Abstract: To address limitations in traditional power quality disturbance identification, such as limited disturbance categories, low noise interference, and missing image features, this paper proposes a multi-channel RES-CBAM model based on combined time-frequency images for comprehensive identification of composite disturbances, considering both multiple disturbance superpositions and high signal-to-noise ratios. Firstly, the multi-channel ResNet18 network is enhanced by integrating a convolutional attention module. Secondly, a time-frequency analysis, encompassing both continuous wavelet transform and synchroextracting transform, is conducted on 21 distinct power quality disturbance (PQD) signals, resulting in the construction of composite image samples that are abundant in feature information. Afterward, these samples are input into the network recognition model to complete the classification of power quality composite disturbances. The proposed method exhibits a remarkable average recognition rate of 99.5983% for 21 types of disturbances, even under a signal-to-noise ratio of 20 dB, demonstrating its robust classification capabilities.