A Short-Term Forecasting Method for Ionospheric TEC Combining Local Attention Mechanism and LSTM Model

Total electron content (TEC) is an important parameter for studying ionospheric variations and space weather. Short-term prediction of the ionosphere plays a significant role in near-Earth space environment monitoring. This study proposes a model that combines a long short-term memory (LSTM) neural network with a localized attention mechanism (LAM), which weights the features using attention weights and connects them to a fully connected output layer. The model is constructed using TEC data from 16 Global Navigation Satellite System (GNSS) observation stations provided by the Crustal Movement Observation Network of China (CMONOC), along with six parameters: Bz, Kp, Dst, F10.7 indices, and hour of the day. The LAM-LSTM model is compared with the LSTM model and the BP model. Experimental results show that the LAM-LSTM model achieves a root mean square error (RMSE) of 1.35 TECU on average for the test dataset, while the LSTM model has an RMSE of 1.54 TECU, and the BP model has an RMSE of 1.74 TECU. The proposed model exhibits good stability in different geomagnetic conditions and different months.