Deep neural network based precise orbit prediction for low earth orbit (LEO) satellites

Bofeng Li,Tianhao Wu,Haibo Ge

High-accuracy orbits are the prerequisite for precise applications with all current Low Earth Orbit (LEO)-borne remote sensing and future LEO-enhanced GNSS (LeGNSS). Owing to the time latency mainly caused by the time consumption of computing precise orbit determination (POD), the predicted orbits of GNSS and LEOs are inevitably required for real-time applications. However, it is rather difficult to precisely predict the LEO orbits due to their perturbation complexity. In this paper, we propose a Deep Neural Network (DNN) based approach of precise LEO orbit prediction, where the errors of reduced-dynamic orbit prediction (RDOP) are further compensated by properly designing DNN-based Sequence-to-Sequence (Seq2Seq) structures. The GRACE Follow-On satellites are taken to numerically validate the efficiency of our method. The results show that with appropriate strategies, 3D RMSEs of 6.9, 14.0 and 22.8 cm can be achieved by RDOP for half-, 1-, 2-h prediction arc. With the compensation of trained Seq2Seqs, the prediction accuracies can be significantly improved by 50–80% for all three directions, where the RMSEs of 5-min, half-hour, and 1-h are 0.6, 2.9, and 6.0 cm for 3D, respectively, and 0.4, 1.8, and 3.7 cm for OURE, respectively. Overall, the predicted orbits with accuracy of 5 cm are achievable for the prediction arc as long as 50 min, which asserted the great potential of proposed DNN-based prediction approach in the high-accuracy LEO orbit prediction.