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

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