Reducing the Sensing Burden: A Sensor-Light Machine Learning Framework for Thermal Comfort Assessment

Thermal comfort shapes occupant health, well-being, and productivity and influences the sustainability and energy efficiency of the built environment. The Predicted Mean Vote (PMV) is the most widely used thermal comfort index, yet four of its six input parameters—globe temperature, clothing insulation, metabolic rate, and air velocity—require specialized, costly equipment or occupant self-reporting, which has long limited its practical large-scale application. This study introduces a machine learning (ML) framework aimed at estimating these four parameters using indoor and outdoor air temperature and relative humidity as its only sensor inputs, complemented by readily available contextual information and individual activity profiles. It exploits the climatic coupling of globe temperature and air velocity to the indoor–outdoor environment and the temperature- and activity-driven behavioral patterns that govern clothing insulation and metabolic rate. The proposed framework achieved strong predictive performance, explaining 85% of the variance in actual PMV values (R2 = 0.85), with a near-zero mean residual (−0.041) and a residual standard deviation of 0.286. Approximately 91% of absolute errors fell below 0.5 PMV units—a deviation unlikely to shift the assigned thermal comfort category. Mapped to thermal comfort categories, predictions reached 80% accuracy, with a macro-averaged precision of 0.81 and recall of 0.80, exhibiting the highest performance for neutral and warm conditions while performing less accurately for cool discomfort. These results suggest that standard temperature and humidity sensors, combined with basic contextual information and individual activity profiles, could support reliable PMV-based thermal comfort assessment, advancing scalable, sensor-light comfort monitoring for energy-efficient, sustainable buildings.