With the rapid transformation of high-carbon power systems to new power systems, the large scale integration of renewable energy sources such as wind and solar, along with the widespread use of power electronic devices, has led to increasing complex and variable power quality disturbances (PQDs). To quickly and accurately detect and capture PQDs and to overcome limitations of traditional disturbance identification methods, such as reduced applicability to complex hybrid PQDs and difficulty in manually selecting features, this paper propose a novel approach for PQDs detection and identification. The proposed approach first employs peak difference guided local difference accumulation to rapidly and accurately detect and capture the PQDs. Then, the improved iterative adaptive kernel regression (IIAKR) method is used for preprocess the captured noisy PQDs, effectively suppressing noise while preserving detailed disturbance features. Finally, the improved visual trajectory circle (IVTC) method transforms the 1-D PQDs into 2-D trajectory circle images with more prominent shape and easier identification features, which are then input to convolutional neural networks (CNN) for autonomous feature extraction and classification. Experimental results show that the proposed approach offers strong noise immunity, high detection rate and classification accuracy for both single and complex PQDs.