Deep learning-based BDS-3 ephemeris and positioning performance prediction

Abstract The 3rd BeiDou Satellite Navigation System (BDS-3) provides global positioning, navigation, and timing (PNT) services. When precise ephemeris are unavailable, accurate prediction is essential. This study employs Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), and Transformer for ephemeris prediction, and proposes two Transformer-based positioning performance prediction methods. Predictions are compared with Single Point Positioning (SPP), using Precise Point Positioning (PPP) based on actual ephemeris as reference. In ephemeris prediction, Transformer outperforms GRU and TCN, achieving cm-to-mm accuracy. Radial orbit prediction is best (Root Mean Squared Error (RMSE) <0.02 m, Mean Absolute Error (MAE) <0.01 m for all except C39); clock error achieves RMSE ∼0.17 m and MAE ∼0.08 m. In the first positioning experiment using WUH2 historical data, Transformer and GRU show comparable PPP prediction, with Transformer achieving RMSE/MAE/R 2 of 0.027 m/0.011 m/0.941 and near-zero mean residuals; TCN degrades significantly. In the second experiment, the Transformer-based correction achieves average RMSE (Up (U)/North (N)/East (E)) of 0.509/0.207/0.427 m and MAE of 0.186/0.070/0.142 m, close to PPP and far superior to SPP, with RMSE improvements of 77.8 %/77.6 %/60.1 % and MAE improvements of 89.5 %/90.3 %/83.3 %. These results show the proposed method provides high positioning accuracy when the precise ephemeris is delayed or unavailable, effectively supporting BDS-3 services.