Abstract: Aiming at the problem of low accuracy of direct prediction due to the non-linear and unstable characteristics of runoff series,a new method based on double decomposition of runoff time series and least squares support vector machine（LSSVM）model is proposed for monthly runoff prediction. Firstly,the complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN）is used to decompose the original runoff time series into a series of intrinsic mode function（IMF）. Then the technique of wavelet decomposition（WD）is utilized to decompose the high-frequency components to extract the implicit information from the original data more effectively. Each component is taken as the input of the LSSVM prediction model optimized by particle swarm optimization（PSO）. Finally,the prediction results of each component are superimposed and reconstructed to obtain the final result. Taking the monthly runoff at Changshui Hydrological Station in Luohe River Basin as an example,the verification results show that the proposed CEEMDAN-WD-PSO-LSSVM combination model improves the accuracy of runoff prediction better than that of the single model. The double decomposition conducted by CEEMDAN-WD is more powerful to detect the information of complex runoff series,and provide a new approach for forecasting the nonlinear and unstable monthly runoff time series.