Abstract: In order to find the ensemble prediction method that can be used to forecast the tracks of tropical cyclones in the Western North Pacific Ocean, 52 cases tropical cyclone track ensemble prediction experiments are made, the 52 cases are from the 8 tropical cyclones that made landfall at China in 2005.MM5 model is used as the experiment model. The model has a horizontal grid spacing of 45 km with 115×115 points and 23 vertical sigma layers. It is run for 72 h. The model domain center is same as the center of experiment tropical cyclone. There are 12 ensemble members. 12 members consist of 11 perturbation members and a control forecast. Two methods are used to create the perturbation members. One method is breeding of growing modes (BGM), in which two 12-h breeding cycles are carried out. The other method is model physics perturbation (MPP), in which members are created by choosing different physics parameterization schemes. The experiment results show that the ensemble mean of BGM is better than the control forecast and the ensemble mean of MPP is worse than the control forecast. For BGM, the ensemble mean for the tropical cyclones whose initial intensity is smaller than 32.6 m/s is more skillful than the ensemble mean for the tropical cyclones whose initial intensity is larger than 32.6 m/s. For MPP, the ensemble mean for the tropical cyclones whose initial intensity is larger than 32.6 m/s is better than the control forecast. According to the different results of BGM ensemble and MPP ensemble, a new method is used to create the perturbation members. In this new method, when the initial intensity of tropical cyclone is smaller than 32.6 m/s, BGM method is used to create the perturbation members, when the initial intensity of tropical cyclone is larger than 32.6 m/s, both BGM method and MPP method are used to create the perturbation members. The ensemble mean of this method is better than the ensemble mean of BGM or MPP. The spreads of BGM ensemble, MPP ensemble and the new method are all too small.