Cross-Regional Methodological Validation of Ensemble and Kernel-Based Machine Learning Models for Automated Building Energy Performance Assessment

Résumé fourni par la source

This paper presents an extended machine learning framework for predicting building Energy Performance Certificate (EPC) scores and evaluates its consistency across regulatory and geographic contexts. Building on two prior single-context studies, the present work integrates four regression paradigms — SVR, Random Forest, XGBoost, and, AdaBoost — trained on the Seattle Building Energy Benchmarking dataset (34,709 records, 46 parameters) and evaluated against real municipal EPC and building-registry data collected from three Turkish administrative districts: Torbalı, Beşiktaş, and Ümraniye. A structured preprocessing pipeline addressing missing data, outlier treatment, correlation-driven feature reduction, and categorical encoding was applied consistently across all data sources. Among the four models, the tree-based ensembles outperformed both the kernel-based SVR and the boosting-based AdaBoost on both datasets, with XGBoost achieving the strongest fit on the Seattle benchmark (R² = 0.801) and again the strongest fit on the Torbalı municipal sample (R² = 0.824). Exploratory analysis of the Turkish municipal data further identifies insulation status, construction era, and heating-fuel type as the dominant levers for emissions reduction, with uninsulated buildings consuming approximately 71% more site energy than fully insulated stock. Unlike earlier single-dataset studies, this work removes optical character recognition (OCR) document ingestion from the scope and instead concentrates on a rigorous, side-by-side comparison of four learning paradigms across independent regional datasets, offering evidence for the predictive ceiling of ensemble learning on tabular energy data and for the conditions under which a smaller, locally collected dataset can achieve results comparable to those from a larger but less homogeneous metropolitan dataset.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.