Physics-Aware Neural Networks for Satellite Tracking

Satellite trajectory prediction is essential for space situational awareness and orbit determination.However, a purely data-driven model is unable to accurately predict a satellite trajectory in highly nonlinear orbital environments.To overcome these limitations, we propose a hybrid approach that integrates a physicsinformed neural network (PINN) with long short-term memory (LSTM) networks to enhance satellite trajectory prediction.The proposed LSTM-PINN model incorporates physics-based constraints into the learning process, thereby improving predictive accuracy and generalization.To evaluate the effectiveness of the proposed approach, we conducted an analysis by using real satellite trajectory data sets and compared its performance with a stand-alone LSTM model, as well as a traditional higher order Runge-Kutta (RK89) numerical propagator.The proposed hybrid model achieved a 34.4% reduction in position mean square error and an 18.1% reduction in mean absolute error, demonstrating enhanced long-term generalization relative to the baseline LSTM model.Overall, results demonstrate that the proposed model effectively captures the temporal dependencies of satellite motion, while offering a practical balance between accuracy and computational complexity.