BiLSTM-Driven Error Compensation for High-Precision Orbit Prediction with Two-Line Elements

With the rapid increase in low Earth orbit space objects, accurate orbit prediction (OP) is becoming critical for satellite safety and space environment management. Traditional physics-based OP methods, especially those relying on publicly available two-line element (TLE) data, often struggle with accuracy due to complex orbital perturbations. However, the availability of long-term, large-scale TLE datasets enables the analysis of underlying orbital error patterns. This study proposes a high-precision OP error compensation method solely based on TLE data. A bidirectional long short-term memory (BiLSTM) neural network is used to model time-varying orbital error trends and correct TLE/simplified general perturbation model 4-based predictions. Experiments on 340 low Earth orbit objects show that the BiLSTM framework effectively captures temporal error patterns, achieving an average accuracy improvement of 82% and a compensation effectiveness exceeding 94% across all targets. To assess the model’s robustness, the impact of training sample size and sampling intervals is examined, revealing that proper data configurations markedly influence compensation performance. Comparative analysis with other deep learning models (recurrent neural network, gated recurrent unit, long short-term memory, and convolutional neural network) further confirms the superior fitting and correction ability of BiLSTM. This study highlights the potential of BiLSTM networks to enhance the accuracy and reliability of TLE-based OP, providing a valuable approach for improving space situational awareness.