Neural network prediction of geomagnetic activity: a method using local Hölder exponents

Local scaling and singularity properties of solar wind and geomagnetic time series were analysed using Hölder exponents α. It was shown that in analysed cases due to multifractality of fluctuations α changes from point to point. We argued there exists a peculiar interplay between regularity / irregularity and amplitude characteristics of fluctuations which could be exploited for improvement of predictions of geomagnetic activity. To this end layered backpropagation artificial neural network model with feedback connection was used for the study of the solar wind magnetosphere coupling and prediction of geomagnetic Dst index. The solar wind input was taken from principal component analysis of interplanetary magnetic field, proton density and bulk velocity. Superior network performance was achieved in cases when the information on local Hölder exponents was added to the input layer.