Research on a Deep Learning Modeling Method of Ionospheric Total Electron Content

Abstract

The current popular deep learning technology of artificial neural networks has been vigorously developed and gradually applied to space weather. Global total electron content (TEC) map data publicly provided by the International GNSS Service (IGS) are 71×73 grids with temporal resolution of 1 hour and spatial resolution of 5°×2.5°, which can be used as sufficient training samples for deep learning. Variation of TEC is closely related to solar activity and geomagnetic activity, so data sets for this article are created based on 2019–2020 TEC map data, as well as the hourly solar radiation index F107 and geomagnetic index Kp together with the corresponding time, and longitude and latitude of each parameter are introduced as supervision information. The long short-term memory (LSTM) network and the multilayer perceptron (MLP) in the deep learning method are used to build a 9-layer deep neural network for training and verification, so that the advantages of the “gate” mechanism of LSTM network in time series modeling and the advantages of MLP in comprehensive consideration and high reliability can be fully brought into play. 70% of data sets are divided into training sets and 30% for validation, which runs in CPU environment. Adam algorithm is used for optimization, and the batch size is set to 24. The training results show that the minimum RMSE is 0.249 TECu, and the maximum RMSE is 4.240 TECu. RMSE of one step prediction is 0.650 TECu, and MAPE is 3.181%.