Progressive filtering and cluster-localized calibration for customized UBEM

Abstract

Urban building energy modeling (UBEM) has become increasingly important in recent years for understanding energy use in urban buildings. Traditional models often rely on a ‘one-size-fits-all’ approach based on a limited set of archetypes, which oversimplifies building diversity and undermines confidence in the results, while generating one archetype per house is computationally prohibitive. To overcome this limitation, this study introduces a “Three-Layer Filtering” approach with dual model that progressively reduces structural mismatch and enables scalable house-level refinement. It first constrains candidate archetypes using taxonomy-based classification, then derives representative archetypes via multi-parameter clustering and enables reusable assignment through a supervised classifier, and finally localizes each assigned archetype to individual houses using sensitivity-guided, surrogate-assisted Bayesian calibration under annual electricity EUI constraints. As a proof of concept, this work is demonstrated on 100 all-electric single-detached houses in Montreal using annual electricity EUI as the sole calibration target. The proposed approach progressively reduces building-level mean absolute percentage error from 80.7% to 23.3% after clustering and further to 4.3% after localized calibration. Aggregate bias is minimized to near zero, with a 95% posterior predictive interval of [−1.7%, 0.2%]. This proposed approach is methodologically general and can be replicated in similar data settings, providing a practical pathway to scalable, building-specific archetype refinement for urban energy analysis.