数据流驱动的电压三相不平衡异常检测研究  

Research on data-stream-driven voltage three-phase unbalance anomaly detection

作　　者：刘波 王大鹏 闫永昶 刘通宇 张园园 袁培森[3] Liu Bo;Wang Dapeng;Yan Yongchang;Liu Tongyu;Zhang Yuanyuan;Yuan Peisen(Supervision and Support Center of Power Supply Service of State Grid Mengdong Power,Tongliao 028000,China;State Grid Inner Mongolia East Power Company Ltd.,Hohhot 010010,China;College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210095,China)

机构地区：[1]国网蒙东电力供电服务监管与支持中心,内蒙古通辽028000 [2]国网内蒙古东部电力有限公司,内蒙古呼和浩特010010 [3]南京农业大学人工智能学院,江苏南京210095

出　　处：《南京理工大学学报》2023年第2期245-253,共9页Journal of Nanjing University of Science and Technology

基　　金：国家自然科学基金(61502236);江苏省农业科技自主创新资金项目(SCX(21)3059);上海市大数据管理系统工程研究中心开放基金(HYSY21022)。

摘　　要：为了提高当前电力计量系统中的电能量时间序列数据异常检测的准确率与效率,该文提出一种数据流驱动的电压三相不平衡异常检测方法。结合已有的面向数据流异常检测思路,实现对时间序列电能量数据的采样。采用基于循环神经网络的长短期记忆神经网络对数据建模预测,计算实际值与预测值之间的误差。采用指数加权平均的滑动窗口对误差进行平滑,实现了在过滤了三相电压数据正常波动的影响下判断某一时间段是否出现异常。试验结果证明,该方法在召回率和综合评价指标等有所提高,验证了该文方法对于电压三相不平衡的异常检测问题的可行性和有效性。To improve the accuracy and efficiency of current electric energy time series data anomaly detection in power metering system,this paper proposes a data stream driven anomaly detection method for voltage three-phase unbalance.Based on existing data stream-oriented anomaly detection motivates,this paper implements sampling of time series electrical energy data.The existing long short-term memory(LSTM)neural network is used to model the data for prediction,and the error between the actual and predicted values is calculated.A sliding window with exponentially weighted average is used to smooth the error,and it is realized to determine whether an anomaly occurs at a certain time period with the influence of normal fluctuations in the three-phase voltage data filtered.The experimental results indicate that the method has improved in terms of recall rate and comprehensive evaluation index,which verifies the feasibility and effectiveness of the method for the anomaly detection problem of voltage three-phase unbalance.

关 键 词：电压三相不平衡 数据流 异常检测 循环神经网络 长短期记忆网络 指数加权移动平均 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]