An explainable deep learning model for energy performance classification and retrofitting recommendations

• Achieved 99.98% test accuracy in classifying building energy efficiency. • Introduced a deep learning model with L2 regularisation and dropout layers. • Balanced EPC dataset using SMOTE for improved fairness and generalisation. • Applied SHAP for explainability and feature importance analysis. • Provided counterfactuals for personalised retrofit recommendations. Energy-efficient building retrofitting is crucial in reducing carbon emissions and enhancing sustainability. This study presents a novel Deep Learning-based Explainable AI model for energy efficiency classification and retrofit recommendation. Our model integrates a neural network with L2 regularisation, dropout layers, learning rate scheduling, and the Synthetic Minority Over-sampling Technique for class balancing, ensuring robust generalisation. The model is trained on an extensive dataset of buildings from the EPC Dataset − Region Lombardy, Italy, classifying structures into energy-efficient (A4) and non-energy-efficient (D-G) classes. The proposed model achieved a test accuracy of 99.95%, surpassing conventional machine learning and hybrid AI approaches in the literature. Additionally, it provides more accuracy by incorporating SHAP-based explainability to interpret model decisions and identify the key factors influencing energy efficiency. Counterfactual explanations provide personalised retrofit recommendations, focusing on insulation, renewable energy adoption, and efficient lighting solutions. The insights from this study provide a transparent, interpretable AI model that supports decision-makers, policymakers, and stakeholders in optimising retrofitting strategies for sustainable urban development.