Prediction and Analysis of Sunspot Activity Based on Multivariable LSTM Network

By adding the length of sunspot cycle, the multivariable input data of Long Short-Term Memory (LSTM) is constructed to predict the change of sunspot number in the next ten years with multiple time steps. According to the length of training data, the data set is divided into two groups of time series segments, namely, 11-slice and 6-slice. The prediction effect of univariate and multivariable on single-time step and multi-time step is compared. Finally, the main results are as follows:(1) Compared with the method of 6-slice, the method of 11-slice has lower Root Mean Squared Error (RMSE). (2) The optimal multi-step prediction is better than the single-step. (3) The accuracy of the starting point, ending point, and the maximum amplitude position of the cycle on the legend proves that prediction effect of the multivariate multi-step method is better than that of single-step method.