Forecasting of Global Ionosphere Maps With Multi-Day Lead Time Using Transformer-Based Neural Networks

Ionospheric total electron content (TEC) is a key indicator of the space environment. Geophysical forcing from above and below drives its spatial and temporal variations. A full understanding of physical and chemical principles, available and well-representable driving inputs, and capable computational power are required for physical models to reproduce simulations that agree with observations, which may be challenging at times. Recently, data-driven approaches, such as deep learning, have therefore surged as means for TEC prediction. Owing to the fact that the geophysical world possesses a sequential nature in time and space, Transformer architectures are proposed and evaluated for sequence-to-sequence TEC predictions in this study. We discuss the impacts of time lengths of choice during the training process and analyze what the neural network has learned regarding the data sets. Our results suggest that 12-layer, 128-hidden-unit Transformer architectures sufficiently provide multi-step global TEC predictions for 48 hr with an overall root-mean-square error (RMSE) of similar to 1.8 TECU. The hourly variation of RMSE increases from 0.6 TECU to about 2.0 TECU during the prediction time frame. Total electron content (TEC) is an important quantity and indicator in the space environment. One benefit of a good TEC estimation is to improve satellite positioning accuracy. Since physical models are limited to available and well-representable driving inputs, recent studies on TEC prediction have leaned toward data-driven approaches, such as deep learning. We need a sequence-to-sequence prediction model for the time-series data, where Transformers have been an excellent architecture in recent years. In this paper, we use a Transformer encoder to build a prediction model and explore the input and output structure of the model, to understand how temporal and spatial dimensions affect the TEC prediction problem. We also discuss the prediction tasks of different time lengths to analyze what the neural network has learned since the given data are based on some integration and fitting methods. Transformer-based neural architectures with careful input layouts are efficient and effective for spatio-temporal data modeling Proposed self-attention layers provide state-of-the-art global predictions with an root-mean-square error of less than 2 TECU without geo-related information Local total electron content extractions from 2D global predictions show a good agreement with the Central Weather Bureau observations for 99.5% of the time