To address the challenge of identifying power quality disturbances (PQDs) in new-generation power systems, a novel PQDs identification method is proposed, which combines improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), dual time-frequency image fusion, and a DenseNet-CPSAMs deep learning model. First, ICEEMDAN is utilized to decompose the PQDs signals and reconstruct their components. Second, corresponding time-frequency images are generated through the synchroextracting transform (SET) and Stockwell transform (ST), which are fused into a 6-channel input tensor. Finally, a DenseNet-CPSAMs deep learning model is introduced, integrating densely connected convolutional networks (DenseNet), channel attention mechanisms (CAM), and parallel spatial attention mechanisms (PSAMs), to achieve multi-scale time-frequency feature extraction and the enhanced disturbance recognition. Compared to the DenseNet-121 model, the DenseNet-CPSAMs method reduces model parameters by 6.5 M, while achieving an average recognition rate of 99.645% for 31 disturbance types under a 20 dB high signal-to-noise ratio. Simulation results demonstrate that the proposed method exhibits strong feature extraction capability, high noise resistance, and superior recognition performance for composite disturbances.