Research On Fault Sample Generation Method Of Wind Turbine Generator Based On Generative Adversarial Nets (GAN)

In the recent years,the energy crisis has intensified and climate change has become more and more obvious.Wind energy,as a clean and renewable energy source,its development and utilization are increasingly valued by countries all over the world.With the continuous maturity of wind power technology,the number of global wind turbines is constantly increasing.Since the downtime of wind turbines will bring serious economic losses,the operation and maintenance of wind turbines is particularly important.In order to reduce the losses caused by wind turbine faults,numerous diagnostic methods have been applied into fault diagnosis.Due to the scale of the fault samples collected in the wind turbine is small,it is difficult to meet the data volume requirements of machine learning and deep learning algorithms,and the small amount of data will result in the accuracy of fault diagnosis no longer improving In order to solve such problems,this thesis proposes a method of wind turbine fault samples generation based on Generative Adversarial nets.The data used in the experiment comes from the SCADA system.The following work is basing on the characteristics of the data and can be summarized as follows,SCADA data preprocessing,feature extraction of wind turbine faults,and the utilization of the Generative Adversarial nets.Firstly,in case of the data collected by the SCADA system in the wind farm contains a lot of noise and some of them is invalid,the preprocessing flow of data cleaning,data transformation and data screening is designed.In the process of data cleaning,the invalidation and irrelevant attributes are cleaned,important data sample information is retained in the data transformation process.After that,we select the data between the cut in and cut out wind speed.Secondly,a method of generating rough fault samples is proposed.Firstly,the fault sample type is set,and the fault tree model is established.According to this specific type,the fault tree is used to determine the sub-events that cause the faults,and the variables related to the faults are obtained.Basing on this,a correlation analysis is performed to obtain the main attribute related to the variable.According to the different conditions,two types of training samples were designed:data samples of normal condition of the WT and rough fault samples.For the rough fault samples,the primary attribute is adjusted numerically,and the related attribute are deleted.Thirdly,we introduce the basic principles of generative Adversarial Nets and design the wind turbine’s GAN model.Finally,putting the missing normal data into the Generative Adversarial network.The internal data distribution of the wind turbine can be studied by generator.After the generative network can generate the data,putting the rough fault sample into the generator to get the optimized fault sample.For the generated fault samples,the experimental part of this thesis carried which using the generated fault samples with the real fault samples.We compare the variance,standard deviation,and the marginal distribution of the data.From the above several indicators,it is obviously showed that the generated fault sample is very close to the real fault sample,and the generator fault sample works well.Then we trained wind turbine condition classifier using the generated fault data.It is clearly showed that as the number of training samples increases,the accuracy of the fault classifier continues increasing.