Physics-informed machine learning for ASHRAE-compliant heat transfer coefficient estimation from short-term IoT monitoring

• ASHRAE-compliant HTC estimation from 7-28 day occupied monitoring achieved. • Physics-informed ML: 53% better than data-only, 45% better than physics-only. • Validated on 28 holdout buildings: 92.9% ASHRAE compliance, R 2 = 0.967, CV-RMSE = 9.8% • Calibrated uncertainty quantification enables risk-stratified deployment. Accurate Heat Transfer Coefficient (HTC) measurement is essential for building retrofit prioritization, yet traditional co-heating tests require 2–4 weeks of vacancy at £3,000–5,000 per building—economically prohibitive for large-scale deployment. This study develops a physics-informed machine learning framework achieving ASHRAE Guideline 14-compliant HTC estimation from 7-28 day occupied monitoring data. The framework systematically integrates SAP 10.2 building physics principles with IoT sensor streams through physics-enhanced feature engineering. A Gradient Boosting model trained on 141 UK dwellings (140 SMETER dataset, 1 Energy House 2.0) achieved CV-RMSE of 9.8%, R 2 of 0.967, and NMBE of −0.6% on holdout testing, with 92.9% of buildings meeting ASHRAE compliance thresholds (CV-RMSE < 15%). Physics-informed features contributed 41% of total model importance, delivering 53% performance improvement over raw monitoring data alone and 69.5% over physics-based calculation alone. Bootstrap conformal prediction provided calibrated uncertainty quantification with 85.7% empirical coverage. Energy House 2.0 validation demonstrated 1.5% prediction error under controlled conditions. Systematic performance boundaries were identified: buildings >100 m 2 with >1,500 kWh annual consumption achieve <8% error, while buildings <50 m 2 with <500 kWh require extended monitoring periods. Leveraging existing IoT infrastructure, this framework enables national-scale building stock characterization supporting evidence-based retrofit prioritization without additional measurement costs