Time-Domain Modeling And Prediction of Global Ionospheric Total Electron Content: Applications of Attention Mechanisms

The global Total Electron Content (TEC) is a critical parameter to present ionosphere morphology, at the same time it significantly affects the propagation of trans-ionosphere radio waves, especially the global navigation satellite signals. The variations in ionosphere can cause satellite signal delays. To address the problem, it is necessary to make accurate prediction of ionosphere TEC so as to strengthen navigation performance. With the rapid development of deep learning techniques, time-series prediction models based on deep learning networks have demonstrated great potential in TEC prediction. Compared to traditional methods, deep learning models have many advantages in capturing complex spatiotemporal dependencies in TEC series. To address the problem, this work compares the performance of five deep learning models in application of ionosphere TEC prediction, focusing particularly on the usage of attention mechanisms on improvement for prediction accuracy. In the experiments, a number of leading deep learning models are evaluated and tested for short-term ionosphere TEC prediction, including the long-short term memory (LSTM) , those attention-based networks like CNN-BiLSTM-Attention, Transformer, PatchTST, the TimesNet based on convolutional kernel was also considered. Two years of TEC series are selected for testing, the year 2014 as the solar maximum in solar cycle 24, and the year 2017 in the solar descending phase of solar cycle24. The advantage of attention-mechanism is analyzed and compared. The results show that the PatchTST model has significant improvement in prediction accuracy, effectively capturing spatiotemporal variations in TEC. Models with attention mechanisms show better performance in prediction accuracy and reliability. By dynamically adjusting the hyper-parameters indifferent parts of the input sequence, the attention mechanism helps to capture variation features that have greater impact on prediction in time series, thus effectively improving the prediction performance in different temporal scales.