Prioritizing building envelope retrofits through data-driven heat flux prediction

Existing buildings account for a major share of space-heating demand, yet retrofit actions are frequently selected arbitrarily due to limited diagnostics, leading to ineffective or economically unjustified interventions. A systematic, data-driven method for identifying the most thermally underperforming envelope components is therefore essential to support scalable retrofit decision-making. This study proposes a workflow that predicts surface-level heat flux and identifies the building-envelope components that should be retrofitted first. Four common Danish residential and office building models—were simulated under winter conditions to generate hourly indoor and outdoor temperatures, surface temperatures, solar radiation, heating energy use, and geometric descriptors. These features were processed within a unified pipeline and used to train two machine-learning models: an Extreme Gradient Boosting (XGBoost) regressor and a deep Multilayer Perceptron (MLP). XGBoost achieved superior accuracy (test RMSE = 11.24 W, MAE = 5.5 W, R 2 = 0 . 92 ), tighter residual distributions, and more stable behavior in the mid-to-high heat-flux regime compared to the MLP, which showed larger scatter and systematic bias during low and high-flux periods. When applied to an independent office building, the model correctly identified the highest-impact walls, and a sensitivity analysis — bringing each wall to current insulation standards — showed that the top-ranked surfaces produced annual heating-energy reductions of 6.49 GJ and 5.03 GJ, whereas the lowest-ranked surfaces yielded only 2.74 GJ. These findings demonstrate that predicted surface heat flux offers an effective, low-instrumentation metric for prioritizing envelope retrofits and that the proposed framework provides a scalable approach for supporting envelope retrofit decision making. • Data-driven surface-level heat-flux prediction using common building data. • Novel physically-informed features: Heat_Drive and A_frac_Q metrics. • XGBoost surrogate outperforms deep MLP across full flux range. • Validation shows top-ranked surfaces yield highest Heating energy savings. • Workflow needs minimal instrumentation, enabling scalable retrofit planning.