Abstract:From the observation data of 96 sounding stations in China, we use 1 887 313 sets of data from 2016 to 2018 as the training set, and 635 337 sets of data in 2019 as the test set. We conduct the comparison test with the exponential, ITU-R exponential, dual-exponential and Hopfield models. The results show that the BP model, which takes the surface meteorological information and the spatial location of the fixed point as input features, achieves the best effect by fully taking into account the various possible influencing factors of the refractivity. Compared with the exponential, ITU-R exponential, double exponential and Hopfield models, the RMSE of the best BP model is decreased by 69.8%, 33.1%, 31.9% and 16.8%, respectively. The BP model not only outperforms the traditional models in overall accuracy, but also has a more uniform distribution of errors in geospatial and longitudinal profiles.