Abstract: In the process of power quality disturbance recognition, the redundancy existing in feature extraction of the disturbance signal may cause complex structure of the recognition model, difficult training and low recognition accuracy. To solve the above problems, this paper proposes a new method of identification and classification based on the complementary ensemble empirical mode decomposition method(CEEMD) and double-layer back propagation neural network (DBPNN). Firstly, an intrinsic mode function (IMF) is obtained by using CEEMD. Secondly, the Pearson correlation coefficient and energy-entropy are utilized to determine the difference and distribution of the disturbance signal in different frequency bands, and the IMF is selected to reduce redundancy based on this kind of differences. Finally, according to the frequency characteristics of different disturbances, a special purpose DBPNN is established, and the frequency band and disturbance type of the disturbance signal are identified, respectively. Simulation results show that, compared with the traditional power quality disturbance identification method, the feature vector redundancy of the proposed method is lower, and the DBPNN recognition model has lower complexity exercise and shorter training time. The average recognition rate of CEEMD-DBPNN can reach 97.78%, and the accuracy is improved by 7.52% compared with the result of traditional neural network and improved by 9.23% compared with the result of support vector machine.