Abstract
Total electron content (TEC), as a crucial ionospheric parameter, has impacts on electromagnetic wave propagation as well as satellite navigation and positioning, and is of great significance in space weather forecasting. Previous prediction efforts using neural network techniques have basically focused on years with relatively low solar activity. In this study, a model combining Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) network has been constructed to forecast the TEC during high solar activities from a single Global Navigation Satellite System (GNSS) receiver at Sanya in Hainan, China. The performance of the CNN-GRU model is compared with the most used empirical models, IRI and NeQuick, and two artificial intelligence models, GRU and SVM. Benefiting from CNN's superior data feature capture capability of convolutional operation, the CNN-GRU model surpasses the original GRU model not only in 1-h-ahead predictions with an RMSE of 4.28 TECU but also in 24-h forecasts, boasting a notably lower average RMSE of 6.94 TECU, undoubtedly also outperforming the remaining models, SVM, NeQuick2, and IRI2020. Furthermore, the CNN-GRU model exhibits stable and excellent performance across different months and hour of the day, even during geomagnetic storms.