Ionospheric VTEC Map Forecasting based on Graph Neural Network with Transformers

Accurate and timely prediction of Total Electron Content (TEC) in the ionosphere is of paramount importance for various applications such as GNSS positioning and navigation, communication systems, and space weather monitoring. While recent years have witnessed the application of various deep learning techniques to this task, these methods often treat vertical total electron content (VTEC) maps as either images or sequences, disregarding the inherent non-Euclidean (spherical) nature of VTEC maps. Addressing this limitation, our study offers a novel perspective by introducing graph structures to represent VTEC data. This paper presents a groundbreaking approach, GNNTrans, which amalgamates the strengths of graph convolutional networks and transformer architectures to predict TEC. GNNTrans adeptly captures the intricate spatial and temporal dependencies intrinsic to VTEC maps. Through an ablation study, the results demonstrate graph structures and Graph Neural Networks (GNN) are superior to conventional Convolutional Neural Network (CNN) methods in extracting non-Euclidean spatial information from VTEC maps, achieving root mean square errors (RMSE) of 2.58 and 2.66. Additionally, experiments demonstrate GNNTrans's supremacy over the CODE one-day forecasting product across various dimensions, reducing the RMSE to 3.34 and 1.49 in 2014 and 2018 respectively, in contrast to C1P's values of 8.74 and 6.41. GNNTrans exhibits remarkable performance in predicting TEC variations across diverse conditions, thus holding promise for heightened accuracy and reliability in ionospheric TEC forecasting.