Algorithm-Enhanced Fuel Poverty Prediction Using Household Characteristics, Socio-Economic Factors, and Clustering Analysis

This paper aims at predicting fuel poverty risk by applying clustering analysis to identify fuel- poor households. More specifically, it presents a novel approach to clustering based XGBoost modeling by integrating data from a representative household survey in England (N=11974). The unobserved heterogeneity in the fuel poverty households survey hides certain relationships between the contributory features in the survey and fuel poverty. We explore the application of the k-prototypes clustering method to group households into heterogeneous clusters. The study segments the entire dataset into three clusters. Four XGBoost models were developed on the entire dataset and on each cluster to predict household fuel poverty. The list of input features includes housing characteristics, socio-economic features, and energy cost variables. The results reveal significant variations in the influence of identified features on fuel poverty across clusters. For instance, Cluster 2, which primarily comprises economically vulnerable households aged 60-74 struggling with adequate despite moderate costs, highlight the necessity for targeted interventions. Furthermore, the Shapley Additive Explanations (SHAP) method is applied to analyze the impact of each feature on fuel poverty per cluster. The current study concludes that the clustering-based XGBoost model is a promising approach to identifying different groups of households at risk of fuel poverty and that policymakers can directly gain insights by determining the most relevant socio-economic features to tackle fuel poverty.