Research on Low Earth Orbit Satellite Orbit Prediction Based on a Two-stage Integrated Model

With the number of low Earth orbit satellites and the expansion of application demands, satellite trajectory prediction is crucial for efficiently operating satellite systems. Traditional trajectory prediction methods rely on classical physical models such as Newtonian mechanics and celestial mechanics. Still, these methods have insufficient accuracy and real-time performance when facing low-orbit satellites’ complex in-orbit environments. Compared with traditional physical models, neural networks can better capture complex changes in trajectories and compensate for the shortcomings of traditional methods. This article analyzes the effects of atmospheric drag, solar radiation pressure, and other factors on the orbital dynamics of low Earth orbit satellites, proposes a trajectory prediction model based on deep neural networks, and uses Long Short-term Memory networks (LSTM) to improve prediction accuracy. To further improve prediction accuracy, this study adopted an ensemble learning strategy, integrating LSTM, GRU, and Transformer models to construct a two-stage ensemble model. The experimental results show that the fusion model significantly surpasses the single model in both short-term and long-term prediction tasks. In short-term forecasting, the combination of LSTM and GRU can efficiently capture rapid changes in trajectory data. In long-term forecasting, Transformer significantly reduces prediction errors through its powerful modeling capabilities. The two-stage integrated model performs excellently in satellite trajectory prediction, with RMSE and MAE reduced by 21.8% and 14.3%, respectively in short-term prediction compared to the LSTM model, and RMSE and MAE reduced by 4.5% and 10%, respectively in long-term prediction.