Temporal convolutional network construction and analysis of single-station TEC model

Abstract

Ionosphere is one of the main error sources of global navigation satellite system (GNSS) precise positioning, and affecting communicate services such as communication, broadcasting, and radar positioning. Total electron content (TEC) is a key parameter to characterize the state of the ionosphere. Establishing a high-precision TEC model and making accurate predictions can effectively improve positioning accuracy and improve communication quality. The traditional TEC model has limited ability to describe the changes of TEC under extreme conditions such as magnetic storms. Based on the temporal convolution network (TCN) model, this paper conducts experiments on TEC grid data in six low latitude regions and six mid latitude regions, and compares them with Long short term memory (LSTM), gated recurrent units (GRU) and bidirectional long short term memory (BiLSTM) models. Results show that the mean average error (MAE) of TCN (1.2385 TECU) is lower in most areas compared with LSTM (1.2727 TECU), GRU (1.2602 TECU) and BiLSTM (1.2767 TECU). And the TCN model shows better performance in the mid latitude regions (0.8778 TECU) than low latitude regions (1.5992 TECU). Then, this paper takes 1st October to 31st December 2021. as an example to calculate the prediction accuracy of the TCN model in the magnetic quiet period and the magnetic storm period. During the sample time, there were 4 weak geomagnetic storms, 1 strong geomagnetic storm, and there was a continuous long magnetic resting period at the same time, with a variety of different geomagnetic activities. The results show that the MAE distribution of the TCN model is more concentrated in the magnetostatic period, and the model error in the mid latitude region is normally distributed between -4-4.5 TECU. During the magnetic storm period, the TCN model has the lowest proportion of errors exceeding 5 TECU, and the proportions in the mid latitude and low latitude regions are 2.8% and 10.4%, respectively, which are better than the comparison model. Finally, we discuss the performance of short-term TEC prediction and the possible causes of obvious errors. The accuracy of the TCN model reaches 1.07 TECU, which is better than the long-term prediction result (1.24 TECU), and the accuracy is the best among the four models. After the detection of TEC anomaly disturbance, we believe that the obvious errors in the three experimental grids in north america are related to hurricane ELSA.