Abstract: Distributed power supply can generate complex power quality disturbances (PQDs) when it is connected to the power grid. In order to improve the accuracy of PQDs signal classification and identification, the Grouping Convolutional Neural Network with Adaptive Feature Enhanced Network (GCNN-AFEN) is constructed. The core of GCNN-AFEN model is as follows: First, S transformation is performed on PQDs signals to form time-frequency matrix images, and CNN is combined with sparse GCNN as the basic framework of feature learning to reduce model parameters and improve the computing speed. Then, through the channel attention mechanism, frequency domain feature enhancement and soft threshold denoising, the AEFN module adaptively learns the correlation between disturbance types and corresponding feature graphs, increases the signal-to-noise ratio, and highlights the deep features that can represent disturbance categories. Finally, full connection layer (FC) and Softmax classifier are used for classification recognition. Simulation results show that the proposed model has high classification accuracy and noise robustness for all kinds of power quality disturbance signals, and can be used for fast identification and classification of power quality disturbances.