TEC-Enhanced Deep Learning Model for Global 3D Ionospheric Electron Density Prediction

Although deep learning has made significant progress in predicting ionospheric parameters, such models often exhibit high accuracy during the training period but perform poorly when extrapolated to extended periods. In this study, we propose a global three-dimensional (3D) ionospheric electron density prediction model based on a Long Short-Term Memory (LSTM) network, which assimilates TEC observations to improve predictive performance, particularly during extended periods. Long-term electron density (Ne) observations from COSMIC-1 (2009–2019) and corresponding TEC derived from Global Ionospheric Maps (GIM) are used for the model training. During the training period, the model with TEC assimilation exhibited higher correlation on the test set and consistently showed lower root mean square error (RMSE) across different latitudinal regions. COSMIC-2 and Swarm-A observations were used to evaluate the models’ performance during the extrapolation period, and the model with TEC assimilation demonstrated significantly improved performance compared to the model without TEC. During the quiet period from DOY 203 to 207 in 2024, the model without TEC showed a pronounced systematic bias relative to COSMIC-2 observations. By contrast, the residuals of TEC-enhanced model closely followed a standard normal distribution, with the RMSE decreasing by 39.1%. During the geomagnetic storm period from DOY 131 to 135 in 2024, the RMSE of the TEC-enhanced model decreased by 25.6% compared to the model without TEC. The electron density distribution predicted by the TEC-enhanced model showed good agreement with Swarm-A satellite observations, whereas the model without TEC failed to capture the characteristic double-crest structure at low latitudes.