Traditional power quality identification needs to use signal processing technology to extract signal features first, and the existing multi-class and multi-label classification modeling methods do not reflect the label correlation between multiple disturbances and single disturbances well, making the composite disturbance classification robust. The stickiness and noise resistance are not ideal. In response to these problems, a one-dimensional convolutional neural network model based on multi-task learning is proposed to identify various power quality disturbances. This structure removes the signal feature extraction stage of the traditional method, divides the disturbance classification task into four sub-tasks, designs the corresponding label coding scheme, and finally outputs a 10-dimensional label vector to complete the multi-task classification. Simulation results show that the method has good recognition accuracy in each signal-to-noise ratio, which shows that this model has strong robustness and anti-noise ability. At the same time, multi-task classification is more accurate than One-hot multi-class and multi-label classification, indicating the effectiveness of this modeling method.
