Enhancing satellite clock bias prediction in BDS with LSTM-attention model

Satellite clock bias (SCB) is a critical factor influencing the accuracy of real-time precise point positioning. Nevertheless, the utilization of real-time service products, as supplied by the International GNSS Service, may be vulnerable to interruptions or network failures. In specific situations, users may encounter difficulties in obtaining accurate real-time corrections. Our research presents an enhanced predictive model for SCB using a long short-term memory (LSTM) neural network fused with a Self-Attention mechanism to address this challenge. This fusion enables the model to effectively balance global attention and localized feature capture, ultimately enhancing prediction accuracy and stability. We compared and analyzed our proposed model with convolutional neural network (CNN) and LSTM models. This analysis encompasses an assessment of the model's strengths and suitability for predicting SCB within the BeiDou navigation system, considering diverse satellites, orbits, and atomic clocks. Our results exhibit a substantial improvement in predictive accuracy through the LSTM-Attention model. There has been an improvement of 49.67 and 62.51% compared to the CNN and LSTM models in the 12-h prediction task. In the case of the 24-h prediction task, the improvements escalated to 68.41 and 71.16%, respectively.