A neural network Dst index model driven by input time histories of the solar wind–magnetosphere interaction

Abstract

A model to forecast 1-hour lead Dst index is proposed. Our approach is based on artificial neural networks (ANN) combined with an analytical model of the solar wind–magnetosphere interaction. Previously, the hourly solar wind parameters have been considered in the analytical model, all of them provided by registration of the ACE satellite. They were the solar wind magnetic field component Bz, velocity V, particle density n and temperature T. The solar wind parameters have been used to compute analytically the discontinuity in magnetic field across the magnetopause, denoted as [Bt]. This quantity has been shown to be important in connection with ground magnetic field variations. The method was published, in which the weighted sum of a sequence of [Bt] was proposed to produce the value of Dst index. The maximum term in the sum, possessing the maximum weight, is the one denoting the contribution of the current state of the near-Earth solar wind. The role of the older states is less important – the weights exponentially decay. Moreover, the terms turn to zero if Bz⪯¡0. In this study, we set up a more comprehensive model on the basis of the ANNs. The model is driven by input time histories of the discontinuity in magnetic field [Bt], which are provided by the analytical model. At the output of such revised model, the Dst index is obtained and compared with the real data records. In this way we replaced those exponential weights in the published method with another set of weights determined by the neural networks. We retrospectively tested our models with real data from solar cycle 23. The ANN approach provided better results than a simple method based on exponentially decaying weights. Moreover, we have shown that our ANN model could be used to predict Dst 1h ahead. We assessed the predictive capability of the model with a set of independent events and found correlation coefficient CC=0.74±0.13 and prediction efficiency PE=0.44±0.15. We also compared our model with the so-called Dst-specification models. In those models, the Dst index was derived directly through an analytic or iterative formula or a neural network-based algorithm. We showed that the performance of our model was comparable to that of Dst-specification models.