An explainable AI framework for prioritising EPC D-G retrofits across England and Wales through emission-driven dwelling typologies and spatial delivery clusters

Meeting the UK’s target to upgrade all homes to Energy Performance Certificate (EPC) band C by 2035 requires urgent action on poor-performing properties, especially those currently rated D to G. These homes contribute disproportionately to residential carbon emissions, yet retrofit strategies often overlook their typological and spatial diversity. This study examines how retrofit needs vary across EPC band D-G dwellings in England and Wales, drawing on 9.7 million records from the national EPC database. A structured method to processing high-cardinality categorical variables was developed, addressing a critical barrier in EPC-based analysis and enhancing both model interpretability and robust feature representation. Carbon emissions were modelled using XGBoost (R2 = 0.82) and interpreted with explainable artificial intelligence (XAI) to identify key emission drivers. Novel SHAP-informed clustering revealed four retrofit typologies, demonstrating improved cluster coherence compared with existing EPC-based methods. Two high-priority typologies emerged: large, exposed and uninsulated homes needing deep fabric-first upgrades, and partially insulated homes with gas boilers suitable for heat transition. Spatial clustering of local authorities identified five delivery environments, characterised by geographically differentiated heat demand profiles, tenure constraints, and delivery scale. The proposed framework improves the transparency and policy relevance of EPC-based modelling, offering actionable insights for locally tailored retrofit planning.