Research on LSTM space target orbit prediction based on attention mechanism enhancement

To address the problem that orbit prediction for space targets is highly dependent on physical models and initial conditions, and errors cannot be completely eliminated, a satellite orbit prediction and correction method is proposed that integrates an attention mechanism (Attention) with a Long Short-Term Memory (LSTM) network. This method uses the LAGEOS satellite as the research object. Using positional error, velocity, and acceleration features extracted from its historical orbit data, a deep learning model is trained to predict the orbit error for the next day and correct the errors in the SGP4 orbit prediction results. Experimental results show that the ATLSTM model outperforms the LSTM and Support Vector Machine (SVM) models in orbit prediction. The model is trained on seven days of data, achieving even better predictions and lowering the prediction error. The residual rates on the X, Y, and Z axes are as low as 3.68%, 4.77%, and 2.37%, respectively, successfully improving the accuracy of satellite orbit error prediction. Furthermore, this study validates the effects of varying the number of neural units and training duration on the model's prediction results. In addition, the model’s generalization capability is validated on the HY-2B satellite. These results demonstrate that combining machine learning with physical models can effectively improve orbit prediction accuracy, providing a new approach for space target orbit prediction.