Scholarlink

Abstract <p>Precise forecasting of ionospheric Total Electron Content (TEC) is critical for safeguarding the reliable operation of satellite navigation and communication systems. However, the severe spatiotemporal heterogeneity exhibited during intense geomagnetic storms poses formidable challenges to existing modeling approaches. To address this, we propose TransTCN-XA, a hybrid deep learning architecture. Featuring an innovative dual-stream decoupled design, the model utilizes a Temporal Convolutional Network (TCN) and a spatial Transformer to extract temporal and spatial features, respectively, while introducing a Cross-Attention mechanism to achieve seamless feature integration. Leveraging Global Ionosphere Maps (GIM) released by the Center for Orbit Determination in Europe, a representative dataset covering representative geomagnetic storm events from 2012 to 2023 was constructed for training and validation. Experimental results indicate that during the main phase of extreme magnetic storms—characterized by a Dst index dropping to − 163 nT—TransTCN-XA exhibits exceptional robustness, achieving an accuracy improvement of approximately 21.9% compared to the Convolutional Long Short-Term Memory (ConvLSTM) benchmark, and maintains robust performance even at the peak of disturbances. Furthermore, the proposed model reconstructs the evolutionary morphology of the Equatorial Ionization Anomaly (EIA) with high fidelity, significantly outperforming comparative models that suffer from structural blurring and deviations as high as 10 TECu. Temporal validation further reveals that the Pearson Correlation Coefficients (PCC) across different latitude bands consistently remain above 0.99, with Mean Absolute Error (MAE) fluctuations maintained below 0.3 TECu. These findings confirm the generalization capability and stability of TransTCN-XA in handling complex space weather events, providing robust technical support for high-precision ionospheric weather monitoring and early warning under extreme conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Global ionospheric TEC prediction for extreme space weather: a spatiotemporal dual-stream deep learning model coupled with cross-attention mechanisms

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Wang Li,
  • Fangsong Yang,
  • Xiujiang Pan,
  • Dongsheng Zhao,
  • Zhen Li,
  • Kefei Zhang

摘要

Abstract

Precise forecasting of ionospheric Total Electron Content (TEC) is critical for safeguarding the reliable operation of satellite navigation and communication systems. However, the severe spatiotemporal heterogeneity exhibited during intense geomagnetic storms poses formidable challenges to existing modeling approaches. To address this, we propose TransTCN-XA, a hybrid deep learning architecture. Featuring an innovative dual-stream decoupled design, the model utilizes a Temporal Convolutional Network (TCN) and a spatial Transformer to extract temporal and spatial features, respectively, while introducing a Cross-Attention mechanism to achieve seamless feature integration. Leveraging Global Ionosphere Maps (GIM) released by the Center for Orbit Determination in Europe, a representative dataset covering representative geomagnetic storm events from 2012 to 2023 was constructed for training and validation. Experimental results indicate that during the main phase of extreme magnetic storms—characterized by a Dst index dropping to − 163 nT—TransTCN-XA exhibits exceptional robustness, achieving an accuracy improvement of approximately 21.9% compared to the Convolutional Long Short-Term Memory (ConvLSTM) benchmark, and maintains robust performance even at the peak of disturbances. Furthermore, the proposed model reconstructs the evolutionary morphology of the Equatorial Ionization Anomaly (EIA) with high fidelity, significantly outperforming comparative models that suffer from structural blurring and deviations as high as 10 TECu. Temporal validation further reveals that the Pearson Correlation Coefficients (PCC) across different latitude bands consistently remain above 0.99, with Mean Absolute Error (MAE) fluctuations maintained below 0.3 TECu. These findings confirm the generalization capability and stability of TransTCN-XA in handling complex space weather events, providing robust technical support for high-precision ionospheric weather monitoring and early warning under extreme conditions.