Comparative Analysis of Four Neural Network Methods for TEC Modeling during Ionospheric Magnetic Storms

The Total Electron Content (TEC) of the ionosphere is an important parameter for describing the ionosphere activities, and much research has been done for the accurate methods for the ionospheric TEC prediction. However, the prediction accuracy of ionospheric empirical models for TEC during geomagnetic storms is still not ideal. To address this issue, this paper aims to assess the performance of ionospheric TEC predicting methods, which involve the LSTM, the BiLSTM, the convolutional neural network-long short-term memory combined with attention mechanism (CNN-LSTM-Attention), and the convolutional neural network-bidirectional long short-term memory combined with attention mechanism (CNN-BiLSTM-Attention). At first, the geomagnetic storm periods are identified by comparing with the threshold of Dst index (≤−30 nT), during the years from 2004 to 2022. Then, four neural network models for the ionospheric TEC prediction are formed, through the combinations of multiple spatiotemporal parameters, such as UTS, UTC, SA, AA, CHS, and SHS. Finally, the accuracy and reliability of the four neural network models are assessed using the reference TEC dataset collected during geomagnetic storms in 2023, and three statistical index, Mean Absolute Error (MAE), The Root Mean Square Error (RMSE), and coefficient of determination R², are utilized. The results show that, the performance of the CNN-BiLSTM-Attention model is superior to the other three models, with MAE ranging from 0.882 to 5.270 TECU, RMSE between 1.175 and 6.983 TECU, and R² values exceeding 0.7. In order to better describe the difference between the predicted values and the reference values, the scatter plots of two datasets are plotted for the fitting of linear regression equations. The slope of fitted function from CNN-BiLSTM-Attention model is very close to the ideal value 1, also indicating a better performance compared to the other models.