Classification of solar wind structures and intense geomagnetic storm alarms with self-organizing maps

Intense geomagnetic storms (D_st < ?100 nT) usually occur when a large interplanetary duskward-electric field (with Ey > 5 mV m^?1) lasts for more than 3 h. In this article, a self-organizing map (SOM) neural network is used to recognize different patterns in the temporal variation of hourly averaged Ey data and to predict intense storms. The input parameters of SOM are the hourly averaged Ey data over 3 h. The output layer of the SOM has a total of 400 neurons. The hourly Ey data are calculated from solar wind data, which are provided by NSSDC OMNIWeb and ACE spacecraft and contain information on 143 intense storms and a fair number of moderate storms, weak storms and quiet periods between September 3, 1966 and June 30, 2002. Our results show that SOM is able to classify solar wind structures and therefore to give timely intense storm alarms. In our SOM, 21 neurons out of 400 are identified to be closely associated with the intense storms and they successfully predict 134 intense storms out of the 143 ones selected. In particular, there are 14 neurons for which, if one or more of them are present, the occurrence probability of intense storms is about 90%. In addition, several of these 14 neurons can predict big magnetic storm (D_st  ?180 nT). In summary, our method achieves high accuracy in predicting intense geomagnetic storms and could be applied in space environment prediction.