Abstract: To improve the prediction accuracy of runoff time series this paper proposes a runoff time series prediction method that combines wavelet packet decomposition（WPD）with singular spectrum decomposition（SSA）-rat swarm optimization（RSO）algorithm-echo state network（ESN）. WPD and SSA is used to decompose the non-stationary runoff time series into several sub-sequences to effectively reduce the complexity of the runoff time series;the principle of the RSO algorithm is introduced,6 typical functions are selected under different dimensional conditions to simulate the RSO algorithm;RSO algorithm is used to optimize hyperparameters such as ESN reserve pool size and sparsity,WPD-RSO-ESN,SSA-RSO-ESN models are established,and WPD-RSO-SVM,WPD-ESN,WPD-SVM and SSA-RSO-SVM,SSA-ESN are constructed,and SSA-SVM are used as comparative analysis models;the monthly runoff time series data from 1957 to 2014 at Jiangbian Street Hydrological Station in Yunnan Province are used to test and compare 8 models. The results show that the RSO algorithm has better optimization accuracy and global search ability under different dimensional conditions. The WPD-RSO-ESN and SSA-RSO-ESN models have predicted average absolute percentage errors of monthly runoff time series for 10 years and 120 months after the example. The average absolute percentage errors are 2.73% and 3.90%,respectively. The prediction accuracy is better than other models under the same decomposition conditions. The RSO algorithm can effectively optimize the hyperparameters of the ESN network and improve the prediction performance of the ESN network. The decomposition effect of WPD on runoff time series data is better than SSA method.