Research on Space Object Orbit Prediction Based on a BiLSTM Neural Network

To reduce the high modeling cost and limited adaptability of traditional dynamics-based orbit prediction methods, this paper proposes a bidirectional Long Short-Term Memory (BiLSTM)-based Simplified General Perturbations 4 residual correction method. The model adopts an enhanced long short-term memory architecture with independent forget gates and input gates to improve the capture of long-term temporal dependencies.

Axis-wise position residuals, velocity components, and the right ascension of the ascending node are selected as input features for orbit prediction. The proposed model was evaluated using HY-2B and WESTPAC as the main low Earth orbit test cases, with Beidou3m1 used as an additional satellite at a higher orbital altitude. Compared with the backpropagation model, the BiLSTM model generally achieved lower root mean square error and mean absolute error values, although the magnitude of improvement varied across satellites and coordinate axes.