Abstract: Recently, blind tropospheric zenith delay (ZTD) models have been developed rapidly since they don't require any measure surface meteorological data, giving GNSS users great convenience. Neural network technology for ZTD models based on sited measured data has made some achievements. Meanwhile, some neural network models for blind ZTD models have been built, but they have some drawbacks:it ignore the ZTD variation with the time and can only forecast ZTD. In view of these shortcomings, this paper constructs an optimization of neural network model of a blind ZTD model. Results show that the proposed neural network models can forecast both ZHD and ZWD respectively and are with good accuracy:for ZHD, BIAS and RMSE are 2.5 mm and 20.6 mm respectively; for ZWD BIAS and RMSE are 2.4 mm and 35.7 mm respectively. In this paper, the ZHD and ZWD precision of the neural network models are also with the world famous blind model- the GPT2w. In addition, compared with the GPT2w, the neural network models in this study have the biggest advantages of usage without large grid data as reserved data but just need to know when to use a trained neural network, whose characteristics provide GNSS users with great convenience.

Key words: ZTD, neural network model, GPT2w model, GNSS