Towards Reliable Ionospheric Total Electron Content Nowcasting

摘要

Robust functioning of the global navigation satellite systems (GNSS) heavily depends upon nonlinear dynamics of ionospheric parameters. Nowcasting the dynamics of ionospheric parameters is challenging problem. Informative parameters selection for the accurate predictive models construction is considered as part of the novel approach based on the machine learning technology. The vertical absolute total electron content (TEC) with a time resolution of 30 minutes is used as experimental data. The data were obtained using phase and pseudo range measurements of TEC at the mid-latitude IRKJ station (52 N, 104 E) for 2014. The results showed that the most informative parameters for TEC nowcasting are the current value of the TEC, the estimated TEC derivative, the local time, the current values of 10.7cm solar flux (F10.7) and symmetric portion of the horizontal component magnetic field (SYM/H), and the exponentially weighted moving averaged TEC values with periods of 12 and 24 hours and SYM/H with periods of 24 and 96 hours, as well as previously received data with somelag, such as the vertical TEC with 12 hours lag, F10.7 with 3 and 15 days lags. The proposed empirical nowcasting models are based on parameters selected by recursive selection of characteristics with determination of their significance using random forest and support vectors methods. Using these important parameters, the linear regression model allows obtaining an estimate on the interval of 4-7 hours with an RMS ~ 4.5 TECU. The machine learning methods such as random forest, support vector method and gradient boosting allow to reduce RMSE to 3-3.5 TECU.