Satellite Orbit Prediction Using an LSTM Residual-Learning and LightGBM Error-Correction Fusion Model

This study proposes a two-stage SGP4-LSTM-LightGBM residual-correction framework for medium- to long-term orbit prediction. The LSTM module learns the temporal evolution of SGP4 prediction errors, while LightGBM further refines the remaining nonlinear residuals, enabling targeted correction with low computational overhead.

Experiments on HY-2B, LARES-2, and BeiDou-3 M1 show that the proposed method significantly reduces RMSE and MAE compared with raw SGP4 and standalone LSTM. Compared with LSTM, prediction accuracy improves by 35%–57%, and compared with SGP4, the improvement exceeds 97%. In addition, ablation experiments, recurrent neural network baseline comparisons, second-stage residual-corrector comparisons, and the Wilcoxon signed-rank test further confirm the effectiveness and statistical reliability of the proposed framework.