Anomaly Detection And Assessment Method For Operational Condition Of Wind Turbines Based On The Multivariate Data Reconstruction

It has important academic significance and application value to study the condition monitoring and assessment approach for wind turbines(WTs),detect the abnormal condition timely and realize quantitative assessment of operation condition for ensuring operational reliability,reducing the O&M costs,formulating scientific and reasonable O&M strategies of WTs,and promoting utilization of large-scale wind power safely and efficiently.Supervisory Control and Data Acquisition(SCADA)system of a wind farm provides a large number of operation data for WTs.However,most data are from healthy operation condition of WTs,while fault data are scarce and even cannot be obtained.Moreover,when the condition parameters of WTs in SCADA system are alarmed due to over limit,the corresponding faults or anomalies have deteriorated.Therefore,how to utilize the large amount of healthy operation data comprehensively and effectively for realizing the operation condition monitoring of WTs is a difficult problem for all wind power researchers to solve.Current operation condition monitoring methods based on SCADA data for operation conditions analysis and judgement of WTs are realized through the analysis on single or little number of condition parameter,which are difficult to comprehensively reflect the operation conditions of WTs.In view of this,the identification and correction of bad data in the healthy operation data of WTs,the abnormal detection and short-term assessment of operational condition for WTs are studied with data-driven approach in the thesis.The main contents are as follows.Firstly,the identification and correction approaches of bad data under normal operation of WTs were proposed.The causes and distribution characteristics of bad data in SCADA system under normal operation of WTs were analyzed.Corresponding bad data identification methods were proposed for different condition parameters of WTs.A combinatorial prediction approach was employed to correct the bad data.The case study indicated that the interquartile range(IQR)-based method is effectively to identify the bad temperature data,and the adaptive mean shift clustering-based bad data identification method shows a better performance on power data than traditional clustering-based methods such as k-means clustering and density-based spatial clustering of applications with noise(DBSCAN)under low wind speed.The combinatorial prediction approach can effectively improve the correction accuracy compared with single prediction approach.Secondly,the multivariate data reconstruction model for condition parameters of a WT was developed.The cross-correlation between the multivariate parameters and short time autocorrelation of each parameter were analyzed.The sliding window stacked multiple denoising autoencoders(SW-SMDAE)model was established to reconstruct the multivariate data of WTs.The analysis on actual SCADA data shown that the cross-correlations between parameters of same components are strong and those of different components are weak.The time series data of each parameter has different degrees of short time autocorrelation.The cross-correlations between the parameters and autocorrelation in each parameter can be captured by SW-SMDAE model simultaneously.Both of the coarse-grained and fine-grained features of inputs can be obtained by training the model with multiple noise ratios,which effectively improves the feature learning and generalization ability of the model.Thirdly,an anomaly detection method for operational conditions of WTs was presented on multivariate data reconstruction.Mahalanobis distance was extracted as condition monitoring indicator from multivariate reconstruction error.The probability distribution function(PDF)of monitoring indicator of WT normal data was estimated by kernel density estimation(KDE).The PDFs of monitoring indicator under different model parameters were analyzed.The thresholds of monitoring indicator and over-limit duration were obtained according to the PDF of monitoring indicator with normal data.The contribution of each parameter to over-limit of monitoring indicator was proposed to realize the detection of abnormal parameters or components of WTs.The detection accuracy of different widths of sliding window,hidden layers and noise ratios were analyzed with actual SCADA data.The case study indicated that the deep network could acquire high-order characteristics,which is more conducive for determining the thresholds and anomaly detection.Meanwhile,it is necessary to select appropriate layers to prevent overfitting.Sliding window processing and multiple noise ratios training method can effectively improve the detection accuracy.The proposed method can detect the anomalies of actual WTs quickly and accurately by analyzing the multivariate data directly compared with the method by analyzing prediction residuals based on the prediction of condition parameters.Finally,an innovative quantitative evaluation approach for operational conditions of WTs was proposed based on Gaussian mixture model(GMM).The GMM of PDF for WT monitoring indicator was established.Minimum Message Length-Expectation Maximization(MML-EM)algorithm was utilized to select the order of model and corresponding model parameters.The analytic expression of PDF for condition monitoring indicator was obtained.The quantitative evaluation indicator of a WT was proposed.The quantitative evaluation of condition of a WT was realized according to the overlap degree between the PDF of monitoring indicators during assessment period and that of the normal operation.The case study indicated that proposed approach can evaluate the operation condition of a WT quantitatively and track the development trends of operation condition effectively.This work is an active exploration for the research of anomaly detection and evaluation of operation conditions for WTs.It has a good practical application value to improve the operational safety,economy and health management level of WTs and reduce the O&M costs.It provides scientific basis for condition monitoring and assessment as well as intelligent O&M of WTs,and a new way for monitoring and evaluating the operational conditions of other large-scale power equipments.