Abstract: In view of the low accuracy of the Bevis model in Hong Kong, the measured data of the radiosonde station in Hong Kong for 20 consecutive years (2000—2019) is used to establish the Bevis refined model of the T_m in Hong Kong considering multi-factors. Firstly, the linear correlation between T_m and surface temperature (T_s), water vapor pressure (e_s) and pressure (P) are analyzed. Secondly, the annual periodic variation of T_m is detected by fast fourier transform. Finally, based on the Bevis model, comprehensively consider the changes of T_s, e_s, P and annual cycle, the Bevis refined model of weighted average temperature (BTm) in Hong Kong is established based on the nonlinear least square method. Considering the values calculated by the numerical integration method of radiosonde data in Hong Kong in 2020 as the reference values, the accuracy of the model is tested and compared with the Bevis global model and GPT2w model. The results show that the BTm model has relatively good accuracy, and its mean bias and root mean square error (RMS) are -0.11 K and 1.64 K, respectively. Compared with the Bevis global model, the GPT2w-1 model and the GPT2w-5 model, the accuracy is improved by 46.4%, 35.7% and 37.2%, respectively.