A Novel Method for Improving LEO Kinematic Real-Time Precise Orbit Determination With Neural Networks

The kinematic approach [i.e., kinematic precise point positioning (PPP)] has been extensively employed in real-time precise orbit determination (RTPOD) of low Earth orbit (LEO) satellites. Nonetheless, this approach typically suffers from a degraded accuracy and robustness due to its sensitivity to global navigation satellite system (GNSS) ephemeris errors and observation quality/geometry. To address this issue, we propose a novel method that incorporates the backpropagation neural network (BPNN) model into LEO kinematic RTPOD (KRTPOD) to improve its accuracy and robustness. The historical KRTPOD results are compared with external orbit products to compute orbit errors. Meanwhile, multiple features are generated from the historical KRTPOD processing. We train the BPNN model to discover the nonlinear relationship between these features and orbit errors. Based on the trained model, orbit errors of LEO KRTPOD can be forecast according to the input features and then serve as the orbit compensation to improve LEO orbits in real time. Data throughout 2022 from six LEO satellites are processed to examine the performance of this method. Our findings reveal that the proposed method can significantly improve the accuracy of LEO KRTPOD by up to 39.0%. Benefiting from this method, a promising accuracy of 3.2 cm can be achieved in LEO KRTPOD. We also demonstrate the superior generalization ability of the new method under different scenarios. Moreover, this method can enhance the robustness of LEO KRTPOD by mitigating the contamination caused by large GNSS ephemeris errors and accelerating convergence.