Enhancing Medium-Orbit Satellite Orbit Prediction: Application and Experimental Validation of the BiLSTM-TS Model

To mitigate the limited accuracy of the Simplified General Perturbations 4 (SGP4) model in predicting medium-orbit satellite trajectories, we propose an enhanced methodology integrating deep learning with traditional algorithms. The developed BiLSTM-TS forecasting framework comprises a Bidirectional Long Short-Term Memory (BiLSTM) network, trend analysis module (T), and seasonal decomposition module (S). This architecture effectively captures sequential dependencies, trend variations, and periodic patterns within time series data, thereby improving prediction interpretability. In our experimental validation, we chose Beidou-2 M6 (C14), GSAT0203 (GALILEO 7), and the Global Positioning System (GPS) satellite named GPS BIIR-13 (PRN 02) as representative satellites. Satellite position data derived from conventional orbital models were input into the BiLSTM-TS framework for statistical learning to predict orbital deviations. These predicted errors were subsequently combined with SGP4 model outputs obtained through Two-Line Element set (TLE) data analysis to minimize overall trajectory inaccuracies. Using BeiDou-2 M6 (C14) as a case study, results indicated that the BiLSTM-TS implementation achieved significant error reduction; mean-squared error along the X-axis was reduced to 0.0309 km2, with mean absolute error of 0.1245 km, and maximum absolute error was constrained to 0.4448 km.