Prediction of SYM-H index by NARX neural network from IMF and solar wind data

Abstract

SYM-H is one of the important indices for space weather. It indicates the intensity of magnetic storm, similarly to Dst index but with much higher time-resolution. In this paper an artificial neural network (ANN) of Nonlinear Auto Regressive with eXogenous inputs (NARX) has been developed to predict SYM-H index from solar wind and IMF data. In comparison with usual BP and Elman network, the new NRAX model shows much better prediction capability. For 15 testing great storms including 5 super-storms of Min. SYM-H < −200 nT, the cross-correlation of SYM-H indices between NARX network predicted and really observed is 0.91 as a whole. For the 5 individual super-storms, the lowest coefficients is 0.91 relating to the super-storm of March 2001 with Min. SYM-H of −434 nT; while for the two super-storms with Min. SYM-H ranging from −300 nT to −400 nT, the correlations reach as high as 0.93 and 0.96 respectively. The remarkable improvement of the model performance can be attributed to such a key feedback from the network output of SYM-H with a suitable length (about 120 min) to the input, which implies that some information on the quasi real-time ring currents with a proper length of history does its work in the prediction. It tells us that, in addition to the direct driving by solar wind and IMF, the own status of the ring current plays an important role in its evolution especially for recovery phase and must properly be considered in storm-time SYM-H prediction by ANN. The neural network model of NARX developed in this paper provides an effective way to achieve it.