Comparative Efficiency of Prediction of Relativistic Electron Flux in the Near-Earth Space Using Various Machine Learning Methods

Abstract

This paper performs comparative analysis of the prediction quality for the time series of hourly average flux of relativistic electrons (E > 2 meV) in the near-Earth space 1 to 24 h ahead by various machine learning methods. As inputs, all the predictive models used hourly average values of the parameters of solar wind and interplanetary magnetic field measured at the Lagrange point between the Sun and the Earth, the values of Dst and Kp geomagnetic indexes, and the values of the flux of relativistic electrons itself. The machine learning methods used for prediction were the multi-layer perceptron (MLP) type artificial neural networks, the decision tree (random forest) method, and gradient boosting. A comparison of the quality indicators of short-term forecasts with a horizon of one to 24 h showed that the best results were demonstrated by the MLP. The horizon of satisfactory forecast accuracy on independent data is 9 h, the horizon of acceptable accuracy is 12 h.