Deep Learning-Based Prediction of Global Ionospheric TEC During Storm Periods: Mixed CNN-BiLSTM Method

The application of deep learning in high-precision ionospheric parameter prediction has become one of the focus in space weather research. In this study, an improved model called Mixed Convolutional Neural Networks (CNN)-Bi-Long Short Term Memory is proposed for predicting future ionospheric Total Electron Content (TEC). The model is trained using the longest available (25 years) Global Ionospheric Maps-TEC and evaluated the accuracy of ionospheric storm predictions. The results indicate that using historical TEC in the solar-geographical reference frame as input driving data achieves higher prediction accuracy compared to that in the geocentric coordinate system. Additionally, by comparing different input parameters, it is found that incorporating the Kp, ap, and Dst indices as inputs to the model effectively improves its accuracy, especially in long-term forecasting where R2 increased by 3.49% and Root Mean Square Error decreased by 13.48%. Compared with BiLSTM-Deep Neural Networks (DNN) and CNN-BiLSTM, the Mixed CNN-BiLSTM model has the highest prediction accuracy. It suggests that the utilization of CNN modules for processing spatial information, along with the incorporation of DNN modules to incorporate geomagnetic indices for result correction. Moreover, in short-term predictions, the model accurately forecasts the evolution process of ionospheric storms. When extending the predicted length, although there are cases of prediction errors, the model still captures the entire process of ionospheric storms. Furthermore, the predicted results are significantly influenced by longitude, magnetic latitude, and local time. To address some limitations in the research on ionospheric storm prediction using deep learning, this paper developed a model called Mixed Convolutional Neural Networks - Bi-Long Short-Term Memory that considers both temporal and spatial information to predict Total Electron Content (TEC). Subsequently, the predicted TEC is used to identify the evolution of ionospheric storms. The data set used spans from 1998 to 2023 and consists of global TEC measurements. The evaluation of the model's performance was conducted using selected data samples. The evaluation reveals that using data in the solar-geographical reference frame, along with geomagnetic indices, enhances the prediction accuracy of the model. Additionally, the model exhibits higher prediction accuracy in low latitude and mid latitude regions compared to high latitude regions, and higher accuracy during daytime compared to nighttime. Meanwhile, the influence of geomagnetic index on long-term prediction is more significant. Development of an improved deep learning model for future Total Electron Content prediction The model enhances accuracy in solar-geographical reference frame by incorporating Kp, ap, Dst indices The evaluation confirmed accurate prediction of ionospheric storm occurrence, disturbance amplitude, and evolution process by the model