Regional TEC Model using Improved Neural Network and Its Application in Single Frequency Precise Point Positioning

摘要

An improved neural network has been developed for estimating the total electron content (TEC) of the ionosphere using observations collected by a GPS network of double frequency GPS receivers. And a single frequency precise point positioning (PPP) software MPEPPP has been developed using neural network to mitigate the ionospheric delay. Ionospheric delay is the main error source of single frequency PPP. It shows large variations which are correlated with the solar activity, geo-magnetic influences etc. Since modeling the ionospheric delay is complicated, while one of the advantages of neural network is that it can fit any complicated surface very well, an improved two-layer back-propagation neural network has been designed to model the ionosphere delay in this paper, using Levenberg-Marquardt training method due to its rapid convergence properties and robustness. TEC values are extracted using polynomial model from the carrier phase observations of base stations. Then the epoch time, geomagnetic latitude and sun-fixed longitude of ionospheric pierce point, zenith distance and the station-satellite distance are set as the input neurons, and the corresponding TEC is the output neuron. The new model is tested and compared with the polynomial model using one week GPS observations of the Chinese Jiangsu province continuously operating reference Stations (JSCORS) covering more than 100,000 km{sup}2, and the distances of the selected base stations are all over 200 kilometers. The authors designed the single frequency PPP software MPEPPP, using the new ionospheric delay model and IGS precise orbit and clock products. Error sources such as troposphere delay, GPS antenna phase center offsets and variations, GPS satellite phase wind-up, earth rotation, earth tide and ocean tide loading etc, are also considered in the software. Numerical results show that the proposed model can achieve a better accuracy than polynomial model.