Gearbox Oil Temperature Prediction for 1.5 MW Wind Turbine Based on Stacking Model

• Published in: 2024 8th International Conference on Power Energy Systems and Applications (ICoPESA) pp. 223 - 229
• Main Authors: Ming, Guangqiang, Cao, Heng, Chen, Fuxin, Ma, Shihao, Wang, Yicheng, Luan, Zeyu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.06.2024
• Subjects: Clustering algorithms; data mining; Data preprocessing; DBSCAN algorithm; grid search method; Machine learning algorithms; Oils; Prediction algorithms; Predictive models; SCADA systems; Stacking; stacking strategy; Wind speed; Wind turbines
• DOI: 10.1109/ICOPESA61191.2024.10743349

Wind turbines are complex systems integrating machinery, electrical, and control, and they often operate in places with harsh environments. This can easily cause the wind turbine to malfunction or even shut down, resulting in significant economic losses and safety accidents. In recent years, big data technology and machine learning algorithms have been constantly tried to be applied to fault diagnosis and prognosis of wind turbines, especially gearbox. In this paper, the oil temperature predicted model of the gearbox is constructed to detect anomalies. First, the wind speed and power in the supervisory control and data acquisition (SCADA) system are used to remove anomaly data points using the density-based spatial clustering of applications with noise (DBSCAN) algorithm, and the grid search method is used to optimize the two input parameters of the DBSCAN algorithm, and then the quartile method is used for further data preprocessing. Second, filter algorithm is conducted to determine the significant features, and these selected features are used to build the predicted model. Third, a fusion model based on the stacking strategy in ensemble learning is established to characterize the normal working status of the gearbox in wind turbines. The application of the presented method to gearbox oil temperature prediction is evaluated by a real-world gearbox normal process. The experimental results show that the proposed method embraces a good ability to predict.