Analysis and comparison of satellite clock error prediction based on various deep learning algorithms

Aiming at the problems of the low applicability of the satellite clock error prediction model and the insufficient combination of the type of the satellite-borne atomic clock and the modeling characteristics in the prediction model, four kinds of neural network models suitable for nonlinear processing are proposed to predict satellite clock error. Firstly, the clock error data is preprocessed. Then, the firefly algorithm models were established by using the back-propagation (FA-BPNN) model, the Elman cyclic (Elman) model, the radial basis function (RBF) model, and the convolutional neural network data of 1 d and 7 d based on the CNN-LSTM model GPS precise clock error data from the Wuhan University International GNSS service (IGS) data analysis center (WHU) are used for clock error prediction At last, the effect of the prediction is analyzed and compared from the point of view of different modeling data and different batches of satellites with the same type of atomic clock and different batches of satellites with different types of atomic clock. The results show that: 1) the modeling accuracy of 1 d clock error data is higher than that of 7 d clock error data, and the RBF model has the greatest influence on the prediction accuracy with the increase of clock error data, and the prediction accuracy changes from sub-nanosecond to tens of nanosecond. 2) the prediction accuracy of the four neural network models is related to the satellite operating time in orbit and the type of atomic clock on board. The prediction performance of the satellites with long operating time in orbit is not necessarily bad, and the prediction performance of different types of atomic clock on different batches of satellites may be the same. The cesium atomic clock type satellite has the best prediction accuracy among the four neural network models.