Wind Turbine Power Curve Abnormal Feature Detection Algorithm based on Improved YOLOv5

The traditional screening of abnormal information in wind turbine power curves mainly relies on personnel with specialized knowledge backgrounds, and this screening method could be more efficient and of varying accuracy. Solving this problem involves proposing an enhanced YOLOv5 target detection model in this paper to identify abnormal features in the wind turbine power curves. Subsequently, it automatically screens fault information in wind turbine power curves, thereby improving the swift localization of problematic units. This study collects operational data from SCADA systems of wind farms nationwide. It employs various processing methods, such as data visualization and anomaly feature annotation, to generate a dataset of anomalous features in the power curves of wind turbines. Regarding enhancement, this paper initially integrates the SE attention mechanism into the final layer of the YOLOv5 backbone network to augment the network's capacity for capturing crucial features. Following this, we substitute the CIoU in the localization loss function with the SIoU function and introduce an angular loss to enhance regression accuracy. On the homemade dataset, the experimentally refined model demonstrates a 3.9% improvement in mAP compared to the conventional model. It also exhibits favorable results in real-world application scenarios, markedly enhancing the efficiency of problematic unit detection and offering substantial support for the intelligentization process of wind power stations.