ED‐ST‐Transformer: A Global TEC Spatiotemporal Prediction Model Based on Multisource Data

The accurate prediction of Total Electron Content (TEC) is crucial for ensuring the safety of space missions and enhancing the stability of satellite navigation. Combining the advantages of Convolutional Neural Networks (CNN) in local feature extraction with the Transformer self‐attention mechanism's ability to capture global dependencies, this paper proposes a spatiotemporal prediction model based on an encoder‐decoder structure—ED‐ST‐Transformer, aimed at improving TEC prediction accuracy. Given the close relationship between TEC, solar activity, and geomagnetic disturbances, this study incorporates auxiliary inputs including the disturbance storm time index (Dst), planetary K‐index (Kp), 10.7 cm solar radio flux (F10.7), and sunspot number. Using multi‐source data from the preceding 24 hr (with a temporal resolution of 2 hr), the model is constructed to forecast global TEC maps for the subsequent 24 hr. To evaluate the performance of the ED‐ST‐Transformer model, this study compares it with the C1PG, ED‐ConvLSTM, and ED‐BiConvLSTM models in terms of overall prediction accuracy, error distribution, visualization effects, and performance in different months and latitude regions under various solar activity years. Additionally, the prediction performance of the model during geomagnetic storms is also analyzed. This outcome demonstrates that ED‐ST‐Transformer provides the most stable and accurate predictions under all conditions, outperforming the other models.