Ionospheric TEC Prediction Model Based on LSTM Spatio-temporal Transformer

The ionosphere is a major source of error for satellite navigation, communication, and other applications, and the Total Electron Content (TEC) of the ionosphere is an important parameter for studying the temporal and spatial variations of the ionosphere, and it is extremely important to accurately predict the ionospheric TEC under different space weather conditions. Existing prediction models, when using auxiliary parameters such as solar activity and geomagnetic activity to improve the performance of ionospheric TEC prediction models, treat the auxiliary parameters as global covariates, ignoring the fact that the auxiliary parameters, although having the same value at each location, have different effects on the ionospheric TEC. To solve this problem, a combined ionospheric TEC prediction model (LSTM-STT) is proposed in this paper, which combines the Space-Time Transformer (STT) with the Long-Short-Term Memory (LSTM) and introduces the space-time attention mechanism. The model adopts the TEC data of China and its surrounding areas from 2000 to 2023 provided by the Center for Orbit Determination in Europe (CODE) of the International GNSS Service organization (IGS), with a time range of 8766 days, and the data are processed by the sliding window method, and the model takes the TEC data of the first 48 hours and the auxiliary parameters as inputs, and the TEC data of the last 24 hours after the prediction are constructed with 8764 samples. A total of 8764 samples were constructed. To verify the performance of the model, experimental prediction analyses were conducted in 2018 (a low solar activity year) and 2023 (a high solar activity year). The results show that the model has an average root mean square error of 1.3981 TECU and an average relative accuracy of 90.2524% on the 2018 test set, and an average root mean square error of 4.6262 TECU and an average relative accuracy of 89.9208% on the 2023 test set, which indicates that the model has good prediction performance.