Novel dynamic thermal comfort prediction in buildings with single-shot vision-based deep learning method and low-cost thermographic imaging

• Published in: Building and environment Vol. 285; p. 113549
• Main Authors: Zhang, Wuxia, Calautit, John, Wu, Yupeng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2025
• Subjects: Built environment; Computer vision-based method; Deep learning model; Dynamic comfort model; Occupancy behaviour; Personal thermal comfort; Thermal imaging; You Look Only Once (YOLO); Deep learning model; Dynamic comfort model; You Look Only Once (YOLO); Occupancy behaviour; Thermal imaging; Built environment; Computer vision-based method; Personal thermal comfort
• ISSN: 0360-1323
• DOI: 10.1016/j.buildenv.2025.113549

•A novel comfort detection method using thermographic imaging.•A dataset of 8,257 thermal images and 340 TSV survey responses trains the model.•Thermographic imaging and TSV survey to build a personal thermal comfort model.•Automated feature extraction eliminates manual input and specialised equipment.•Intra-subject testing achieves high accuracy, demonstrating adaptability. Occupants’ thermal comfort is essential for individual well-being and energy efficiency in buildings. Widely used thermal comfort models, such as predicted mean vote (PMV), require input parameters that are often difficult to measure accurately without specialised equipment, and they overlook individual variability. Additionally, occupants’ comfort levels change over time due to dynamic human and environmental factors. A real-time thermal comfort prediction method is therefore crucial for building energy management systems and improving occupants’ thermal satisfaction. With advancements in computer vision, machine learning methods have been increasingly applied to thermal comfort prediction, particularly for individual comfort models. Unlike traditional machine learning approaches that require manual feature extraction, this study proposes a deep learning method that automatically extracts relevant information from thermal camera data, providing a more efficient, non-intrusive solution for real-time comfort prediction. A public dataset was established comprising 8527 thermal images and 340 thermal sensation votes from 22 participants. The model was evaluated through intra-subject and cross-subject experiments. Results showed that the Deep Learning Predicted Comfort (DLPC) method achieved an average MAE of 0.87 in cross-subject tests, outperforming the PMV model, which had an average MAE of 1.15 when both were evaluated against subjective thermal sensation votes. In intra-subject evaluations, DLPC achieved an even lower average MAE of 0.51, compared to the PMV model’s MAE of 0.93 (both relative to TSV). The method demonstrated its ability to predict thermal comfort without requiring manually defined features or explicit skin temperature measurements, reducing the reliance on specialised expertise. While still in the early stages, the approach demonstrates potential for enabling real-time building energy system control through low-cost, image-based comfort assessment, enhancing both occupant comfort and energy efficiency. [Display omitted]