In response to the problem of feature aliasing in the recognition of composite power quality disturbances (PQD), this paper constructs a frequency decoupling layer based on discrete cosine transform (DCT) as the input layer of the deep network, and a deep frequency attention network (DFAN) suitable for composite PQD recognition is proposed. Firstly, based on DCT theory, a frequency domain decoupling layer is constructed to decompose the PQD signal, reducing feature aliasing between different disturbances. Secondly, PQD components with different frequencies are used as input for multiple channels, and high-dimensional features containing different frequency information are extracted using convolutional layers. Then, attention mechanism is used to adaptively allocate weights for each frequency channel, achieving frequency component selection and noise removal. Finally, the calculation process of the proposed frequency domain decoupling layer is optimized. Some parameters are calculated and saved in the network structure during network construction, which avoids repeated calculations during the forward propagation process of network training and reduces the training time of the network. The simulation experiments show that compared to other common PQD recognition methods, the proposed method can effectively handle the mixed problem of multiple PQDs, and has higher accuracy in identifying composite PQDs.