Forecasting Global Ionospheric TEC Using Deep Learning Approach

Abstract

Global ionospheric total electron content (TEC) maps are widely utilized in research regarding ionospheric physics and the associated space weather impacts, so there is a great interest in the community in short‐term ionosphere TEC forecasting. In this study, the long short‐term memory (LSTM) neural network (NN) is applied to forecast the 256 spherical harmonic (SH) coefficients that are traditionally used to construct global ionospheric maps (GIM). Multiple input data, including historical time series of the SH coefficients, solar extreme ultraviolet (EUV) flux, disturbance storm time (Dst) index, and hour of the day, are used in the developed LSTM NN model. Different combinations of the above parameters have been used in constructing the LSTM NN model, and it is found that the model using all four parameters performs the best. Then the best performing LSTM model is used to forecast the SH coefficients, and the global hourly TEC maps are reproduced using the 256 predicted SH coefficients. A comprehensive evaluation is carried out with respect to the CODE GIM TEC. Results show that the first/second hour TEC root mean square error (RMSE) is 1.27/2.20 TECU during storm time and 0.86/1.51 TECU during quiet time, so the developed model performs well during both quiet and storm times. Moreover, typical ionospheric structures, such as equatorial ionization anomaly (EIA) and storm‐enhanced density (SED), are well reproduced in the predicted TEC maps during storm time. The developed model also shows competitive performance in predicting global TEC when compared to the persistence model and two empirical models (IRI‐2016 and NeQuick‐2).

What sustained multi-disciplinary research can achieve: The space weather modeling framework

“…The model results show that the first/second hour TEC root mean square error (RMSE) is 1.27/2.20 TECU during storm time and 0.86/1.51 TECU during quiet time. Comparing with the CODE GIMS, the RMSE of the LSTM prediction is 1.06/1.84 TECU for the 1st /2nd hour, while the RMSE errors from the IRI-2016 and NeQuick-2 models are around 9.21/5.5 TECU, respectively (Liu et al, 2020). Moreover, typical large-scale ionospheric structures, such as equatorial ionization anomaly and storm-enhanced density are well reproduced in the predicted TEC maps during storm time.…”

“…Spherical harmonic (SH) fitting is often used in constructing the GIM map. We applied an LSTM neural network method (LSTM/NN, Hochreiter and Schmidhuber, 1997) to forecast the 256 SH coefficients, which are then used to construct the GIM maps (Liu et al, 2020). The model results show that the first/second hour TEC root mean square error (RMSE) is 1.27/2.20 TECU during storm time and 0.86/1.51 TECU during quiet time.…”

Proceedings - 3rd Congress in Geomatics Engineering - CIGeo

“…Areas of interest go from the core GNSS goal, improving localization (Hosseinyalamdary 2018; Diggelen 2019) to all other bias effect presented in Eq. ( 1) like multipath (Diggelen and Wang 2020; Quan et al 2018), ionosphere (Orus 2018;Jiao et al 2007;Liu et al 2020;Linty et al 2019) and troposphere (Benevides et al 2019) effects. These works demonstrate how ML can contribute to address problems addressed in a different way.…”

Classification of High Density Regions in Global Ionospheric Maps With Neural Networks

“…McGranaghan et al (2017) demonstrated successful application of network analysis to an ionospheric dataset and discovered connections between multi-scale TEC properties and solar influences. Machine learning shows promise in TEC forecasting (Liu et al, 2020). Chen et al (2019) demonstrated a deep learning based approach to fill in missing information in TEC maps.…”