A Comparative Study on Different Machine Learning Approaches with Periodic Items for the Forecasting of GPS Satellites Clock Bias

Abstract Accurately predicting satellite clock deviation is crucial for improving real-time location accuracy in a GPS navigation system. Therefore, to ensure high levels of real-time positioning accuracy, it is essential to address the challenge of enhancing satellite clock deviation prediction when high-precision clock data is not available. Given the high frequency, sensitivity, and variability of space-borne GPS satellite atomic clocks, it is important to consider the periodic variations of satellite clock bias (SCB) in addition to the inherent properties of GPS satellite clocks such as frequency deviation, frequency drift, and frequency drift rate in order to improve SCB prediction accuracy and gain a better understanding of its characteristics. In recent applications, deep learning models have demonstrated significant improvements in handling time-series data. This paper presents four machine learning prediction models that take into consideration periodic variations. Specifically, we utilize precision satellite clock bias data from the International GNSS Service forecast experiments and assess the predictive effects of various models including backpropagation neural network (BPNN), wavelet neural network (WNN), long short-term memory (LSTM), and gated recurrent units (GRUs). We analyze clock bias prediction across different time scales and scenarios compared with the quadratic polynomial model. The results indicate that the WNN model incorporating periodic variations outperforms the standard quadratic polynomial model in terms of predictive accuracy. This highlights the promising potential of deep learning models in forecasting satellite clock bias.