Modeling and prediction of TEC based on multivariate analysis and kernel-based extreme learning machine

Radio wave propagation of Global Navigation Satellite System (GNSS) signals via an ionospheric medium offer the opportunity to monitor ionospheric weather nowcasting and forecasting services. The GPS-TEC observations of 20 years are taken into consideration over the Japan region at the GridPoint (134.05 degrees E and 34.95 degrees N). Multivariate Singular Spectrum Analysis (MSSA) is described in this article as a new model for nowcasting and ionospheric prediction. The MSSA algorithm includes a) Time series decomposition, b) reconstruction of approximate components that retain useful components and remove noise components and c) forecast of new data points by a kernel-based extreme learning machine (KELM). An essential modification of MSSA is to maximize the joint variance of all the variables based on Vautard and Ghil (1989) approach. The proposed MSSA achieves high-level now casting illustration at different seasons and solar activities. The first MSSA mode constitutes 99% of the overall variance and characterizes the solar activity variation of the TEC. The RMSE between observed and MSSA model TEC values is 1.52 TECU for the period (1997-2016) and the correlation coefficient is 0.99. Further, MSSA is used as a pre-processing tool for TEC prediction based on KELM. The performance of MSSA-KELM is evaluated in seven cases of different solar periods. The average error measurements during the seven cases are 0.70 (MAE), 5.23 (MAPE), and 0.99 (MSD) respectively. The model achieved higher forecasting accuracy and the lowest training time.