Transferability limits of PMV as a thermal-comfort benchmark across climatic and building-operation contexts in China

Indoor thermal-environment assessment underpins occupant comfort, HVAC control, building energy use, carbon performance, and design optimization, making the reliability of widely used comfort benchmarks an engineering issue beyond thermal sensation alone. Predicted mean vote (PMV) remains a standard physical benchmark, yet its correspondence with field-observed thermal sensation vote (TSV) may vary across climatic, building, and operational contexts. This study evaluated PMV as a benchmark-transfer problem using 15,760 field records from 29 Chinese cities. Cold, hot-summer/cold-winter, and severe-cold zones formed the confirmatory domain; Mild (211 records) was exploratory and hot-summer/warm-winter (14 records) coverage-only. Standardized PMV was recalculated for 8220 complete cases following ISO 7730 and assessed alongside physical and contextual models using repeated city-grouped nested validation, 2000 city-cluster bootstrap replicates, strict city-disjoint leave-one-zone-out transfer, sensitivity analyses, and non-China external validation. Standardized and database PMV showed strong agreement (Pearson r = 0.979; MAE = 0.120) but were not interchangeable. Canonical Macro-F1 was 0.469 for M1, 0.506 for M2, and 0.512 for M3; however, the paired M3–M1 difference was 0.043 and its 95% bootstrap interval crossed zero (−0.007 to 0.075). Strict transfer gains varied across climates (+0.012 to +0.099), while external rankings changed between complete and common-support samples. PMV–TSV residuals were most clearly directional in the Cold subset (mean −0.304; 95% CI −0.563 to −0.199). These findings show that contextual information can improve in-domain prediction but does not guarantee transportability. PMV should therefore remain the physically interpretable reference, with contextual correction used only when target-domain support and transfer performance are explicitly demonstrated.