ATS-UNet: Attentional 2-D Time Sequence UNet for Global Ionospheric One-Day-Ahead Prediction

The ionosphere contains many charged free particles that significantly affect radio signals passing through it. Due to limitations in observation technology, high-precision ionospheric forecasting has attracted widespread attention. Deep learning is well-suited to address its high-dimensional nonlinearity. In this study, we propose the attentional 2-D time sequence UNet (ATS-UNet) model based on the UNet network combined with an attention mechanism. Using this model and global ionospheric observation data from 2008 to 2020, we develop a one-day forecast model for the global ionosphere. We select the 2015 and 2020 as test data and compared the ATS-UNet model with various state-of-the-art (SOTA) models currently used in global ionosphere forecasting. These models include adaptive autoregressive (AAR), long short-term memory (LSTM), ConvLSTM, and the UNet model. In 2015, the root mean square error (RMSE) value of the ATS-UNet model's forecast results decreases by 14% compared to the AAR model, 6% compared to the LSTM model, 4% compared to the ConvLSTM model, and 3% compared to the UNet model. In 2020, the RMSE value of the ATS-UNet model decreases by 7%, 8%, 6%, and 2%, respectively, when compared to the same models. The results demonstrate that the ATS-UNet model can effectively improve prediction accuracy.