This paper proposes a novel method to identify coherent generator groups using wavelet transform multi-scale space energy distribution feather and improved self-organizing neural networks. Firstly, the identification criteria of coherent generator groups are defined and then the features of the unit power angle rocking curve are extracted using multi-scale spatial energy wavelet distribution method. Furthermore, the time domain, frequency domain and wavelet energy feature vectors are used as inputs of growth-oriented self-organizing neural networks to obtain grouping of different precision by adjusting the threshold λ. Finally, the recognition results on the IEEE-39 bus system, considering the features of only time-frequency domain and both the wavelet energy and time-frequency domain, are compared. The results show that the proposed method taking into account the feathers of both the wavelet energy and time-frequency domain can obtain higher accuracy.