Research on Space Object Orbital Prediction Using Long Short-Term Memory Neural Networks

Addressing the challenges posed by the complex and dynamic space environment and the significant errors in space object orbit prediction, this study proposes a method to correct orbit predictions based on the Simplified General Perturbations 4 (SGP4) model. This method utilizes a Long Short-Term Memory (LSTM) neural network to predict positional errors. Focusing primarily on the “Ajisai” satellite as a case study, the LSTM model learns from historical orbital characteristics—including positional error, velocity, and acceleration—to predict positional error for the next day and optimize orbit predictions. Experimental results demonstrate that the LSTM method outperforms the SVM (Support Vector Machine) method, achieving good performance. The orbit errors in the [Formula: see text], [Formula: see text], and [Formula: see text] axes were reduced to 7.14%, 6.77%, and 8.39% of their original values, respectively, using the LSTM approach. We further investigated the impact of the number of hidden layer units on model performance. Additionally, to validate the model’s generalizability, we tested its predictive accuracy using orbital data from the low-Earth orbit satellite Larets. This method provides a valuable reference for research aimed at improving the prediction accuracy of space object orbits.