Abstract Aiming at the obvious periodic errors of the meteorological parameters derived from the GPT3 model, we use the seven radiosonde stations in the Yangtze river delta region during 2015 to 2019 as references to analyze the seasonal periodic changes of meteorological residuals of the GPT3 model. We use the Emardson-H model to establish a new improved GPT3 model. The results show that the root mean square (RMS) of pressure, temperature, vapor pressure and weighted average temperature (Tm) of the GPT3 model are 5.09 hPa, 3.90 K, 4.01 hPa and 4.54 K respectively compared to the reference values of radiosondes. Their mean RMS of the improved GPT3 model based on Emardson-H are 4.64 hPa, 3.53 K, 3.73 hPa and 3.27 K respectively, improved by 0.45 hPa, 0.37 K, 0.28 hPa and 1.27 K in comparison to the GPT3 model. Therefore, the precision of the improved GPT3 model based on Emardson-H is better than that of the GPT3 model.