Research On Image-based Wind Power Curve Abnormal Data Cleaning And Recognition Method

As a kind of renewable energy,the large-scale deployment of wind energy in the energy and electrical industries has prompted researchers to study data-driven wind power monitoring and control systems.At present,monitoring and data acquisition systems are widely used in data collection in application scenarios such as wind turbine performance evaluation,condition monitoring,and anomaly detection.Among them,the wind power curve describes the relationship between wind speed and power,which is an important technical indicator for the analysis of the performance and working conditions of the wind turbine.However,in the actual environment,the wind power curve data contains different types of abnormal data due to changes in wind speed and direction,which will adversely affect the analysis of the operation status and characteristics of the wind turbine.Therefore,the identification and cleaning of abnormal operation data of wind turbines is a current research hotspot.The research work of this paper mainly focuses on the subject of abnormal data cleaning of wind power curves,and proposes a method for cleaning and identifying abnormal data of wind power curves based on image thresholding.The main research content is divided into the following parts:(1)By establishing a mapping between continuous data space and discrete image space,the wind power curve data is converted into binary digital images.The new idea of converting wind power curve data cleaning problems into image segmentation problems is proposed,which can more intuitively clean the abnormal data.The amount of calculation can be reduced through information compression.(2)Based on the binary digital image,a feature image based on the median of the multi-directional distance metric is further constructed.The feature image where the higher gray value means the higher the confidence of being normal data can distinguish normal data from abnormal ones.(3)Constructing an energy space that combines class uncertainty theory and Fourier descriptor shape dissimilarity which can search for the optimal threshold based on the principle of minimizing the energy function and perform image thresholding on the feature image.In the original data space,all the discrete data points are labeled through the mapping between the data and pixel,and the abnormal data is further divided into three types based on the priori rule and the horizontal linear feature extraction algorithm.The method proposed in this paper is compared with the data-based and the image-based benchmark cleaning algorithm.The data set uses real data collected from 37 sets of wind turbines from two wind farms.The comparison of time consumption and classification metrics proves the versatility and effectiveness of the method proposed in this article.Based on the data cleaned by the proposed algorithm,regression modeling is performed,and the performance of several models for modeling wind power curves is compared.Through experiments,it is found that the five-parameter logistic function curve is most suitable for wind power curve modeling.