Forecast of Ionosphere GNSS TEC using CNN and LSTM neural network

Ionospheric total electron content (TEC) is an essential parameter for describing the ionosphere state and thus for space weather monitoring, and there has being great interest in the community in predicting short-term GNSS TEC variations. Nowadays, deep learning techniques, such as the convolutional neural network (CNN) and the long short-term memory (LSTM) neural network, are very promising and has been widely used in various aspects. CNN and LSTM are suitable for extracting spatial feature and temporal-dependent sequence, respectively, so in this study a combination of CNN and LSTM architecture is used for spatio-temporal feature extraction to predict 2D TEC sequence. Specifically, we first define the CNN layers as time series of 2D TEC data over China sector, and then wrap them into a Time-Distributed layer as input to the LSTM layer. In addition, multiple datasets describing the external drivers, such as solar wind speed, dynamic pressure, IMF By and Bz, are also concatenated into the LSTM layer as additional inputs. The final output of the CNN-LSTM network is the predicted TEC for the next few hours at all GNSS stations.