The power quality problem in the context of the energy internet is becoming more and more prominent. However, there are several problems in the traditional power quality disturbance (PQD) identification process, such as the signal feature extraction is complex, the algorithm recognition ability is insufficient, and it is difficult to differentiate composite disturbance, etc. Thus a new method—Res-GRU-AT, combining hybrid component multi-scale time-frequency diagram, residual neural network (ResNet), gated recurrent units (GRU) network and attention (AT) mechanism, is proposed for power quality composite disturbance identification. First, the PQDs signals are decomposed at multiple scales using singular spectrum decomposition (SSD) and successive variational modal decomposition (SVMD) respectively to obtain the hybrid components. Then the hybrid components are analyzed by Hilbert-Huang transform (HHT) to obtain the multi-scale time-frequency diagram. Secondly, multi-scale time-frequency diagrams are deeply extracted, strengthened, and recognized using the Res-GRU-AT model. The Res-GRU-AT model can perform feature fusion by using the spatial feature extraction capability for 2D images of ResNet and the temporal feature extraction capability of GRU. Then the feature-weighted enhancement is done by AT to improve the recognition capability of PQDs. Simulation results of different schemes show that the proposed method has strong feature extraction capability, good noise immunity, and high recognition rate of composite perturbation.