Research On Fault Prediction And Recognition Of Wind Turbines Based On Data Driven

The wind turbine has the characteristics of high failure rate.In order to improve the reliability of wind turbine operation,this paper takes the SCADA system of the wind power central control center as the data source,takes the wind turbine produced by a company as the research object,and uses the data-driven method to conduct fault prediction and identification research on the wind turbine.This paper provides reference information for the intelligent operation and maintenance of wind farms.The detailed research work of this paper is as follows:Firstly,this paper analyzes the fault phenomenon of wind turbine.Through the wind turbine status code in the SCADA system,the data of the wind turbine in the running state is filtered.Through the database operation,the system abnormal values such as zero and null values are eliminated,and the accumulated abnormal data in the wind power curve is also eliminated.This paper writes a data screening algorithm based on the idea of density clustering,and then removes sparse abnormal data to obtain a structured sample set.Secondly,in view of the high-dimensional and heterogeneous characteristics of SCADA variables in wind turbines,a condition monitoring model for wind turbines based on an improved stacked autoencoder is proposed.By using the denoising autoencoder and the sparse autoencoder as the unsupervised pre-training model of the stacked autoencoder,the feature learning ability of the condition monitoring model is improved.By combining the momentum gradient descent method with the small batch gradient descent method during model training,the training accuracy of the condition monitoring model is improved.By analyzing the operation principle of the wind turbine,the monitoring variables of each main component are selected,and component-level state monitoring models are constructed by using an improved stack self-encoder.Through the condition monitoring model,the wind speed power model and the cabin vibration model are constructed,and the two models and the condition monitoring model of each component are used to correct the index balance and index fusion,and finally the comprehensive index reflecting the condition of the whole wind turbine is obtained.Thirdly,in view of the uncertainty of state information caused by wind and wind power conversion during the operation of wind turbines,a cloud model-based wind turbine fault prediction and identification model is proposed based on the condition monitoring model.Based on the similarity measurement process of the coincidence degree of the cloud model,the comprehensive index of the wind turbine is converted into a probabilistic health index.The health index is further graded through the membership of the cloud model to obtain a qualitative value of the health status of the wind turbine.Through the ratio of each component index to the index sum,the key components that cause the change of the health status of the wind turbine are identified.Finally,through the use of different gradient descent methods in the state monitoring process for comparative experiments,the superiority of the improved gradient descent method is verified.Through comparative experiments on the condition monitoring methods of wind turbines,the superiority of the improved stacked autoencoder for processing SCADA data of wind turbines is verified,and the validity of the condition monitoring model is verified by using fault data.Through the comparison experiment of cloud model similarity measurement methods,the superiority of the proposed method in the classification of cloud models and the time cost of the algorithm is verified,and the validity of the fault prediction and recognition model is verified by using the fault data.