PQD recognition using two-dimensional time-frequency spectrograms and an improved YOLOv5

As the penetration rate of renewable energy sources increases in new-type power systems， so too does the complexity of the grid structure， leading to more diverse and complex power quality disturbance （PQD）. To accurately identify power quality， a method for PQD image recognition has been proposed， utilizing a two-dimensional time-frequency spectrograms and an improved YOLOv5. Initially， PQD data is projected onto a two-dimensional time-frequency spectrograms using the S-transform. This approach allows for detail-oriented representation of disturbances in terms of time， frequency， and amplitude via imagery. Subsequently， a YOLOv5 training network is constructed that integrates atrous spatial pyramid pooling （ASPP） structure and attention mechanisms. This design broadens the receptive field of the feature map， facilitating a comprehensive extraction of the disturbance image features， and enables PQD classification recognition through image detection methods. Finally， the accuracy and robustness of the disturbance recognition are validated using simulation data. The results evidence that this method offers a high degree of recognition accuracy. Moreover， the integration of the image recognition method enhances the visual representation of the PQD recognition results.