Prediction of solar activity time series using LSTM artificial neural network

Abstract

A numerical model for predicting the parameters of solar activity — the number of sunspots R and the radioflux at a wave of 10.7 cm F10.7 ahead for 27 days — has been built. The numerical model uses an artificial neural network (NN) with LSTM (Long short-term memory) layers. For both the number of sunspots and the radioflux, the model predicts the levels and limits of variation of the values for 27 days. The average absolute prediction error of the model is less than 6 %. The real-time model is implemented on the site http://aurora.pgia.ru and can be an addition to long-term forecasts of other INTERNET resources.