Improvements to global ionospheric forecasting with a recurrent convolutional neural network

Abstract

Single-frequency GNSS users are reliant on estimates of the Total Electron Content (TEC) along lines of sight to navigation satellites to correct for ionospheric propagation delay and the resulting positioning errors. The parametric correction methods in use (Klobuchar’s algorithm for GPS and the NeQuick-G model for Galileo) can compensate for a large fraction of the delay but are hindered by using only a few daily coefficients to describe the ground truth ionosphere state. This loss of state information is particularly detrimental during periods of high deviation from baseline TEC patterns, e.g. solar weather events. This work describes an autoregressive RNN/CNN approach for spatiotemporal TEC forecasting from windowed historical map products, preserving local temporal and geospatial dependence between samples. By leveraging a large dataset spanning from 2000-2020 and applying convolutional transformations over both the temporal and spatial dimensions of the data, this model exhibits improved performance for time horizons up to 48 hours, compared to neural network-based approaches described in the literature to date.