Machine Learning with Administrative Data for Energy Poverty Identification in the UK

Energy poverty continues to be a critical challenge, and this requires efficient and scalable identification methods to support targeted interventions. The Low Income Low Energy Efficiency (LILEE) indicator and previously the Low Income High Costs (LIHC) indicator have been used by the UK government to monitor national energy poverty levels. Yet due to their reliance on complex, time-intensive data collection processes and estimations, these indicators are not suitable for identifying energy poverty in specific households. This study investigates an alternative approach to energy poverty identification: using machine learning models trained on administrative data, data that could reasonably be available to governments for all or most households. We develop machine learning models using data from the English Housing Survey that serves as a proxy for administrative data. This data is selected to closely resemble what might be available in national administrative databases, incorporating variables such as household socio-demographics and building physical characteristics. We evaluate multiple classification algorithms, including Random Forest and XGBoosting, applying resampling and class weighting techniques to address the inherent class imbalance in energy poverty classification. We compare model performance with a ‘benchmark’ model developed by the UK government for the same goal. Model performance is assessed using the metrics of accuracy, balanced accuracy, precision, recall, and F1-score, with SHapley Additive exPlanations (SHAP) values providing the interpretability of the predictions. The best-performing model (XGBoosting with class weighting) achieves higher balanced accuracy (0.88), and precision (0.51) compared to the benchmark model (balanced accuracy: 0.77, precision: 0.24), demonstrating an improved ability to classify energy-poor households with fewer data constraints. SHAP analysis reveals household income and dwelling characteristics are key determinants of energy poverty. This research demonstrates that machine learning, trained on existing administrative datasets, offers a feasible, scalable, and interpretable alternative for energy poverty identification, enabling new opportunities for efficient targeted policy interventions. This study also aligns with recent UK government discussions on the potential for integrating administrative data sources to enhance policy implementation. Future research could explore the integration of real-time smart meter data to refine energy poverty assessments further.