Deep Learning‐Based Ionosphere Model Using Multi‐Channel Self‐Attention ED‐SA‐ConvGRU

ABSTRACT

The Equatorial Ionospheric Anomaly (EIA) occurs within the ionosphere on both sides of the Earth's magnetic equator, typically at magnetic latitudes of approximately ± 10° to ± 20°. In this region, the spatiotemporal evolution of ionospheric total electron content (TEC) is highly complex, leading to anomalous enhancements in electron density. These ionospheric disturbances become particularly pronounced during geomagnetic storms, posing significant challenges for high‐precision prediction. To address this prediction challenge, we introduce a deep learning model termed an Encoder‐Decoder with Self‐Attention Convolutional Gated Recurrent Unit (ED‐SA‐ConvGRU), where multiple physical indices were incorporated, including solar activity indices (F10.7, Solar Radio Flux at 10.7 cm; SSN, Sunspot Number) and geomagnetic indices (Kp, K‐index; Dst, Disturbance Storm Time Index), with three distinct input combinations constructed. This study utilised GNSS data from 70 stations of the Australian Regional GNSS Network (ARGN) from 2023 to 2025. Test results indicate that, compared to the GRU, ConvGRU, and ED‐ConvGRU models, the proposed ED‐SA‐ConvGRU model reduced the Root Mean Square Error (RMSE) by 16.8%, 2.4%, and 6.8%, respectively. Moreover, the prediction accuracy of the ED‐SA‐ConvGRU model was further enhanced through the incorporation of physical indices. Specifically, Input Combination III (historical TEC + geomagnetic indices Kp and Dst + solar activity indices F10.7 and SSN) yielded the lowest RMSE. These findings indicate that the ED‐SA‐ConvGRU model has competitive performance in ionospheric TEC prediction over the low‐to mid‐latitude EIA region.