Modeling and Personalization of Thermal Comfort Using Machine Learning and Transfer Learning Approaches

Abstract

Thermal comfort is a multifactorial phenomenon influenced by environmental, physiolog- ical, and contextual factors. As smart building environments evolve toward personalized control systems, data-driven models have emerged as key enablers of adaptive comfort prediction. However, existing models often struggle to generalize across users, time, and deployment contexts. This thesis investigates the problem of user preference transferability and dynamic com- fort modeling using real-world datasets enriched with multivariate sensor data. It em- phasizes user-centric modeling through feature engineering, temporal embeddings, and the use of sequential deep learning architectures. A particular focus is placed on explor- ing CNN-LSTM networks, which can jointly learn spatial representations and temporal dependencies inherent in thermal comfort signals. Additionally, the study explores model adaptation techniques—including instance reweight- ing and feature distribution alignment—that aim to improve generalization across users and climates. These approaches are evaluated with respect to their capacity to transfer learned comfort representations while minimizing degradation in predictive performance. In evaluating the predictive models, emphasis is placed on real-world deployment scenar- ios, including model simplicity, data sparsity, and on-device computation limits. These considerations help ensure that the insights derived from this research remain applicable to practical smart building implementations and personalized control systems.