Ineffective thermal comfort prediction can lead to challenges, including discomfort for building occupants and substantial energy waste caused by unnecessary excessive cooling or heating in buildings. The Predicted Mean Vote (PMV) model has been widely applied for thermal comfort management in air-conditioned buildings due to its theoretical validity and practical applicability. Among the critical inputs required for PMV calculations, the metabolic rate of occupants is identified as the most crucial parameter, as it directly influences the heat balance model that forms the foundation of the framework. However, existing methods for measuring metabolic rate encounter significant obstacles, such as operational difficulties in real-world situations or technical inaccuracies arising from oversimplified assumptions. This chapter presents an innovative approach to enhance the PMV model’s precision in thermal sensation prediction by indirectly estimating the metabolic rate. Specifically, the metabolic rate is formulated as a linear function of two key environmental variables: room air temperature and air velocity, with the model explicitly considering the effects of physiological adaptation mechanisms. To optimize the model parameters, a variable metric algorithm is used as an optimizer, which systematically minimizes the difference between the PMV predictions and the subjective thermal sensation votes collected from human subjects. The proposed method is validated through a series of experimental setups, including controlled tests in environmental chambers designed to simulate a stratum-ventilated classroom and an aircraft cabin, as well as field experiments conducted in an actual air-conditioned building using data from the ASHRAE database. The results demonstrate notable improvements: the method enhances the accuracy and robustness of PMV in thermal sensation prediction by over 52.5% and 41.5%, respectively. Fundamentally, this approach develops a grey-box modeling framework through model calibration, which outperforms traditional black-box models based on machine learning algorithms in terms of interpretability and generalizability.