Prediction of ionospheric total electron content data using spatio-temporal residual network

Abstract

Since the ionospheric total electron content (TEC) exhibits complex spatial and temporal behaviour including the seasonal and solar activity dependencies, resulting in large scale differences across different regions of the world, developing an effective deep-learning-based predictive model, capturing the complex behaviour of TEC is much needed. We present a novel deep learning-based model to predict the TEC using the spatio-temporal residual network (ST-ResNet). The TEC data used in the study were obtained from 185 GPS receiver locations, covering the latitude region from 10° N to 80° N and the longitude region from 110° W to 34° E. We considered TEC data of 154 days each corresponding to the solar maximum (2014) and the solar minimum (2020) years for training the network together with the exogeneous interplanetary magnetic field (IMF) data, including the magnetic field components By, Bz, plasma flow speed (Vp), and proton density (Np). A total of 44,352 maps (collected at the rate of 288 maps per day) were pre-processed by grouping them into closeness, period, and trend channels, which associate with the near-time, short-term, and long-term trends in the TEC data, respectively. After training, the model was tested for one day. The objectives of the study are: i) To develop and evaluate the performance of the ST-ResNet model, ii) to assess its performance on TEC data corresponding to mid-latitude region, under different solar conditions and iii) to compare the ST-ResNet model with other predicted models. For the entire study region, the root mean squared error (RMSE) was found to be 3.13 TECU for the solar maximum year and 1.23 TECU for the solar minimum year. Similar trends were observed in the mid-latitude region (30° N-60° N), with RMSE values of 2.31 TECU and 1.58 TECU for the solar maximum and solar minimum years, respectively. ST-ResNet model showed an improved accuracy over the long short-term memory (LSTM) deep learning network by about 23% and 31% for the entire study region and by 49% and 19% for only mid-latitude region, during the solar maximum and minimum years respectively, suggesting the superior performance of ST-ResNet over the LSTM network. A comparative observation of results of ST-ResNet model with those of back propagation and international reference ionosphere-2016 (IRI-2016) model also showed the superiority of the former. Results confirm the suitability of the ST-ResNet model for predicting the spatial and temporal behavior of TEC across different locations during various solar conditions.