Type recognition and parameter identification of power quality disturbances using improved 1D-YOLO and multi-task learning

Accurate recognition and parameter identification of Power Quality Disturbances (PQDs) are crucial for effective disturbance management. This research presents a model built on an improved One-Dimensional You Only Look Once (1D-YOLO) framework, incorporating multi-task learning to unify three tasks: type recognition, time localization, and characteristic parameter identification. The proposed model utilizes a Path Aggregation Network (PANet) for advanced multi-scale feature fusion, while decoupled heads enable concurrent analysis of 10 PQD elements. The case study results indicate that the model achieves over 99.5% mean average precision and classification accuracy for both element-level and signal-level recognition tasks. Additionally, it precisely identifies starting-ending times, as well as key parameters, with multi-task learning further enhancing the performance of individual tasks.