Neural network-based nonlinear prediction of magnetic storms

The best known manifestations of the solar wind impact on the Earth's magnetosphere are the geomagnetic storms. At middle latitudes, the magnetic storm is best described in terms of the horizontal component of the geomagnetic field. We use the method of neural networks which is suitable for nonlinear dynamical systems modeling and for the nowcasting as well as forecasting of the magnetic storms. For constructing a neural network model, a multivariate method for determination of the inputs is applied first. This method enables us to reduce the original 17 input variables to two variables, the so-called principal components. The performance of the model is characterized by the correlation coefficient rho. Its mean value is 0.93 considering two principal components and time history 6 h, Our interest is also focused on the more than I h ahead forecasting of the geomagnetic index D-st. Another question investigated here is how the inclusion of the history of D-st index into the input matrix influences the predicting ability of the network. (C) 2002 Elsevier Science Ltd. All rights reserved.