Machine learning based thermal response study of university students under summer outdoor military training

Visual perception significantly influences human thermal comfort. However, few studies have examined its role in predicting thermal comfort, especially among university students undergoing outdoor military training. This study investigates the outdoor thermal comfort of such students and evaluates the performance of different algorithms and features in predicting their thermal comfort. A questionnaire survey was conducted to investigate the thermal and visual perceptions of freshmen at Tianjin Chengjian University during their military training, yielding 1,754 valid responses. Microclimatic conditions were monitored continuously throughout the study. The collected data were then used to develop thermal comfort prediction models based on eight machine learning algorithms. The results demonstrate that models utilizing the visual dataset achieved superior predictive performance. Notably, the thermal acceptability model attained the highest accuracy of 68.2%. Compared to a model using only PET as an input feature with Logistic Regression, this represents a maximum accuracy increase of 34.1%. Furthermore, Bayesian Optimization enhanced the accuracy of the thermal sensation, acceptability, and comfort models by 2.2%, 0.7%, and 2.4%, respectively. Among all features, the visual comfort vote was identified as the most significant in predicting thermal comfort. This study concludes that visual comfort is a critical factor in predicting the thermal comfort of students during military training. The findings provide a scientific basis for the design and planning of outdoor open spaces.