Transfer learning with unsupervised domain adaptation for personal thermal comfort prediction

Abstract

Maintaining comfortable indoor environments is essential for occupant well-being and energy efficiency. However, predicting personalized thermal comfort remains challenging owing to the limited availability of labeled data. This study presents a novel approach to personalized thermal comfort prediction by applying unsupervised domain adaptation techniques. We leverage ASHRAE Global Thermal Comfort Database II as the source domain and the BCA testbed dataset as the target domain to address the challenge of adapting the models to new subjects with no labeled data. Specifically, we compare three domain adaptation methods: Correlation alignment (CORAL), Deep subdomain adaptation network (DSAN), and Dynamic adversarial adaptation network (DAAN). Our results demonstrate that applying deep domain adaptation (DSAN and DAAN) significantly enhances the prediction accuracy for both thermal preference and thermal acceptability tasks, with improvements of up to 12—15% compared with the base (CNN-LSTM) model. Furthermore, our proposed method maintains a high prediction performance, even when the available training dataset is substantially reduced. The maximum reduction in accuracy is at most 2.28% with approximately 10% of the unlabeled training dataset, demonstrating its robustness in real-world scenarios with limited environmental features. This study underscores the potential of unsupervised deep domain adaptation for improving personalized thermal comfort predictions and contributes to more efficient indoor climate control and a comfortable environment.