Machine Learning Modelling of Geostationary SBAS Satellite Orbit Errors Observed from an Equatorial West African Station

Geostationary satellites used by satellite-based augmentation systems are continuously visible from equatorial Africa, but their raw broadcast ephemeris accuracy has received limited assessment in West Africa. This study measured and modelled the broadcast orbit errors of three SBAS GEO satellites, comprising EGNOS S23 and S36 and GAGAN S27, observed with a u-blox ZED-F9P receiver at Ogbomoso, Nigeria, from January to March 2026. Broadcast positions were propagated using a second-order Taylor model and compared with interpolated IGS Final precise orbits. Atmospheric drag and piecewise polynomial models were used as mathematical baselines before comparison with Random Forest, XGBoost, support vector regression and long short-term memory models. S23 and S36 maintained stable mean 3D errors near 1.27 m, with standard deviations of approximately 0.55 m. S27 recorded mean errors near 2.6 m in January and February and an isolated March maximum of 97.250 m. The atmospheric drag and polynomial baselines produced RMSE values of 1.329 and 1.293 m, respectively. LSTM achieved 1.282 m, while XGBoost, Random Forest and support vector regression reduced RMSE to 0.272, 0.250 and 0.171 m. The staged residual scheme achieved the lowest RMSE of 0.164 m, representing a 92.8% reduction from the uncorrected value of approximately 2.278 m, and produced a 2DRMS of 1.144 m. The results show that GEO broadcast orbit errors contain nonlinear, satellite-specific patterns that data-driven models can correct effectively, provided anomalous satellites are identified before use.