Ionospheric VTEC Maps Forecasting Based on Graph Neural Network With Transformers

The precise and timely forecasting of vertical total electron content (VTEC) in the ionosphere is crucial for navigation, communication systems, and space weather monitoring. Recent research has applied deep learning to predict VTEC maps, treating them either as images or sequences using convolutional neural networks (CNN), recurrent neural networks, or Transformers. However, these approaches overlook the intrinsic non-Euclidean (spherical) nature of VTEC maps. To address this, our study proposes a novel spatial-temporal graph neural network (GNN) framework, termed GNNTrans. GNNTrans leverages a graph convolutional network to capture the spherical characteristics of VTEC maps. It integrates external factors, such as the Dst-index and ap index, through an attention mechanism for focused feature extraction and employs a transformer mechanism to model temporal patterns. Two variants of GNNTrans, isotropic and pyramid, were explored to determine the optimal structure. The pyramid model emerged as the top performer, achieving a root-mean-square error (RMSE) of 2.52 total electron content units (TECU). The isotropic model also outperformed the homogeneous CNN model in handling the spherical nature of VTEC maps, achieving 2.58 TECU compared to the CNN model's 2.65 TECU. Furthermore, GNNTrans surpassed the CODE one-day forecasting product across various dimensions, reducing the RMSE to around 3.3 and 1.3 TECU in 2014 and 2018, respectively, compared to C1P's 4.5 and 1.8 TECU. In addition, insightful visualizations and analyses shed light on GNNTrans's mechanisms, enhancing our understanding of its predictive capabilities. Overall, GNNTrans demonstrates remarkable performance, offering enhanced accuracy and reliability in predicting VTEC across diverse conditions.