基于滑动窗口和聚类算法的变压器状态异常检测

Anomalous State Detection of Power Transformer Based on Algorithm Sliding Windows and Clustering

在线监测数据在变压器运行状态中发挥出重要角色,而传统的异常状态检测一般基于阈值判定法,难以及时发现变电设备的异常情况,甄别噪声数据。针对上述问题,根据变压器在线监测数据中异常值特点,提出了一种基于滑动窗口和聚类算法的变压器状态异常检测方法。首先,利用时间序列和滑动窗口对多维的在线监测数据流进行筛选,记录异常点的发生时间和类型,建立候选异常数据集合的判断模型;其次,基于无监督的k-means聚类方法建立多元特征量数据点的异常检测模型,并用于在线监测实时数据的异常检测,判断异常时刻与异常类型。通过某变电站的油中气体数据对本文算法进行了验证,结果表明,该方法可以实时检测在线监测数据流中因运行状态变化而产生的趋势异常,并祛除少量传感器噪声或突变值的影响,具有较高的实用价值。

The online monitoring of power transformer is an important guarantee of operation situation. The traditional method is based on the threshold in guidelines or related documents of state grid ,and the anomalous situation can not be detected in time. Aiming at solving these problems, we proposed a fast anomalous state detection method of power trans- formers based on sliding windows and clustering. First, the time series and sliding windows were used to extract features of multi-dimensional state data. The anomalous time and type were recorded in order to establish the detecting model of anomalous data. Then, an anomaly detecting model were established by the unsupervised k-means clustering method. This model was used for detection of the multi-dimensional online monitoring data so that the anomalous time and type were detected. The measured data of oil gas of one transformer in substation were used to validate the method. The re- suits reveal that the anomalous trends of online monitoring data stream can be detected in time, which are caused by the changing of operation situations. And the disturbance of few noise data and abrupt change point can be eliminated. Over- all, the method is valid for the anomalous state detection.