Abstract: Aiming at the problems of complex signal feature extraction, insufficient algorithm recognition and difficult discrimination of composite disturbances in the process of traditional power quality disturbances (PQDs) recognition, a method of composite disturbance recognition based on the Gram angular fields (GAF) and the depth residual network (ResNet) is proposed. Firstly, the one-dimensional time series PQDs signals are standardized and polar coded, and then the dual channel GAF method is used to retain the signal timing features and map them into two-dimensional images to form a two-channel image training set with sufficient information and obvious features. On this basis, ResNet is used for further feature extraction to construct a network framework suitable for the composite PQDs classification. Simulation results show that this method has a strong feature extraction, good anti-noise performance and high recognition rate of composite disturbance.