To address the challenges of multi-type composite disturbance feature coupling and the fragmentation between static and dynamic characteristics in single-image representations for PQD (power quality disturbance) identification in new-type power systems, a novel composite PQD identification method based on dual-modal feature fusion and a collaborative attention-driven architecture was proposed. Time-series signals were first encoded into complementary images using GASF (Gramian angular summation fields) and MTF (Markov transition fields), overcoming the information limitations inherent in single-encoding modes. Subsequently, a dual-branch collaborative attention-driven network was designed to perform parallel feature extraction from the two types of encoded images, achieving complementary enhancement and synergistic integration of static trends and dynamic details through dual-modal feature fusion. Furthermore, a multi-SNR mixed training strategy was introduced to enable the model to learn noise-invariant features, thereby significantly improving its robustness under 20~40 dB noise interference. Experimental results demonstrated that the proposed method attained an average classification accuracy of 97.42% on a test set containing 10 single and 22 composite disturbances under 20 dB noise, confirming its high accuracy and strong robustness even in noisy environments.