A machine learning approach for total electron content (TEC) prediction over the northern anomaly crest region in Egypt

Abstract

In this paper, we developed an empirical model for predicting Total Electron Content (TEC) on a single station receiver located within the equatorial anomaly crest region at Helwan – Egypt (29.86°N, 31.32°E & MLAT 27.02°N). The neural network algorithm was developed from several inputs representing diurnal, seasonal, and solar cycle variations, solar and geomagnetic activity, and solar wind parameters using TEC data collected by the SCINDA GPS receiver. The TEC data collected for about nine years from 2009 to 2017 comprises almost the whole solar cycle 24 allowing the developed model to be validated and tested for different solar activity conditions. The 2009 – 2016 dataset was randomly divided into 70% for training, and 30% for validation during model development while the 2017 dataset was reserved for testing the model during independent validation. The validation was done by comparing TEC from the developed model and TEC derived from the GPS scintillation and TEC monitor (GPSTEC). Knowing that TEC modeling over the equatorial anomaly crest regions is challenging since it is characterized by high temporal and spatial variations, the developed model reconstructed diurnal GPSTEC from 2014 to 2016 with a root mean square error (RMSE) of 4.0 TECU and 3.2 TECU for high and low solar activity phases respectively. The neural network (NN) modeled TEC was highly correlated with GPSTEC, that is 0.93 and 0.94 for low and high solar activity respectively. In comparison, the International Reference Ionosphere (IRI)-2016 model predicted the general trend of GPSTEC in 2017 with RMSE of 5.43 TECU better than the developed TEC model with RMSE of 5.74 TECU. However, the NN model predicted diurnal GPSTEC for selected quiet and disturbed days, and seasonal TEC better than the IRI model in 2017. Therefore, the developed model was able to capture temporal ionospheric TEC for diurnal and seasonal variations during low solar activity. It was established that the performance of the empirical model depends on the size and distribution of the data used for model development. It’s worth noting that despite the improved performance of the NN algorithm in predicting diurnal and seasonal TEC variation, the NN-modeled TEC performed poorly in reconstructing the nighttime anomalies exhibited in the GPSTEC. The inclusion of the IRI-NmF2 as network input also adds to the improved performance of the network.