Modelling of magnetic storms count data using nonlinear poisson regression

Abstract

© 2019 IEEE. The paper proposes a prediction model of dynamics of time series of magnetic storms number. The model used nonlinear Poisson regression. The investigated time series were converted from Dst index data for the 1964-2018 time interval. An artificial neural network was used to build a nonlinear autoregressive model. The training procedures were adapted using statistical properties of the investigated time series. It is shown that fluctuations of the number of geomagnetic storms are close to the Poisson distribution. Thus, to estimate the prediction efficiency, we proposed a special quality measure based on the analysis of the standard deviation ratio of the estimated model parameters. The model was used to forecast the number of magnetic storms for a week in advance. It was shown that the prediction accuracy was 20% higher compared to the traditional approaches to training of artificial neural network systems. A similar approach can also be successfully used to forecast dynamics of rare events number in atmospheric and solar-terrestrial physics.