optimize LSTM hyperparameters for EPI forecasting. SHAP is used to interpret feature contributions to model-predicted EPI, and TWFE regression is used to provide complementary panel-data evidence on factor–EPI associations. The results show that the SSA-LSTM model outperforms benchmark machine learning and deep learning models in out-of-sample prediction performance. SHAP-based interpretation indicates that variables such as GDP, energy intensity, and power generation per capita contribute strongly to prediction variation, with notable regional heterogeneity. TWFE results are broadly consistent with several key patterns identified in the SHAP analysis. Overall, the proposed framework provides an accurate and interpretable provincial energy poverty prediction approach and offers a useful empirical reference for energy poverty monitoring and policy discussion.