Physics of auroral phenomena : proceedings of the 40th annual seminar, Apatity, 13-17 March, 2017 / [ed. board: N. V. Semenova, A. G. Yahnin]. - Апатиты : Издательство Кольского научного центра РАН, 2017. - 143 с. : ил., табл.

“Physics o f Auroral Phenomena”, Proc. XL Annual Seminar, Apatity, pp. 74-75, 2017 © Polar Geophysical Institute, 2017 h t/jlu Geophysical \ J J y Institute THE CAUSAL RELATIONSHIP BETWEEN THE DYNAMICS OF HIGH-LATITUDE GEOMAGNETIC ACTIVITY AND TYPE OF SOLAR WIND MAGNETIC CLOUD N.A. Barkhatov, S.E. Revunov, Yu.A. Glavatskij Nizhni Novgorod State Pedagogical University, Nizhni Novgorod, Russia Abstract. By the neural network (ANN) method establishes cause-effect relationships of dynamics of high-latitude geomagnetic activity (for AL index) with the type o f Solar wind magnetic cloud caused by the parameters of magnetized plasma coronal ejection. As a tool for analyzing nonlinear dependencies a recurrent neural network of Elman type is used. The neural network model is based on the search for optimal physically connected input and output parameters characterizing the effect on the magnetosphere of a specific plasma stream such as a magnetic cloud. The success o f restoring the dynamics of AL on the data used as the established nonlinear AL connection with the parameters of the cloud is characterized. Interplanetary magnetic clouds (IMC) as the studied fluxes like most geoeffective coronal formations are chosen. There is a variety o f configurations ofmagnetic clouds and methods of their influence on the terrestrial magnetosphere, which depends, among other things, on the impact parameters of the cloud. However, a required feature is the rotation of the interplanetary magnetic field (IMF) vector inside the cloud, which ensures the appearance of a geoeffective negative Bz component. The latter, however, does not mean that IMCs always was cause for global magnetic storms, but they often include substorm processes [1]. The structure o f fast magnetic clouds is noticeably complicated by the appearance o f a shock wave and a turbulent region behind it. In connection with this, it is of interest for the degree of participation of the elements of the structure of magnetic clouds in the formation of high-latitude geomagnetic activity to establish. In this study, as in [2], we apply a neural network approach using a recurrent neural network o f the Elman type. As before, we propose the creation o f a fast neural network model based on the search for optimal physically connected input and output parameters characterizing the effect on the magnetosphere of a specific plasma flux, depending on the type o f magnetic cloud. However, the feature of this research is the use o f different neural network architectures is presented. The study using minute data corresponding to the observation intervals of 52 interplanetary magnetic clouds recorded in 1998-2012 was performed. The parameters o f Solar wind were analyzed for each IMC interval: the concentration N and the plasma velocity V and the components of the vector В (Bx, By, Bz) of the interplanetary magnetic field (IMF) in the GSM coordinate system, as well as the Dst and AL values o f magnetic activity indices. All data with a 1-min resolution from http://cdaweb.gsfc.nasa.gov is taken. The analyzed IMC intervals into two samples: group 1 - fast clouds with shock waves and a turbulent region (33 cases) and group 2 - slow clouds without shock waves and turbulent regions (19 cases) were divided. In addition, the data intervals (group 3 - 7 0 cases) corresponding to isolated magnetospheric substorms according to indications of the AL index without specifying the type of plasma flow, but certainly not associated with magnetic clouds were analyzed. The performed neural network experiments to the search for optimal physically connected input and output parameters characterizing the effect on the magnetosphere of the magnetic clouds under consideration are devoted. In this case characteristic times o f the necessary prehistory o f dynamics o f the cloud parameters for launching substorms were determined. In numerical experiments, a neural network with an external feedback loop was used. This architecture allows to reinforce learning by synthesized within ANN sequences o f the AL-index (Fig. 1). Inputs x and z allow to model two different depths o f the prehistory (H and P). Under the depth of prehistory is meant an additional number of parameters at the input o f the neural network, simulating a time delay. The external feedback loop is shown in bold lines, on it the sequences y*(t), fed to the main input, are synthesized. At the input z, the depth of the prehistory is P=60 minutes. Such a delay on the outer loop was chosen on the basis of studying the effect of Solar wind energy storage to provide a substorms process [4]. At the input x, the depth of the prehistory H can vary. The only one output neuron у generates a sequence o f AL index values. The search for the optimal depth of the prehistory at input x for the input sequences of IMF components from 30 to 90 minutes in 10 minute increments was carried out. An objective assessment o f the quality o f the recovery o f the AL index was carried out by calculating the classical correlation coefficient R and the efficiency of PE recovery [3] between the real (target) and neural network generated values. As a result, it was found that 90 minutes of the background prehistory o f the By and Bz dynamics of the IMF components proved to be most effective for modeling sequences of AL values. The numerical experiments performed with ANN for group 1 produced average values o f R = 0.80, PE = 77% and for group 2 - R = 0.92, PE = 81%. 74