A Hybrid EMD-ICA-DLinear Multi-View Representation Model for Accurate Satellite Orbit Prediction in Space

Abstract

Accurate prediction of the on-orbit positions of Low Earth Orbit (LEO) satellites is essential for mission success, operational efficiency, and safety. Nevertheless, the non-stationary nature of orbital data and sensor noise presents significant challenges for accurate prediction. To address these challenges, we propose a novel forecasting model, EMD-ICA-DLinear, which combines trend-residual representation with EMD-ICA in an innovative manner. By integrating the TSR (Trend, Seasonality, and Residual) framework with the EMD-ICA dual perspective, this approach provides a comprehensive understanding of time series data and outperforms traditional models in capturing subtle nonlinear relationships. When predicting the orbital position of the Fengyun-3C satellite, the model uses MSE and MAE as evaluation metrics. Experimental results indicate that the proposed EMD-ICA-DLinear hybrid model achieves MSE and MAE values of 0.1101 and 0.1567, respectively, when predicting the orbital position of the Fengyun-3C satellite 6 h in advance, representing reductions of 37.87% and 19.85% compared to the best baseline model, TimesNet. This advancement enhances satellite orbit prediction accuracy, supports operational stability, and enables timely adjustments, thereby improving mission efficiency and safety.

Keywords

• LEO satellites
• orbital prediction
• deep learning
• empirical mode decomposition
• independent component analysis
• DLinear