Adaptive thermal comfort (ATC) models can be used to achieve energy savings by dynamically specifying the room set point temperature. The accurate ATC models can assist in indoor environmental control and building energy-saving design. Existing models mainly include adaptive heat balance and adaptive regression models. Due to the complicated PMV index of adaptive heat balance in practice often predicting worse than simple indices (temperatures), the adaptive regression models (such as ASHRAE 55 and EN 16798 models) are more widely used. This study focuses on the regression model, which is a regression equation with outdoor temperatures as the independent variable and neutral temperatures as the dependent variable. The neutral temperature can be calculated using the Griffiths method. The operative temperature is the representative value of the dependent variable of the model, and there are many kinds of representative values of the independent variable (outdoor temperatures). According to the outdoor test data, different processing methods (such as natural daily mean, monthly mean, 7-day running mean, 15-day running mean, 21-day running mean, and so on) result in different representative values of outdoor temperatures. The representative values of the independent variable of the existing models are not uniform, and the models built with different representative values vary greatly. How to determine the appropriate representative values to establish the optimal model? Little research has been done on the above question. There are two traditional methods for establishing ATC models. The first method is that the representative value of the independent variable is directly adopted from existing studies, which has the drawback of not verifying the applicability and effectiveness of the representative value. The second method is to select a reasonable representative value based on the maximum coefficient of determination (R2) to determine the optimal ATC model. The main defect of this method is that it does not consider the estimation standard error (Se) of the equation. The purpose of establishing ATC models is to evaluate and predict. Determining the optimal model should consider not only the degree of prediction (R2) but also the prediction error (Se). Few studies have applied the Se index to determine the optimal ATC model. In this paper, R2 and Se are used as evaluation indexes to determine the optimal ATC model. Since the evaluation indexes are not unique, the optimal solution cannot be obtained using a simple comparison. Multi-objective optimization analysis is required to determine the optimal ATC model. The purpose of this study is as follows. A method to optimize the ATC model is proposed, and a yearround thermal comfort survey of seven dormitory buildings in Bengbu, China, is used as an example to implement and validate the proposed method. The research methods and process are as follows. Firstly, the thermal comfort survey was carried out using simultaneous instrumental tests and questionnaires. The buildings were all operated in natural ventilation mode during the survey. The survey was conducted over one year, in which indoor and outdoor thermal parameters were measured. The outdoor parameters included air temperature and relative humidity, and the testing instrument was placed in a louver box 1.5 meters above the outdoor ground. The questionnaire mainly consists of two parts: objective information and subjective feelings. Then, the literature was consulted to summarize all the representative values of outdoor temperature, and ATC models were built for eachrepresentative value of outdoor temperature. A total of 17 kinds of models were obtained. In addition, classification ATC models were also developed separately by gender and season, as ATC models are affected by gender and season. Afterward, a multi-objective optimization method (grey relational analysis) was used to evaluate all ATC models and ultimately determine the optimal ATC model. Finally, the different models were compared. The optimal annual ATC model obtained from the study was compared with the most internationally recognized ATC models (ASHRAE 55, EN 16798, and GB/T50785) to verify the accuracy of the optimal model in this study. The classification ATC models were compared with the annual model to verify the accuracy of the classification model. The following conclusions are drawn. The optimal ATC model of this study can explain 69.6% of the information in the measured data. In comparison with existing international ATC models, it is found that the optimal ATC model determined in this study has the best prediction accuracy and stability, indirectly confirming the effectiveness of the method proposed in the paper. The difference in predictive performance between the classification ATC model and the annual model is significant. The classification model has better predictive performance than the annual model. Compared with the annual model, the prediction accuracy of ATC models established by seasons improved by 7.7% (winter), 13.6% (summer), and 8.5% (transition season), respectively. This study enriches the adaptive thermal comfort theory and provides a valuable reference for accurately establishing ATC models.