Abstract:Firstly, we evaluate the accuracy of the meteorological parameters estimated by the GPT3 model using the data of 49 radiosonde stations adjacent to GNSS stations in China from 2017 to 2018. Secondly, combining the meteorological parameters estimated by the GPT3 model and 49 GNSS stations, we calculate the daily mean PWV, and evaluate the accuracy by radiosonde stations adjacent to GNSS stations. Finally, the results are obtained through the experiment: 1) In China, the accuracy and stability of the GPT3 model with 1° resolution are better than those with 5° resolution. The air pressure, temperature and Tm annual bias are 0.73 hPa, 1.34 K and -1.67 K, and the annual RMSE are 4.21 hPa, 3.75 K and 4.15 K. 2) The accuracy of PWV based on temperature inversion by GPT3 model combined with Bevis empirical formula is similar to that by GPT3 model Tm, and the PWV obtained by the two methods and PWV obtained by the sounding data show good consistency; furthermore, the accuracy of PWV in Tibet Plateau and northwest China is better than that in the southern and northern regions.