基于大数据分析的输变电设备状态数据异常检测方法  

An Method for Anomaly Detection of State Information of Power Equipment Based on Big Data Analysis

作　　者：严英杰[1] 盛戈皞[1] 陈玉峰 江秀臣[1] 郭志红 杜修明 

机构地区：[1]上海交通大学电气工程系,上海市闵行区200240 [2]国网山东省电力公司电力科学研究院,山东省济南市250002

出　　处：《中国电机工程学报》2015年第1期52-59,共8页Proceedings of the CSEE

基　　金：国家863高技术基金项目(SS2012AA050803);国家电网公司科技项目(520626140020)~~

摘　　要：传统的阈值判定方法难以准确检测输变电设备的状态异常,该文提出一种基于时间序列分析和无监督学习等大数据分析的异常检测方法,从数据演化过程、数据关联的全新角度实现异常检测。通过时间序列模型和自适应神经网络对历史数据潜在的特征进行挖掘,并将数据对时间的动态变化规律用转移概率序列表示。针对多维的监测数据,运用无监督聚类方法简化各参量之间的相关关系,从而避免参量间相关性难以确定的问题。提出异常检测体系,并使之适用于输变电设备状态监测数据流,实现数据流中异常的快速检出。最后结合运行实例验证了提出方法的有效性,表明本方法能快速检测出设备的异常运行状态。To detect the anomaly state of power equipment, the traditional method threshold value determination is unable to ensure the accuracy. This paper proposed a method for anomaly detection of state data of power equipment based on big data analysis from time series analysis and unsupervised learning, thus a new perspective of data association and data evolution was achieved. Mining the potential features through time series model and self-organized maps, the method put the original data series into the transition probability series. To simplify the relationship between the multidimensional state sequences, the unsupervised learning was used to form several clusters. The method proposed the anomaly detection framework which has a rapid detection speed and is applicable for the state data flow. At last, the effectiveness of the method is verified by being combined with running instances and the result shows that the abnormal operating state can be rapidly detected.

关 键 词：大数据 异常检测 时间序列 神经网络 无监督聚类 

分 类 号：TM76[电气工程—电力系统及自动化]