Data-Driven Building Retrofits: Predicting Energy Savings with AI for Sustainable Project Management

Achieving the European Union’s climate and energy targets requires large-scale improvements to the performance of existing residential buildings. Traditional methods for assessing retrofit strategies, although accurate, are often time-consuming, data-heavy, and unsuitable for early decision-making. This paper introduces XGreen, a predictive tool designed to estimate the outcomes of retrofit interventions in residential buildings in Lombardy, Italy, using only limited and readily available data from energy performance certificates (EPCs). The study investigates whether regional EPC records can be used to provide reliable predictions of two key outcomes: non-renewable energy performance (EPgl,nren) and the achievable post-retrofit energy class. The tool was developed using a curated dataset of more than two million entries from Lombardy’s CENED 2.0 + register. Fourteen models based on the XGBoost algorithm were trained: seven regressors to predict post-retrofit EPgl,nren, and seven classifiers to estimate the final energy class, using the regressor outputs as inputs. Results show that the regressor associated with transparent envelope upgrades achieved the best performance, with a mean absolute percentage error of 6.9%. All regressor errors were within the lowest 3% of their respective EPgl,nren distributions, indicating high reliability. Among the classifiers, the combined-intervention model achieved the strongest results, with 88.3% accuracy in predicting buildings upgraded to class A4. By providing rapid, data-driven feedback on different retrofit options without the need for full building simulations, XGreen supports faster, more cost-effective planning. The tool offers a scalable framework to guide policymakers, auditors, and owners in meeting EU renovation targets and advancing building decarbonisation.