Abstract: This paper proposed a novel method to predict non-stationary time series,based on the empirical mode decomposition and fuzzy support vector machine.Firstly,uisng EMD,the non-stationary time series are decomposed into single modal components,reducing the prediction signal nonlinear complexity.Then,using the fuzzy support vector machine,each intrinsic mode function is predicted.Finally,the results predicted by each intrinsic mode function are superimposed to obtain the final forecast.Using Lorenz and sunspot month smooth value sequence with noise as the experimental data,our method was compared with BP neural network prediction and SVM prediction method by experiments.And this method has stronger adaptability to the sequence signal with isolated points and noise,and better prediction accuracy.

Key words: Non-stationary time series,Empirical mode decomposition,Fuzzy support vector machine,Combination forecast