Research And Application Of Wind Turbine Fault Warning Method

With the increase of the running time of wind turbine,the power generation efficiency gradually decreases and the failure frequency rises,which causes economic losses to wind farms.Therefore,how to effectively mine the data value of the current wind power system and carry out the fault early warning of wind turbines has certain practical significance and economic value for the improvement of operation and maintenance efficiency and the reduction of power loss caused by faults.Because most of the Data collected by the wind farm Supervisory Control and Data Acquisition(SCADA)system is unlabeled,it is difficult to model specific failure patterns.In this paper,the typical faults of the main parts of wind turbine are analyzed.Based on the data,the health value and warning level of the fault warning under a specific fixed collection period are determined and applied.(1)Based on word frequency-inverse text Frequency algorithm(Term frequency-inverse Document Frequency,TF-IDF),the early-warning model of temperature anomaly and control anomaly of variable paddle motor was studied and designed.First,SCADA is used to collect the historical normal operation data of 1/2/3 temperature and 1/2/3 speed of the cleaned variable-paddle motors,as well as the normal baseline of off-line training TF-IDF.Then,the actual baseline of TF-IDF was constructed by using the actual operation data under the fixed acquisition cycle.By analyzing the Mahalanobis distance deviation between the normal baseline and the actual baseline,the health value and warning level of the early-warning model of the temperature anomaly of the variable propeller motor and the abnormal control fault under the corresponding acquisition cycle were determined.(2)Based on multiple linear regression,the early-warning models of gearbox bearing temperature anomaly,gearbox oil temperature anomaly,generator bearing temperature anomaly and generator stator winding temperature anomaly are studied and designed.Firstly,the change mechanism of temperature objects is analyzed and studied.By introducing the first law of thermodynamics and the law of conservation of energy,the linear regression algorithm is used to establish the temperature prediction model based on the historical normal operating state.Then,based on the temperature prediction model,the temperature residuals of the actual operating data under the fixed acquisition period were calculated.Finally,the health values and warning levels of the fault warning models such as abnormal gearbox bearing temperature,abnormal gear oil temperature,abnormal generator stator winding and abnormal generator bearing temperature were determined by residual analysis.(3)The blade icing fault warning model was studied and designed based on GBR(Gradient Boosting Regression)and XGBoost(e Xtreme Gradient Boosting).According to the characteristics that the active power output changes significantly when the aerodynamic airfoil is changed by blade icing,the wind speed-active power regression model based on the historical normal operating state is established using the GBR algorithm.Then,based on the regression model,the predicted active power corresponding to the actual wind speed under the fixed acquisition period is calculated.According to the mechanism of blade icing,ice labeling was carried out on the characteristic data of the fault warning model,and combining with Bagging integrated learning method,binary probability prediction based on XGBoost algorithm was carried out,so as to judge the health value and warning level of blade icing fault warning under the corresponding collection period.