Neural network prediction of the AE index from the PC index

It is shown that although the power spectra of the AE and PC data are quite similar they show differencies in structure function analysis. While the A E time series has a clear drop in the slope of the structure function (SF) after the first 2 hours, the slope of the SF of the PC data decreases gradu ally and at a little longer time scale. It is also shown by using 15-min av eraged data, that both SFs are periodic with a clear diurnal variation. The PC time series seems to have a more pronounced periodicity, probably becau se it is measured at a single station at Thule. The AE index has been derived from the PC index for 7.5 minutes ahead by di fferent methods. All these predictions gave normalized mean square errors ( NMSE) of the order of 0.22-0.25 during wintertime. The corresponding correl ation coefficient was 0.88 at the best. This shows that the AE time series can be derived quite accurately from the PC data, at least during wintertim e, when the field-aligned current are the main source of the PC index. (C) 1998 Elsevier Science Ltd. All rights reserved.