Abstract:Using multi-layer perceptron(MLP) on sounding data from 2015 to 2017 in southwest China as the experimental data, we establish the weighted mean temperature (Tm) model for the area. We use meteorological parameters (surface temperature, water vapor pressure) and non-meteorological parameters (elevation, latitude, and day of year) as model input factors, and use the Tm calculated by the numerical integration method as the learning target. We apply the neural network model for iterative training to obtain the Tm in southwest China. Using the Tm data of the sounding stations in 2018 as reference values, we verify the accuracy of the MLP model and compare it with the Bevis and GPT3 models. The results show that the mean annual RMSE and annual bias of the MLP model are 1.99 K and 0.15 K, respectively. Compared with the Bevis model and the GPT3 model, the mean annual RMSE is reduced by 1.36 K (40.6%) and 1.51 K (43.1%), respectively. The bias drops by 0.70 K(82.4%) and 1.04 K(87.4%), respectively, and the accuracy and stability of the model in different elevations, latitudes and seasons in southwest China are better than the Bevis and GPT3 models.