There are many types of power quality disturbances in a power system, and the characteristics of the disturbances are complex. The validity of feature extraction directly affects the recognition accuracy. In order to ensure the validity of feature extraction, the dimension of the eigenvector is usually sacrificed, but the high dimension of the eigenvector will increase the complexity of the recognition model and reduce the speed of recognition. From the above considerations, a combined reconstruction feature extraction method is proposed based on energy entropy and power spectral entropy. First, the power quality disturbance signal is processed according to its characteristics and modified ensemble empirical mode decomposition (MEEMD). Second, the disturbance features are extracted by combining energy entropy and power spectrum entropy, so that a high-precision and low-dimension eigenvector is constructed. Finally, the disturbance signals are identified by a double-layer back propagation neural network (DBPNN). Simulation and experimental results show that, compared with the single feature extraction method, the dimension of the eigenvector, the complexity of the recognition model and the difficulty of recognition are reduced in the combined reconstruction feature extraction method proposed. The accuracy rate is higher and it has a certain degree of anti-noise.