A Forecasting Model of Ionospheric foF2 Using the LSTM Network Based on ICEEMDAN Decomposition

To further improve the short-term forecasting ability of the critical frequency of the ionosphere F2 layer (foF2), a sample entropy (SE) optimized deep learning (DL) long-short-term memory (LSTM) forecasting model based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is proposed. The ICEEMDAN-LSTM model uses the foF2 hour-level time series data of Dourbes station from 2009 to 2019 for training and verification and realizes a single-step high-precision foF2 time series forecast. Through the statistical analysis of the observation of foF2 parameters and the forecast results of the model, the ICEEMDAN-LSTM model can predict foF2 parameters well during the geomagnetic calm and storm periods. Moreover, the proposed model outperforms others in predicting foF2 time series under diurnal and seasonal variation. In the high solar activity year, the root mean square error (RMSE), relative root mean square error (RRMSE), mean absolute error (MAE), and $R<^>{2}$ evaluation indexes of the ICEEMDAN-LSTM model are 0.19 MHz, 4.33%, 0.13 MHz, and 0.99, respectively, and they are 0.22 MHz, 5.54%, 0.14 MHz, and 0.95 in the low solar activity year. The ICEEMDAN-LSTM has the highest forecast accuracy in different solar activity years and is almost unaffected by solar activity. Meanwhile, the prediction performance of ICEEMDAN-LSTM is also verified by observatories in other regions, with high forecasting accuracy. The above shows that the ICEEMDAN-LSTM model has good applicability and usability, and the forecast accuracy of foF2 short-term forecasting can be improved further.