基于速度-关联约束的风电机组风速感知异常数据识别方法

Cleaning Method of Wind Speed Outliers for Wind Turbines Based on Velocity and Correlation Constraints

该文以风速时空关联特性为理论依据,针对风速数据单独清洗构建一种基于速度-关联约束的异常风速数据识别方法。分析了风电场典型异常风速的产生原因和分布特征,根据数据的变化趋势,将异常风速概括为突变型异常数据和渐近型异常数据两类;为提升风速数据清洗方法的准确性,提出一种基于二元形态分割算法的风速数据时序区间分割方法,将全局风速序列在时序上划分为多段分布独立的局部风速子序列,分别对每段风速子序列构建速度-关联约束条件,实现异常风速数据的识别。验证结果表明,所提方法能够有效识别风电场各类异常风速数据,清洗效果好、效率高,具有普适性和鲁棒性。

The wind speed sensors of wind turbines,due to long-term exposure to the harsh operating environment,are prone to instability or misalignment,resulting in a large amount of abnormal data in the collected raw data.The existing data cleaning methods for wind turbines,which primarily focus on the mapping relations between wind speed and wind power in power curves,are designed for wind power generation performance evaluation or wind power curve fitting.Few anomaly detection methods of wind speed data are proposed for the performance evaluation and health management(PEHM)for wind speed sensors.Consider the fact that different applications have different boundaries and requirements for data cleansing,this paper proposes a wind speed data cleaning method based on temporal and spatial correlation characteristics to deeply mine various abnormal data,especially the weak fault data located in or around the wind power curve.The anomaly detection results are helpful to find more information of wind speed sensers behaviors,which can be furtherly employed for PEHM.Firstly,the typical abnormal wind speed data in wind farms and their temporal features are summarized,which are classified into the sudden-changes and gradual-changes wind speed outliers.Then,a combined method using velocity and correlation constraints is proposed to detect various outliers.Since the local features of wind speed data are prone to partial coverage and annihilation when using long time scale data to model their temporal relations and correlations,this paper proposes a change-points detection algorithm based on shape binary segmentation to perform temporal interval partitioning for the global wind speed series.Thus,the velocity and correlation constraints are constructed for each sub-series in temporal and spatial dimensions to conduct data cleaning tasks.The experiment results on two datasets,including a simulated dataset and an actual wind farm dataset,show that,the proposed method is capable of detecting various abnormal wind speed data for wind turbines.The precision rates of the proposed anomaly detection method for sudden-changes and gradual-changes abnormal wind speed data reach 95.57%and 100%,respectively.Statistics on the abnormal data indicate that the gradual-changes abnormal wind speed data holds the dominate proportion among the detected outliers and the rate of abnormal data reaches 21.87%.An explanation for that is the wind capture performance of the selected sensor decrease severely,instead,the anti-electromagnetic performance is perfect.Interestingly,there are many scatters inside the wind speed and power curve that are detected as abnormal wind speed data by the proposed method.However,such scatters are detected as normal data for power curves fitting and wind power generation performance evaluation,which furtherly proves that the boundaries and definitions for outliers vary for different data category and digital applications.The following conclusions can be drawn from the case analysis:(1)The shape binary segmentation-based change-points detection method is capable of effectively identifying the moments when the temporal features of wind speed series change significantly.The temporally segmented local wind speed sequences are mor suitable to correlations modelling with higher accuracy and reliability.(2)The proposed data cleaning method based on velocity and correlation constraints provides accurate detection for various abnormal wind speed data in wind farms,especially for the weak fault data.(3)This paper presents a general data cleaning framework for time series based on spatial and temporal correlation characteristics,which takes the differences of the sensitivity to abnormal data in practical applications into account and can accurately identify various abnormal wind speed data with precision rate of 95.57%and 100%for sudden-changes and gradual-changes outliers.