Prediction and analysis of non-stationary runoff extreme sequence based on ESMD combination prediction model

Abstract With increasingly severe climate changes and intensified human activities, it is more and more difficult to predict the non-stationary extreme runoff series accurately. In this research, based on the ‘decomposition-prediction-reconstruction’ model, an instantaneous frequency distribution map was used to measure the effect of empirical mode decomposition (EMD), ensemble empirical mode decomposition, complete ensemble empirical mode decomposition and extreme-point symmetric mode decomposition (ESMD) in dealing with mode mixing; appropriate prediction methods for each component were selected to form a combined prediction model; and the advantages of a combined prediction model based on ESMD were compared and analyzed with the following results acquired: (1) ESMD can address the mode mixing problem with EMD; (2) particle swarm optimization-least squares support vector machine, autoregressive model (1) and random forest are suitable for high-/medium-/low-frequency components and the residual components R; (3) the results of the combined prediction model are better than those of the single ones; and (4) the prediction effect of the combined prediction model is the best under ESMD decomposition, and the prediction errors of the runoff extreme value sequence can be reduced by about 58–80% compared with the three other decomposition methods. Moreover, as demonstrated in this study, the combined prediction model based on ESMD can effectively predict the non-stationary extreme runoff series, while providing reference for forecasting other non-stationary time series.