GPS Navigation Satellite Clock Bias Prediction Based on NBeats Neural Network

Satellite clock bias is one of the crucial factors affecting the accuracy of navigation positioning. Establishing a high-precision clock bias prediction model is of great significance for improving the accuracy of satellite navigation positioning. Currently, there are numerous models for predicting satellite clock bias, but a unified model that is both stable and highly accurate is still lacking. In this paper, a new neural network clock bias prediction method based on the NBeats model is proposed. The NBeats model uses stacked fully connected layers to construct stacks, and then specifies the characteristics of different stacks using the classical STL time series decomposition method. Finally, the outputs of each stack are aggregated to model time series data. Experimental comparisons demonstrate that the proposed model outperforms traditional models such as QP, GM(1,1), and ARIMA in the short to medium term (within 12 hours) satellite clock bias prediction. Additionally, compared to a neural network model based on first-order differenced data (LSTM), the proposed model shows improvement and greater stability in forecasting result, which can reach 0.45ns RMSE with 21% improvement for 3-hours SCB prediction.