Multi-source domain adaptation for personalized thermal comfort prediction using wearable sensors

• A field study integrates wearable sensing and subjective thermal feedback with data analysis. • SMAT model is developed for personal thermal comfort prediction without labeled target data. • Class-conditional alignment with LMMD-based source selection reduces negative transfer. • SMAT shows superior accuracy and robustness compared to non-adaptive and single-source models. Thermal comfort is significant to human well-being and energy efficiency in buildings. However, personal thermal comfort prediction using data-driven models remains a challenge due to the heavy reliance on large volumes of labeled data for model training. To address this limitation, this study proposes a novel Selective Multi-source Adaptation for Thermal comfort (SMAT) model framework, which uses multi-source unsupervised domain adaptation to predict personal thermal comfort without requiring any subjective label from the target subject. A real-world field study was conducted with 16 subjects, collecting 1919 valid data samples comprising wearable sensor readings and environmental measurements. Comprehensive data analysis was performed to evaluate different indoor environments, exposure durations, clothing insulation levels, and thermal variables with subjective responses. SMAT was developed and evaluated based on the field experiment data, achieving the highest average prediction accuracy of 71.00 %, outperforming both non-adaptive model (CNN-LSTM, PMV etc.) and single-source domain adaptation model (DSAN). To further validate the generalization capability of the proposed approach, an additional climate chamber dataset involving 12 subjects was utilized. On this dataset, SMAT attained an accuracy of 88.71 %, confirming its robustness under controlled environmental conditions. These results demonstrate that SMAT not only reduces the need for subjective data collection, but also maintains high predictive performance across diverse individuals and settings, highlighting its potential for practical deployment in real-world settings.