Toward smarter HVAC control: Machine learning reveals hidden drivers in thermal comfort databases

Interpretable thermal-comfort modeling requires more than accuracy: it requires clarity about which inputs matter and how missing data policies reshape those conclusions. We present a quality-aware sensitivity framework that couples cross-dataset quality labeling with perturbation analysis to quantify how model predictions of thermal sensation respond to changes in inputs. We transfer quality labels from the China Thermal Comfort dataset to previously unlabeled ASHRAE records, then perturb inputs with ± 10% multiplicative noise (numericals) or random category swaps (categoricals) and summarize effects as the absolute mean shift in predicted thermal sensation per feature. Crucially, we predeclare and compare three missingness policies: (A) rule-based imputation (mean radiant temperature [MRT] → air temperature [Ta] where missing), (B) NaN-native learning (keep null MRT), and (C) measured-only (drop null MRT). The Ta:MRT sensitivity ratio (air temperature leverage relative to mean radiant temperature) shifts markedly across policies: 2.533 [2.531, 2.536] under (A), 4.462 [4.458, 4.467] under (B), and 1.895 [1.893, 1.897] under (C), with lower error but conservative radiant underestimation for (A). Anthropometric and contextual variables remain influential, and simple-perturbation and Monte-Carlo rankings strongly agree (e.g., numerical Kendall’s τ b ≈ 0.82; Jaccard@10 ≈ 0.82). The framework yields policy-aware, uncertainty-quantified sensitivity that supports lean sensing (prioritize MRT where feasible) and control design (report policy + Ta:MRT + CIs alongside accuracy) for energy-efficient, occupant-centric HVAC.