Using Long Short-Term Memory Neural Network for Satellite Orbit Prediction Based on Two-Line Element Data

Abstract

The space environment is becoming increasingly crowded, raising the likelihood of collisions between satellites. Accurate prediction of satellite orbits is crucial for space transportation and communications. This article proposes an orbit prediction method based on the long short-term memory (LSTM) neural network algorithm and two-line elements (TLEs). The effectiveness of the proposed method was validated and evaluated by selecting space objects from different orbits [low Earth orbit (LEO), medium Earth orbit, and geostationary Earth orbit]. Six months of TLE data for these space objects were collected. The predicted orbits for LEO using the LSTM and Simplified General Perturbation Version 4 methods were compared with reference orbits derived from precision orbits released by the International Laser Ranging Service. Calculations were performed every two days for six months of data, and the results indicate that LSTM can improve the orbit prediction accuracy of these satellites by at least 20% over half a month.