A hybrid active learning framework for personal thermal comfort models

Personal thermal comfort models are used to predict individual-level thermal comfort responses to inform design and control decisions of buildings to achieve optimal conditioning for improved comfort and energy efficiency. However, the development of data-driven thermal comfort models requires collecting a large amount of sensor-related measurements and user-labelled data (i.e., user feedback) to achieve accurate predictions, which can be highly intrusive and labour intensive in real-world applications. In this work, we propose a hybrid active learning framework to reduce data collection costs for developing data-efficient and robust personal comfort models that predict users' thermal comfort and air movement preferences. Through the proposed framework, we evaluated the performance of two active learning algorithms (i.e., Uncertainty Sampling and Query-by-Committee) and two labelling strategies (Independent and Joint Labelling strategies) to achieve the optimal reduction in user labelling effort for personal comfort modelling. The effectiveness of the proposed framework was demonstrated on a real-world thermal comfort dataset involving 58 participants collected over 10 working days with 2,727 responses under 16 thermal conditions. The final results showed a 46% and 35% reduction in labelling effort for the thermal comfort and air movement preference models, respectively, with increasing reductions occurring over time and when encountering new users. Through the insights gained in this study, future studies on data-driven thermal comfort models can adopt active learning as a viable and effective solution to address the high cost of data collection while maintaining the model's scalability and predictive performance.

Automated summary

In this study, a hybrid active learning framework was proposed to reduce data collection costs for developing efficient and robust personal comfort models. Two active learning algorithms and two labelling strategies were evaluated to achieve optimal reduction in user labelling effort. The results showed significant reduction in labelling effort for thermal comfort and air movement preference models, with increasing reductions over time and with new users. This study highlights the potential of active learning as an effective solution for the high cost of data collection in data-driven thermal comfort modelling.