A novel hybrid Machine learning model to forecast ionospheric TEC over Low-latitude GNSS stations

Abstract

Understanding the ionospheric variability caused by the solar and geomagnetic space weather can be achieved using the ionospheric Total Electron Content (TEC) parameter. Developing a novel artificial intelligence-based machine learning model that can forecast the TEC parameter would primarily benefit Global Positioning System (GPS) users. This paper holds one such a novel hybrid empirical model that exploits nonlinear autoregressive neural network with external input (NARX) for a one hour in advance TEC forecasting over the Hyderabad (17.45°N,78.47°E) and Bengaluru (12.95°N, 77.68°E) GPS stations. The 11-year GPS data spans from 2009 to 2019 corresponding to the 24th solar cycle obtained over these GPS stations operated by the International GNSS Service (IGS) center. The performance of this multiple input data model, which accepts TEC, Ap index, Solar flux index, time of the day, geographical coordinates, and others, have been evaluated during both High Solar Activity (HSA) year 2014 and Low Solar Activity (LSA) period 2019. A comprehensive evaluation of the proposed model has been carried out in comparison with the International Reference Ionosphere 2016 (IRI-2016) model, NeQuick-G model, Auto Regressive Moving Average (ARMA), Neural Network (NN), and combined NN with ARMA models. Results reveal that the proposed hybrid ML model could forecast the ionospheric well. The proposed hybrid ML model's root-mean-square error (RMSE) ranges in 0.5–1 TECU during the LSA period and 1.63–2.2 TECU during the HSA period. The proposed model's Maximum Absolute Percentage Error (MAPE) is between 1.6% and 4.5%, with a 0.99 correlation coefficient (R 2 ) value consistently. Thus, the proposed ML model provides promising ionospheric TEC forecasting results.