Improving Satellite Orbit Prediction Accuracy Using a Hybrid CNN-LSTM-Transformer Model

This study proposes a hybrid convolutional neural network–long short-term memory–Transformer (CNN–LSTM–Transformer) model for future SGP4 orbit error correction under complex perturbation conditions. The model learns the temporal evolution of historical orbital features and previously observed SGP4 residuals to predict future orbit error corrections, which are then applied to the original SGP4-propagated orbit to improve prediction accuracy.

Specifically, the CNN extracts local disturbance-related features, the LSTM captures short-term temporal dynamics, and the Transformer models long-range dependencies in orbital evolution. For the HY-2B satellite, the three-axis position RMSE is reduced from approximately 500 m to approximately 50 m, achieving an improvement of more than 90%. For the Etalon-1 and BeiDou-3M1 satellites, independently trained models achieve RMSE reductions exceeding 94%, demonstrating the applicability and robustness of the proposed approach across different orbital regimes and prediction timescales.