A short-term prediction model of global ionospheric VTEC based on the combination of long short-term memory and convolutional long short-term memory

The ionospheric vertical total electron content (VTEC) is an essential parameter for studying the ionosphere's dynamic variations, and its short-term forecast is essential for some research and applications. In this study, we attempt to combine the long short-term memory (LSTM) network and the convolutional LSTM (ConvLSTM) to obtain more stable and reliable VTEC prediction results with fewer data. First, in the data preprocessing stage, the time series stationarity test, difference processing, and correlation analysis were performed on the spherical harmonic (SH) coefficient time series. Then, the LSTM + ConvLSTM model is built by combining the LSTM network and ConvLSTM. Finally, the VTEC prediction performance of the model under different geomagnetic conditions is evaluated. The results show that the LSTM + ConvLSTM hybrid model has better forecasting performance than the single LSTM and ConvLSTM models. The root-mean-square error (RMSE) values between the LSTM + ConvLSTM model's predicted VTEC during quiet, weak, moderate, and strong geomagnetic storms and the CODG VTEC are 0.69, 0.80, 0.91, and 1.10 TECU, respectively. Even during strong geomagnetic storms, 99.60% of the differences are within ± 5 TECU, and the model still has reliable results. The average Structural Similarity Index measure (SSIM) indices under various geomagnetic activity levels are 0.905, 0.895, 0.894, and 0.863, respectively. Compared with the traditional ionospheric prediction products, the performance of the LSTM + ConvLSTM model is improved in different degrees for different levels of geomagnetic storm periods. During strong geomagnetic storms, the performance improvement in the model is most obvious, with an RMSE reduction rate of more than 76% and an average SSIM index improvement rate of more than 80%.