ML-Driven Sensitivity Analysis for Lean HVAC: New Insights from Large-Scale Comfort Data

Hongshan Guo, Yu Chang, Difeng HU

Keywords: Thermal comfort modeling, global sensitivity analysis, Lean sensing for HVAC, Variance-based feature attribution

TL;DR: We use LightGBM and global sensitivity analysis on 148k comfort records to show that mean radiant temperature is the top driver of thermal sensation, enabling leaner, more explainable HVAC control strategies.

Abstract:

Identifying the physical and contextual drivers
of occupants’ thermal sensation is essential for
lean sensing and explainable HVAC control.
We merge and harmonise 148,148 steady-state
records from the ASHRAE Global Thermal Com-
fort Database v5 and the China Thermal Comfort
Dataset, then train a LightGBM regressor selected
via PyCaret in a no-imputation workflow that ex-
ploits the model’s native NaN handling. Five-fold
cross-validation yields an RMSE of 0.67 TSV
units. Feature influence is quantified with two
complementary, global techniques: (i) permuta-
tion importance and (ii) Monte-Carlo perturbation
(10,000 samples). Both agree that anthropomet-
ric variables dominate (height ≈0.048, weight
≈0.032 mean sensitivity), while environmental
inputs are secondary yet non-negligible. Notably,
the mean radiant temperature (MRT) and air tem-
perature (Ta) show comparable leverage, with an
effective sensitivity ratio of MRT : Ta ≈1.5 : 1.
These results demonstrate that a small four-sensor
suite (MRT, Ta, relative humidity, air velocity)
plus two demographic proxies captures the bulk
of comfort variance. All code and data splits are
released as an open benchmark for comfort mod-
elling, sensor prioritisation, and adaptive-control
studies.

Submission Number: 5