Thermal performance of post-disaster housing and its impact on occupant comfort: An integrated ML-ABM approach

Abstract

This study addresses the deterioration of thermal performance in post-disaster rural housing envelopes in hot-humid climates. By integrating a synergistic framework of Random Forest (RF), Shapley Additive Explanations (SHAP) interpretability analysis, and Agent-Based Modeling (ABM), we reveal the dynamic mechanisms by which architectural features regulate cooling dynamics via thermal sensation mediation. Empirical data from 231 post-disaster reconstructed dwellings in Mianzhu, Sichuan Province, China, demonstrate that roof insulation (28.3 % SHAP contribution), Window-to-Wall Ratio (WWR) (25.9 %), and exterior wall insulation (24.7 %) are core drivers of TSV. High WWR (>40 %) combined with single-pane glazing (ϕ=+0.23) significantly increases thermal discomfort risks, while double-glazed insulation (U ≤ 2.8 W/m²K) with moderate WWR (35 %-45 %) effectively mitigates overheating. ABM simulations reveal non-insulated buildings exhibit a temperature rise rate of 1.51 °C/h, triggering air conditioning (AC) activation 0.92 h earlier than for optimized structures. Envelope optimization reduces daily AC usage from 5.75 h to 3.2 h (-44.3 %) while increasing fan utilization by 6.6 %. The proposed Machine Learning-Agent-Based Modeling(ML-ABM) framework is validated and transferable to diverse building typologies (e.g., urban residences, offices) in hot-humid climates. The findings identify dual energy-saving pathways; these being passive thermal attenuation via inertial modulation, and behavior-guided active cooling, both validating the efficacy of "daylighting-insulation-ventilation" co-design. Further, findings provide empirical evidence of the benefits of transitioning from single-parameter optimization to dynamic system control in hot-humid climate retrofits.