Research On Condition Monitoring Method Of Wind Turbine Based On Normal Behavior Modeling

In order to achieve the "3060" goals about emission peak and carbon neutralization,one of the core measures is to continuously promote the use of green energy such as wind energy and solar energy to replace traditional fossil energy.Under the background of the gradual reduction of new energy price subsidies,how to improve the operation and maintenance level of wind farms and strengthen the informatization,automation and intelligence of wind turbine condition monitoring technology has become one of the important research topics in the wind power industry.Using the historical and real-time data of supervisory control and data acquisition(SCADA)system with low acquisition cost and rich types of measuring points as the data source,this paper studies the wind turbine condition monitoring method based on normal behavior modeling.The main research contents of this paper are as follows:1.The relevant information of the normal behavior modeling method are introduced in detail,including three core contents:data processing,data-driven modeling,and residual analysis.Data processing includes abnormal data detection and processing,feature selection,etc.Data-driven modeling is to establish the model that can remember and learn the behavior and characteristics of the monitored equipment during normal operation based on the normal operational data.Residual analysis is to use the residuals between the estimated and the measured value of condition variables to calculate the quantitative condition monitoring indicators,and further calculate the discrete quantity representing the operation condition of the equipment in combination with the warning threshold.2.For the data processing of normal behavior modeling:an abnormal SCADA data detection method based on wind turbine technical parameters,status code,and the statistical characteristics of operational data is proposed,which can effectively detect most abnormal data with low complexity.A feature selection method based on maximum conditional mutual information is proposed,which can effectively measure the redundant information between selected features and the features to be selected,so as to obtain the feature subset with less redundant variables and higher modeling accuracy.3.For the data-driven modeling of normal behavior modeling,multivariate state estimation technology(MSET)is studied,which belongs to the nonparametric method:some properties of matrix DT⊕D in MSET are proved.An improved MSET based on ensemble learning and soft clustering is proposed,which can not only retain the advantage of high estimation accuracy,but also make full use of a large amount of data to train the parameter regression method which is built as the combiner to improve the accuracy,and solve the problem that MSET can not use the data outside the memory matrix.A dynamic MSET based on k-nearest neighbor is proposed,which can online select some data close to the current input data from the training set to construct the dynamic memory matrix.Compared with the fixed memory matrix with the same scale,the accuracy of the proposed algorithm is significantly improved.An improved MSET based on incremental learning is proposed,which can add new normal data to the training set and delete the redundant data online,so that the proposed algorithm can maintain high accuracy and low false alarm rate in long-term operation.Meanwhile,a novel dynamic memory matrix construction method is proposed to improve the online calculation speed.4.For the residual analysis of normal behavior modeling,the residual analysis method based on relative entropy is studied:for continuous relative entropy,a residual analysis method based on Box-Cox transformation is proposed,that is,the residuals are transformed first to improve the normality of data,and then calculate the relative entropy of normal distribution,which can improve the trend and fault sensitivity of condition monitoring curve.For discrete relative entropy,an incremental calculation method suitable for sliding window data is proposed,which significantly reduces the computational complexity and improves the online calculation speed of relative entropy.