基于变分贝叶斯推断的DPGMM风电机组异常数据识别研究

Abnormal Wind Turbine Data ldentification Using a Dirichlet Process Gaussian Mixture Model Based on Variational Bayesian Inference

为了准确识别和剔除风电机组在实际运行过程中产生的异常数据,以便为功率预测等工作提供有效的数据支持,通过分析风电机组运行数据散点在风速-功率(v-P)坐标系中的分布特征,提出了基于变分贝叶斯推断的狄利克雷过程高斯混合模型异常数据识别方法。将试验机组E17实测数据散点沿水平功率方向以一定间隔划分区间,采用能自适应确定最佳分量个数的狄利克雷过程高斯混合模型对每一个功率区间内的数据散点进行聚类,结合各高斯分量置信椭圆参数及数据散点在v-P坐标系中的分布特征,对试验机组E17各功率区间内的高斯分量及其聚类散点进行异常标识。结果表明:该模型克服了传统高斯混合模型需要人为确定分量个数的缺点,能够对风电机组异常数据进行准确识别。

In order to accurately identify and eliminate abnormal data generated by wind turbines in the actual operation process,so as to provide effective data support for power prediction and other work,an abnormal data identification method using a Dirichlet process Gaussian mixture model(DPGMM)based on variational Bayesian inference was proposed by analyzing the distribution characteristics of wind turbine operational data points in a wind speed-power(α-P)coordinate system.The measured data scattered points of test unit E17 were divided along the horizontal power direction at a certain interval,and the data scattered points in each power range were clustered by using DPGMM that can adaptively determine the optimal number of components,the confidence ellipse parameters of each Gaussian component and the distribution characteristics of data scattered points in the P coordinate system were combined,so as to identify anomalies in the Gaussian components and their clustering scattered points in each power range of the test unit E17.Results show that the model overcomes the shortcoming of the traditional Gaussian mixture model that the number of components needs to be determined manually,and the new model can accurately identify abnormal wind turbine data.