Improved Starlink Satellite Orbit Prediction via Machine Learning with Application to Opportunistic LEO PNT

• DOI: 10.33012/2024.19892
• Publication Date: 2024-10-15T14:54:29Z
• Authors: Paul El Kouba; Samer Hayek; Joe Saroufim; Zaher M. Kassas; Evan Fakhoury
• Keywords: Orbit (dynamics)

A machine learning (ML) framework for improved Starlink low Earth orbit (LEO) satellites' prediction is presented. The exploits newly published SpaceX ephemerides files containing relatively errors during the first eight hours of release. This assumes two stages: (i) data processing stage that uses to learn error between given and propagated using simplified general perturbations (SGP4) model, which are subsequently used train a time-delay neural network (TDNN); (ii) forecasting over certain period time where estimated correct SGP4-propagated orbits. Simulation results presented showing ML approach achieved mean satellite position velocity 177 m 0.86 m/s, respectively. In contrast, ephemerides' were 2,535 2.75 An unknown receiver could use forecasted TDNN-improved localize itself Doppler measurements from overhead satellites. showcase improvement in stationary localization upon relying on ephemerides. Fusing 19 satellites 5-minute via an extended Kalman filter, if rely orbits itself, initial 6.5 km gets reduced 2.3 km, whereas TDNN-corrected reduces 53 m.