From data to dignity: Understanding and predicting fuel poverty in the United Kingdom with machine learning

This study applies a data-driven framework to assess fuel poverty risk in England using the 2019 English Housing Survey. A clustering-prediction-interpretation approach combines k-prototypes clustering, XGBoost modeling, and Shapley Additive Explanations (SHAP) to identify household vulnerability profiles and the key drivers of fuel poverty. The study segments the entire dataset into three clusters: households with moderate energy costs, younger employed households with high appliance and lighting costs, and older low-income economically inactive households, who emerge as the most vulnerable group. Four XGBoost models—one for the entire dataset and three cluster-specific models—were developed to predict household fuel poverty. Predictive performance was high across models, and SHAP analysis revealed that the influence of socio-economic, housing, and energy-lighting cost features varies significantly across clusters. The results demonstrate that fuel poverty is heterogeneous and cannot be effectively addressed by uniform national schemes. Instead, differentiated interventions—such as targeted financial transfers for vulnerable households, retrofitting programs, and appliance-efficiency policies—are required. Methodologically, the study highlights the value of integrating clustering with predictive modeling to uncover hidden vulnerability patterns that conventional static indicators often overlook. While the analysis focuses on England and uses cross-sectional data, the proposed framework is adaptable to other contexts and can be extended with longitudinal data to strengthen generalizability. • Three distinct household clusters reveal heterogeneous fuel poverty risks in England. • Cluster 2 (ages 60–74, low income, single occupancy) faces the greatest vulnerability. • Housing type, dwelling age, and energy costs are strong predictors of fuel poverty. • Cluster-based XGBoost outperforms single models, improving prediction accuracy. • Findings call for targeted policy interventions tailored to specific household groups.