Research And Application Of GNSS Tropospheric Delay Model

Under the background that GNSS promotes the development of space-time information acquisition technology,high-precision GNSS positioning has become the focus of future development.Tropospheric delay error has always been a stumbling block for GNSS positioning accuracy to be improved again.Model correction is the most frequently used solution at present,but there is still room for improvement in accuracy.After investigating the research contents,it is found that most of the current industry tends to build the empirical model of ZTD,while the traditional empirical model requires massive grid values,which requires facilities with better performance and increases the difficulty of using the model.Based on the gradual development of machine learning into the field of time series analysis,this paper introduces the LSTM neural network into the ZTD model to create a new model with higher precision and basically applicable to the world.Considering that ZTD is the key factor affecting PPP accuracy and convergence time,different ZTD correction models were substituted into PPP to explore practical engineering examples,and the effect of the model in actual positioning was verified.The main contents and research work of this paper are as follows:1.The relative theories of atmosphere are analyzed,and three traditional ZTD models are compared from the perspective of modeling principles.Three mapping functions,NMF,VMF1 and GMF,were loaded into the PPP solution respectively,and their effects on PPP were compared and analyzed.The experimental results show that NMF has the worst effect on PPP,GMF and VMF1 have better positioning effect than NMF.2.The existing ZTD model has the problems of unstable prediction effect,only applicable to some areas,and low precision level.Taking advantage of LSTM’s ability to capture long-term time-varying information effectively,a new model based on LSTM neural network was constructed by extracting the ZTD truth value provided by IGS center to train LSTM model.The model is designed for a single station and does not require meteorological parameters actually measured.The experimental results show that :（1）The average prediction accuracy of LSTM model is 8.99 mm among the eight globally distributed experimental stations,which is 35.0% and 20.7% higher than that of the existing BP model and GA-BP model,respectively.MAE and MAPE were lower than those of the two models.In general,the accuracy and stability of the LSTM model are significantly improved compared with the comparison model.（2）Among the eight stations,62.5% of the stations whose ZTD prediction accuracy of the LSTM model basically reached mm level.This indicates that the LSTM model is basically suitable for the prediction of ZTD in the global region.3.Most of the existing studies focus on how to build a new ZTD model,but neglect to verify its effect in actual positioning.The ZTD models based on LSTM,BP and GA-BP were loaded into the corresponding modules of PPP practical engineering calculation examples,and the corresponding station coordinates in IGS circular solution files were viewed as the true values.The influence of the new model on PPP effect was analyzed,which was mainly divided into two aspects: convergence time and position solution accuracy.The experimental data show that :（1）The average position resolution accuracy of the LSTM model at E,N and U is 1.68 cm,1.71 cm and 2.82 cm respectively,which is significantly higher than that of the latter two models.The average convergence time is 16.65 min,which is reduced by 27.8% and 21.1% compared with the other two models respectively in 6 stations.（2）PPP effect is directly proportional to the accuracy of ZTD model.