Physics-informed neural network for predicting orbital parameters of Low Earth Orbit satellites using two line element dataset

This research introduces a novel Physics-Informed Neural Network (PINN) framework using Long Short-Term Memory (LSTM) architecture as the baseline model for the precise prediction of orbital parameters of Low Earth Orbit (LEO) satellites, relying exclusively on publicly available Two-Line Element (TLE) datasets. The proposed method is validated for near-circular orbit with a near zero eccentricity in LEO orbits, which is a characteristic of operational LEO constellations such as Iridium, Orbcomm, Starlink etc. The main goal is to address the challenges linked to infrequent updates of TLE data, which generally decreases propagation accuracy with time. The proposed method efficiently incorporates fundamental physical laws of orbital mechanics into the neural network’s learning framework, improving prediction accuracy, stability, and generalizability relative to traditional solely data-driven approaches. Results from the experiments conducted utilizing historical TLE records of the IRIDIUM 7 and ORBCOMM FM26 satellites reveal that the integration of physics-based constraints significantly reduces prediction errors for satellite position and velocity. For IRIDIUM 7 the position and velocity estimates are improved by 87% and 70% respectively, which are 50% and 72% respectively for ORBCOMM FM26 satellite. The proposed PINN also demonstrates notable improvement over the baseline LSTM, with most R 2 improvement for inclination and mean anomaly predictions by approximately 59.5% and 59.1%, respectively. These improvements are crucial for high-precision applications such as Positioning, Navigation, and Timing (PNT) as well as space traffic management in the increasingly crowded LEO environment, where high fidelity satellite ephemerides are important.