Research On Diagnosis And Prediction Method Of Wind Turbine Blade Icing Based By Data Driven

The traditional energy crisis has made many countries pay more attention to the development of non-polluting renewable energy such as wind power.In order to obtain wind energy more efficiently,wind turbines are also increasingly installed in high-altitude areas or sea areas with abundant wind energy resources,but high-altitude areas are closer to wet clouds,excessive humidity on the sea surface,etc.Which makes the failure of wind turbine blade icing more frequent and has not been solved very well.This problem is also obvious in China,because the northern region where the wind power industry is the most concentrated area and is also the most vulnerable area to wind turbine blade icing failure.Based on the monitoring data of SCADA(Supervisory Control And Data Acquisition)of real wind turbines,this paper uses data-driven method to diagnose and predict wind turbine blade icing.The main research work is as follows:(1)In view of the different correlation degree between various monitoring data and wind turbine blade icing,three kinds of features with the highest contribution rate are selected from 26 kinds of characteristic data by dynamic PCA algorithm as the key features of wind turbine blade icing fault diagnosis and prediction.Combining with industry experience and data visualization,8 kinds of features with the lowest correlation degree are eliminated,and various monitoring data and wind turbine blade icing are quantitatively analyzed.The characteristic relationship between the faults leads to some data characteristics which are most relevant to the icing of the wind turbine blades and are conducive to the subsequent analysis.(2)Due to the large amount and fast change of SCADA monitoring data of wind turbines,the unbalanced proportion of data in various working conditions makes it difficult to process fault feature data.In view of the insufficiency of SMOTE(Synthetic Minority Oversampling Technology)in balancing data without considering boundary conditions,an improved over-sampling balancing algorithm SC-SMOTE(Safe Circle Oversampling Technology)is proposed,optimizing the data set of working conditions.Based on the data set processed by SC-SMOTE algorithm,a fault diagnosis method for wind turbine blade icing based on k-Nearest Neighbors classification is proposed.Compared with the SMOTE algorithm,the experimental results show that the method is effective in the diagnosis of wind turbine blade icing fault and improves the accuracy of diagnosis.(3)In view of the difficulty in predicting the icing fault of wind turbine blades,a new method for predicting the icing fault of wind turbine blades based on a variety of machine learning algorithms is proposed.Firstly,based on the above-mentioned features which are most relevant to the icing of fan blades,the prediction experiments are carried out by Elman artificial neural network and SVR(Support Vector Machines for Regression).The predicted characteristic values(trends)of various features in the future are input into BP(Back Propagation)artificial neural network trained with a large number of label data.The network is classified and judged to predict whether the wind turbine blades will freeze in the future.Finally,the experimental results show that the prediction error of Elman artificial neural network is smaller and the method is effective.