AI-driven thermal comfort management in public waiting halls: Passive sensing and adaptive HVAC control

Maintaining thermal comfort in large public waiting halls is challenging due to rapidly varying occupancy, spatial non-uniformity, and tight energy budgets. We propose an AI-enabled, occupant-centric control framework that passively estimates per-person features from synchronized RGB and thermal-infrared (TIR) streams. The framework first identifies occupancy-invariant, physically grounded features—metabolic rate, clothing insulation, action state (sit/walk) and intensity (walking speed)—and combines them with environmental drivers (air temperature, mean radiant temperature, air velocity) to inform thermal prediction. A lightweight vision pipeline (detection, tracking, optical-flow speed, and TIR torso temperature) extracts these features in real time without individual profiling. A sequence model then predicts short-horizon Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfied (PPD), and a comfort-focused regulator maps these forecasts to native HVAC setpoints (temperature and air speed) under device constraints. In a simulated waiting-hall scenario (10-40 occupants), the calibrated predictor attains PMV-TSV MAE , the controller keeps – of votes within the ASHRAE comfort band , and typically achieves within – min after density changes. These results show that coupling edge-efficient vision with sequential comfort prediction and actionable HVAC control can improve crowd-level comfort in complex public spaces while preserving privacy. • Occupant-centric. • YOLO detection. • Optical-flow motion estimation. • Sequential Mamba estimation.