Improving LEO short-term orbit prediction using LSTM neural network

Abstract

Orbit prediction of low earth orbit (LEO) satellites is of paramount importance for LEO-augmented navigation. Currently, the most widely used approach for satellite orbit prediction in navigation domain is the dynamical propagation method, which necessitates a good understanding of orbital dynamics. However, this method is plagued by the rapid error accumulation as prediction time increases due to our limited knowledge of the complex orbit dynamics. An effective solution to this challenge is employing the machine learning algorithm, which is data-driven and requires no explicit physical knowledge, in orbit prediction of LEO satellites. We focus on improving LEO short-term (less than 120 min) orbit prediction using the long short-term memory (LSTM) neural network. To this end, we have constructed datasets of the entire year 2019 from seven LEO satellites to conduct the experiments. Historical orbit prediction errors, derived from the comparison between the dynamical-propagation-based predicted orbit and external precise orbit, along with multiple satellite status and environment features are trained to forecast future orbit prediction errors, which will subsequently serve as the compensation for improving the dynamical-propagation-based predicted orbit. Our findings reveal that the LSTM model can improve the accuracy of predicted orbit by more than 30 % for most LEO satellites with a maximum percentage around 75 %. Benefiting from the LSTM model, the prediction time for obtaining 5-cm accuracy of predicted orbit can be extended to (41.2, 42.0, 31.2, 37.9, 30.0, 86.3, 108.1) min for GRACE-C/D, Swarm-A/B/C, and Sentinel-3A/3B satellites, respectively. Additionally, generalization tests between different LEO satellites suggest that the LSTM model exhibits a commendable generalization ability when orbit prediction time is less than 30 min. As the prediction time increases, the model trained using datasets from one LEO satellite is more suitable for forecasting orbit prediction errors of multiple LEO satellites with comparable orbital altitude and orbital plane.