Integrating LDA clustering and autoencoder-based transfer learning for thermal comfort prediction

Understanding and predicting indoor thermal comfort is crucial for occupant well-being and energy-efficient building operation. However, personalized comfort models require high-quality data for extended periods of time, presenting significant challenges for practical implementation. This study presents an integrated approach that combines topic modeling and deep learning to enhance personal thermal sensation prediction. This study leverages the large ASHRAE Global Thermal Comfort Database II and a smaller experimental dataset to transfer knowledge from public datasets and improve personal thermal comfort model prediction. Latent Dirichlet Allocation (LDA) is used to cluster thermal sensation data into latent topics, identifying patterns in different individuals’ data ​from the experiment. An autoencoder is then trained on the public dataset’s features to extract deep representations, which are transferred to personal thermal sensation prediction models​. An ensemble classification model is built using the enriched feature set, and its performance is compared against baseline models which were based on experimental data. Results indicate that the enhanced ensemble models outperform baseline models in accuracy, precision, recall, and F1-score. This integrated methodology demonstrates the effectiveness of combining clustering and feature transfer from public datasets for personalized thermal comfort prediction, paving the way for more reliable and adaptive personal comfort models in building systems.