Optimized BiLSTM-Based Short-Term Prediction of BeiDou Satellite Clock Bias for High-Precision GNSS Applications

Accurate satellite clock bias prediction is essential for improving the positioning accuracy and reliability of Global Navigation Satellite Systems (GNSS). However, the nonlinear, time-varying, and complex temporal characteristics of BeiDou satellite clock bias make accurate forecasting a challenging task for conventional prediction approaches. This study proposes an optimized Bidirectional Long Short-Term Memory (BiLSTM)-based deep learning framework for short-term prediction of BeiDou Navigation Satellite System (BDS) satellite clock bias. The proposed model is evaluated against traditional prediction methods, including Quadratic Polynomial (QP), Periodic (SA), Grey Model GM (1,1), and conventional Long Short-Term Memory (LSTM) networks. Experiments are conducted using high-precision BDS-3 satellite clock bias data obtained from the International GNSS Service (IGS), covering different satellite types, including Medium Earth Orbit (MEO), Inclined Geosynchronous Orbit (IGSO), and Geostationary Earth Orbit (GEO) satellites with different atomic clock configurations. The model performance is investigated under single-day and multi-day prediction scenarios with forecasting intervals of 1-hour, 6-hour, and 12-hour. Prediction accuracy is evaluated using Root Mean Square Error (RMSE), error range, and standard deviation analysis. The experimental results demonstrate that the optimized BiLSTM model consistently achieves superior prediction accuracy and more stable error characteristics compared with traditional and unidirectional deep learning models. The proposed approach effectively captures complex temporal dependencies in satellite clock bias sequences and provides a reliable solution for high-precision GNSS timing and navigation applications.