Prediction of Ionospheric TEC Based on the NARX Neural Network

Abstract

Effective prediction of ionospheric total electron content (TEC) is very important for Global Navigation Satellite System (GNSS) positioning and other related applications. This paper proposes an ionospheric TEC prediction method using the nonlinear autoregressive with exogenous input (NARX) neural network, which uses previous TEC data and external time parameter inputs to establish a TEC prediction model. During the years of different solar activities, 12 datasets of 3 stations with different latitudes are used for experiments. Each dataset uses the first 120 days for training and the next 20 days for testing. For each test dataset, a sliding window strategy is adopted in the prediction process, wherein the TEC of future 2 days are predicted by the true TEC values of the previous 2 days. The results show that in the year with active solar activity (2011), the TEC prediction with the NARX network can improve the accuracy by 32.3% and 43.5%, compared with the autoregressive integrated moving average (ARIMA) model and the 2-day predicted TEC product, named C2PG. While in the year with calm solar activity (2017), the prediction accuracy can be improved by 20.7% and 22.7%.

Prediction of ionospheric total electron content data using NARX neural network model

“…where local time, ℎ ranges from 0-23. The sin and cos functions help in normalizing the time input in the range (-1,1) [1]. -Solar activity: solar activity significantly influences TEC values, with lower TEC during solar minimum years and higher TEC during solar maximum years [21].…”

“…This influence is directly proportional to the density of electrons between the satellite transmitter and the receiver which quantifies the total electron content (TEC). Ionospheric TEC exhibits complex spatial and temporal fluctuations along with certain solar and geomagnetic related ionospheric disturbances [1]. The daily diurnal variations, seasonal variations, 27-day variations, 11-year solar cycle variations, geographic latitude, and some events like geomagnetic storms which generate sudden disturbances in the ionosphere are some obvious variations of ionospheric TEC [2].…”

“…The ability to predict symmetric horizontal (SYM-H) storm time index using NARX showed better accuracy compared to back propagation and recurrent Elman networks [18]. NARX has also been used for TEC prediction using time, day of the year and season as exogenous inputs [1].…”

“…We have studied TEC prediction at different latitudinal regions namely equatorial (0 0 -10 0 N), low-latitude (10 0 -30 0 N), mid-latitude (30 0 -60 0 N), and high-latitude (60 0 -80 0 N), during both the minimum solar activity year (2008) and maximum solar activity year (2014). We compare the results of our NARX model (henceforth called NARX1) with the NARX model developed for TEC prediction by Guoyan et al [1] (henceforth called NARX2).…”

Comparison of Two Neural Network Architectures for Local Prediction of Ionospheric Total Electron Content

“…Such approaches employ a machine learning algorithm feed by historical data that "learn" from the data, thus creating a model able to perform a TEC prediction from a time series of precedent TEC values. Considering the same articles, 2/3 refer to local TEC prediction (like Guoyan et al, 2021or Ruwali et al, 2020, and the remaining, to TEC map prediction (like Li et al, 2020). Considering machine learning, half use Multilayer Perceptron (MLP) neural networks, and the remaining, mostly Recurrent and/or Convolutional Neural Networks (RNN/CNN).…”