FDLSTM: A satellite orbit prediction model with frequency-domain feature fusion

Current satellite orbit prediction methodologies confront significant challenges. Physical models struggle due to the complexity of the space environment, and deep learning approaches often lack rigorous analysis regarding their viability for on-orbit deployment. Furthermore, both approaches face difficulties in directly supporting collision prediction tasks. To address these issues, this paper proposes a novel LSTM model enhanced with frequency-domain analysis, termed FDLSTM (Frequency-Domain enhanced LSTM), for satellite orbit prediction. Importantly, the FDLSTM model also offers a robust foundation for satellite collision prediction. The model employs a frequency-domain filtering module to decompose the orbital data into high-frequency and low-frequency features. Leveraging LSTM’s capacity to capture long-term dependencies and incorporating a Feedforward Neural Network (FFN) to augment generalization capability, FDLSTM effectively processes these decomposed features. Experiments on real TLE data demonstrate that compared with conventional LSTM, FDLSTM achieves significant improvements in both the coefficient of determination (R 2 ) and mean absolute error (MAE) for three-dimensional position and velocity prediction. This model enables more precise satellite orbit prediction, thereby establishing a solid foundation for collision risk assessment. The predicted trajectories exhibit high consistency with actual values, validating the effectiveness of the frequency-domain analysis module. These results confirm that FDLSTM maintains both stability and accuracy, demonstrating substantial application potential in satellite collision prediction tasks. • Proposed FDLSTM model integrating frequency-domain analysis for enhanced satellite orbit prediction. • Validated effectiveness of frequency-domain filtering module through ablation studies. • Demonstrated practical potential for satellite collision prediction tasks.