Neural network-based ionospheric modeling and predicting—To enhance high accuracy GNSS positioning and navigation

Abstract

As a highly variable space plasma, the sophisticated monitoring and expression of the ionosphere are critically important for the performance of Global Navigation Satellite System (GNSS). To satisfy the requirement of more precise positioning and navigation, it is necessary to model and predict the ionosphere in a more accurate way. In this paper, the potential advantages of neural networks in ionosphere modeling and predicting are explored. Radial Basis Function Neural Network (RBF-NN) is used for modeling and the Long- and Short-Term Memory Neural Network (LSTM-NN) is used for predicting. Results validate that performance of the RBF-NN is superior to the traditional spherical harmonic and polynomial model. The average root mean square error of one-day modeling residuals by RBF-NN is 1.2951 Total Electron Content Unit (TECU, 1 TECU = 1016 electron/m2), which is reduced by 0.4172 and 0.2710 TECU respectively, compared with the traditional spherical harmonic and polynomial model. The predicting results show that the LSTM-NN has better short-time performance, compared with the finite Fourier series model and the autoregressive integrated moving average model. Finally, one-day predicted products are used as priori constraints for single-frequency static and dynamic precise point positioning (PPP). Positioning results show that constraint effect of the predicted products is equivalent to that of the International GNSS Services (IGS) final products. Compared with the results without ionospheric constraint, the PPP convergence time is obviously reduced, and the positioning performance in the horizontal direction is significantly improved.