Improving High-Precision BDS-3 Satellite Orbit Prediction Using a Self-Attention-Enhanced Deep Learning Model.

Area of Science:

• Satellite Geodesy
• Deep Learning Applications
• Navigation Systems

Background:

• Precise Global Navigation Satellite System (GNSS) orbit prediction is vital for real-time positioning.
• Current BeiDou Navigation Satellite System-3 (BDS-3) orbit prediction accuracy lags behind GPS and Galileo.
• Traditional dynamic modeling shows limited improvement for BDS-3 orbit prediction.

Purpose of the Study:

• To introduce a novel data-driven methodology, SCINet-SA, to enhance BDS-3 ultra-rapid orbit prediction.
• To improve the accuracy and reliability of BDS-3 orbit predictions.
• To address the disparity in prediction accuracy compared to other GNSS.

Main Methods:

• Developed a deep learning model, Sample Convolution and Interaction Network with Self-Attention (SCINet-SA).
• Modeled temporal characteristics of orbit differences between BDS-3 ultra-rapid and final products.
• Incorporated a self-attention mechanism to capture long-range temporal dependencies.

Main Results:

• SCINet-SA demonstrated superior performance in enhancing BDS-3 ultra-rapid orbit prediction accuracy.
• Achieved highest average relative improvement (IMP) in 3D Root Mean Square (RMS) error across 1, 7, and 15-day horizons.
• Reported significant IMPs, ranging from 7.78% to 38.91% for 1-day predictions.

Conclusions:

• SCINet-SA effectively refines ultra-rapid orbit estimates by predicting discrepancies.
• The self-attention mechanism enhances long-term prediction capabilities and mitigates latency.
• The proposed methodology represents a significant advancement in BDS-3 orbit prediction accuracy.