This paper proposed a new method to recognize coherent generator groups using curve similarity in power system. Firstly, the real-time perturbed trajectory of rotor angle is used as an original data, which can be obtained from wide-area measurement system (WAMS). According to the definition of discrete curve similarity, the power angle trajectory of each generator is divided into several sections. The method uses genetic algorithm to determine the best similarity distance between track segments, which can keep the local features of the track segments. Then all segment similarity distances are integrated into overall similarity between generators. The overall similarity is taken as clustering index. Then with the hierarchical clustering method, it can achieve clustering coherent generators in multi machine system. This method is simple and feasible, and is not restricted by the parameters of the system model and the fault type. Finally, the simulation analysis results of IEEE39 node system show that the proposed method is effective.