Research on Orbit Prediction Based on Multi-source Satellite Orbit Characteristics and Multi-head Attention Mechanism-LSTM Model

This strategy enables the model to capture satellite-specific error evolution and common orbital error patterns across similar satellites. Experimental results show that residual rates in the X, Y and Z directions are 0.24%, 0.16% and 0.20%, respectively, with root mean square errors of 1.17 m, 0.96 m and 0.75 m, outperforming Long Short-Term Memory(LSTM), Back Propagation Neural Network(BP) and Support Vector Machine(SVM) methods.

Further experiments analyze the effects of training-satellite number, neural units and sampling interval on model performance, and verify its applicability to different prediction durations, space targets and GLONASS. The results demonstrate that multi-head attention mechanism-LSTM(MHALSTM) improves SGP4 orbit prediction error correction and supports high-precision space target prediction and space situational awareness.

However, its performance remains constrained by orbital similarity, sample size and complex low-Earth-orbit perturbations, and its long-term stability in complex environments requires further verification.