Prediction of Isolated Substorms by a Package of Parallel Neural Networks

A neural network forecast of substorms caused by the impact of solar wind plasma flows on the Earth’s magnetosphere has been performed. For this, recurrent neural network models were created based on physical cause-and-effect relationships of the dynamics of high-latitude geomagnetic activity (according to the AL index) with the parameters of the interplanetary magnetic field (IMF) and solar wind plasma (SWP). Two parameters are used as input sequences: the bz -component of the IMF and the integral parameter Σ[ NV 2 ], taking into account the prehistory of the process of pumping the kinetic energy of the solar wind into the magnetosphere, where N and V are the plasma density and solar wind velocity, respectively. The forecast of the AL index according to SWP and IMF for 10 min, etc. with 10 min discreteness individually by an individual artificial neural network (ANN) for each point corresponding to the dynamics of the AL index was completed. This means that the prediction of a continuous series of values AL index is achieved by a parallel running of the ANN package. The number of ANNs in the package is determined by the duty cycle of the required predictive series of the AL index, while taking 90 min of the history of input parameters in each of the networks into account provides a prediction of the values AL index with an accuracy of ~80%.