Hybrid Physics-Informed Residual Learning for Robust BDS-3 Satellite Clock Bias Prediction.

Area of Science:

• Geodesy and Satellite Navigation
• Machine Learning in Geospatial Science

Background:

• Reliability of real-time precise point positioning (RT-PPP) depends on continuous precise satellite clock products.
• Interruptions in BeiDou Navigation Satellite System (BDS-3) clock products degrade positioning performance.
• Accurate satellite clock bias (SCB) prediction is crucial for maintaining RT-PPP services during outages.

Purpose of the Study:

• To develop a robust framework for predicting BDS-3 satellite clock bias (SCB).
• To enhance the continuity and reliability of precise positioning services under product outage conditions.

Main Methods:

• A hybrid framework combining a physics-informed neural network (PINN) and a long short-term memory (LSTM) network.
• PINN models the physically consistent trend component of SCB increments using clock dynamical constraints.
• LSTM learns and predicts the residual sequence capturing short-term stochastic variations.

Main Results:

• The proposed PINN-LSTM framework significantly outperforms conventional models (QP, ARIMA, CNN-LSTM, LSTM-Attention) in SCB prediction accuracy.
• Achieved improvements of up to 58.5% in prediction accuracy for 12h and 24h forecasting horizons.
• Demonstrated reduced long-horizon error accumulation, improved robustness across different satellite clock types, and stronger generalization.

Conclusions:

• The hybrid PINN-LSTM framework effectively supports the continuity and reliability of BDS-3 precise clock products.
• The proposed method shows strong potential for improving real-time precise point positioning (RT-PPP) applications.
• This approach offers a robust solution for mitigating the impact of satellite clock product outages.