LSTM-driven Prediction of Orbital Parameters for Accurate LEO Opportunistic Navigation

A novel Recurrent Neural Network (RNN)-based algorithm using Long Short-Term Memory (LSTM) networks has been developed to enhance the accuracy of Position, Navigation, and Timing (PNT) applications by predicting orbital parameters of Low Earth Orbit (LEO) satellites. Traditional models, such as Simplified General Perturbations 4 (SGP4), depend heavily on Two-Line Element (TLE) data, which are often updated infrequently and thus lead to the problem of error propagation in LEO satellites position and velocity estimation. The proposed model uses historical TLE data to train LSTM model, which then predicts orbital parameters for future epochs. Bayesian search technique for optimized hyper-parameter search has also been implemented in the model to ensure high-precision predictions. Using the historical TLE data of an Starlink satellite, it has been shown that the proposed method significantly reduces the Root Mean Squared Error (RMSE) in position and velocity estimates by up to 54% and 73% respectively compared to outdated TLE when an updated TLE is not available. The algorithm can be useful in the intervals between TLE updates to mitigate error propagation issues, allowing to obtain better satellite position and velocity estimates for positioning purpose. By integrating LSTM with traditional orbital mechanics, the proposed algorithm offers an improved approach for LEO satellites' orbital element prediction for propagating the position and velocity of a LEO satellite which is required for positioning with LEO satellites.