An XGBoost-Based Method for Improved Orbit Prediction With an Orbit-Separate Modeling Strategy

Wenbin Huang, Rui Tang, Guangzhi Qu, Feng Zhang

Abstract: This article introduces a novel approach to improving the accuracy of satellite orbit prediction (OP) by combining XGBoost with the extended Kalman filter (EKF). While EKF is a well-established method in orbit determination, it falls short when dealing with OP due to the absence of observation data during the prediction period. To the best of the authors' knowledge, this article is the first attempt to apply XGBoost, a leading ensemble learning method, to address this challenge. Furthermore, we adopt the orbit-separate strategy rather than the orbit-mixed strategy for prediction. This choice is based on the belief that the former outperforms the latter, given that it treats each orbit uniquely. This belief aligns with the understanding that each orbit follows its own physical laws, influenced by distinct perturbations, such as gravity, atmospheric drag, radiation pressure, and the gravitational pull of the Sun and Moon. The experimental results, conducted on five Starlink satellites, highlight the effectiveness of the XGBoost-based ensemble learning method when modeled with the orbit-separate strategy. It successfully captures the differences between predicted values from the physical model and true values simulated by the EKF. Notably, when compared with the state-of-the-art model employing an orbit-mixed strategy, the XGBoost model with an orbit-separate strategy demonstrates an overall improvement in prediction accuracy.