Abstract: As the penetration of renewable energy increases rapidly, the power quality disturbance (PQD) is becoming more and more complex, making it difficult for traditional methods to accurately identify the PQD and locate the time interval. To address this problem, this paper proposes a PQD point classification and time interval identification method based on the incorporation of multi-level attention mechanism. The classification model is constructed by using convolutional neural network (CNN) with the local feature attention mechanism (LFAM) and the dual-scale attention mechanism (DSAM). LFAM tracks changes in amplitude by analyzing the envelope and selectively amplifies local features in the signal waveform using weighted techniques. On the other hand, DSAM facilitates the model in identifying the significance of features from both the channel and neuron perspectives. Finally, each sampling point is classified in the form of multiclass-multioutput, based on which the time interval is also identified. To validate the effectiveness of the proposed method, a simulation dataset with 63 PQD types is established. The average classification accuracy of the proposed model is 99.10% in a 30dB white noise environment, and the time-detection errors are all in the millisecond range, which has better generalization performance and robustness than other deep learning models. Additionally, a hardware platform utilizing an AC power supply is developed to assess the performance of the model. The model achieves an average accuracy of 99.03% on this platform, further verifying the reliability of the proposed method.