Research On Generalized Model For Anomaly Detection And Operational Risk Assessment Approach Of Wind Turbines

Influenced by harsh natural environment and random fluctuations of wind energy,wind turbines(WTs)suffer operation conditions with complicate variation characteristics.Because operation conditions influence on WT condition parameters significantly,traditional condition assessment approaches based on thresholding are so difficult to be used in practical applications.In order to improve the operational reliability and safety of WTs,the work focuses on anomaly detectionand operational risk assessment of WTs.The main contents are shown as follows.Firstly,the method for development of the WT condition parameter prediction models was presented.Based on the correlation between the condition parameters and the envioronmental parameters,the condition parameter classification method was proposed.The back-propagation neural network(BPNN)based prediciton models were developed for the envioronmentally sensitive parameters.The performance of the BPNN model was compared with the least squares support vector machines(LS-SVM)based prediction model and the nonlinear state estimation technique(NSET)based prediction model.The performance of the predicition model trained by different type of sample,such as the WT's current data,the WT's historical data and the other similar WTs' data were compared.The quantification method of the sample similarity degree was studied.The influence of the sample distribution of the target condition parameters on the prediction accuracy was studied.Secondly,the generalized model for WT anomaly detection was presented.The WT classifition method was studied based on the statistical distribution of the target condition parameters of the sample data.The prediction performance was used to quantify the effectiveless of anomaly detection for the developed prediciton models.A fast prediction model selection method was proposed for the prediciton models trained by the WT's historical data and the other similar WT's data.The probability distribution functions(PDFs)of the prediction error were developed based on kernal density method.A novel index called abnormal degree index was proposed based on the prediction error distribution.Finally,the fuzzy systhetic evaluation method was used to detect the WT anomalies.The case study indicated that the proposed WT anomaly detection method could integrate the anomaly detection results of multiple prediction models,which shown a better performance to detect the WT anomalies.Thirdly,the WT short-term reliability prediction model was presented.Based the anomaly detection results,the WT conditions were clarified into three categories.The Markov model was used to describe the transition process of the WT condtions.The calculation method for the time varing conditon transition probability was proposed.For the evnironmentally senesitive parameters,the parameter exceedance protection operation model was developed based on the distribution of the condition parameter prediction error.For other condition parameters,theparameter exceedance protection operation model was developed based on the seting time of the protection relay.Finally,by integrating the abnormal information of WTs,the protection relay operation probability and the statistical outage probability of each condition parameter,a WT short-term reliability prediction model was established.The case study shown that the prediction accuracy of the proposed WT short-term reliability prediction model was much higher than thatonly based on the statistical data of WT outages.Finally,the WT operational risk assessment method was studied.Based on the analysis on the failure rate and maintence costs of the WT components,the WT fault classification method was proposed.The relationship between the WT faults and the conditon parameters was studied and the fault chracteristices indices were developed.By using the the fault chracteristices indices as the input parameters,the fault prediciton model of WTs were developed based on the particle swarm optimization(PSO)based LS-SVM.The accuracy of the proposed prediciton model was compared with the BPNN based,the radial basis function neural network(RBFNN)based,and the SVM based model.Finally,the WT operational risk assessment method was proposed based on the predicted WT short-term outage probability and the fault prediction results.The case study shown that the propoesed method was effctive in assessing the operational risk of WTs.The work is an active exploration for research on smart operation and maintenance of WTs and possesses value in practice application to improve the operational reliability and to reduce maintenance costs of WTs.It could provide the foundation and technical reference for thesecurity assessment of wind power collection systems.