This paper presents a new method to identify coherent generator groups in power system based on projection pursuit. Projection pursuit algorithm is introduced to model wide-area measured time series and analyses high-dimensional data in low-dimensional subspace. It could seek and extract key projection vectors reflecting generator coherent features and identify the coherency of generators according to projection directions of generators. The presented technique could realize real-time identification of coherent generators, in which grouping is based on measured data avoiding the impact of model parameters. It proves that the composition of principal components has corresponding relationship with system oscillation mode. Finally, China Southern Power Grid is used as testing system to verify the feasibility and effectiveness of the method.