基于伪标签的无监督电力数据异常检测框架
Pseudo-Lable Based Unsupervised Anomaly Detection Framework for Energy Data
林卫伟 1白向阳 2孔军1
作者信息
- 1. 江南大学物联网工程学院,江苏 无锡 214122
- 2. 江苏鼋博群智能技术有限公司,江苏 无锡 214026
- 折叠
摘要
电力系统容易受到各种异常用电行为的影响,异常检测可以提高电力系统的可靠性.针对现有异常检测模型存在阈值选择困难、精度低等问题,提出了一个基于伪标签的无监督异常检测框架PLAD.框架由伪标签提取模块PLE和基于重构误差的异常检测模块READ组成.PLE模块中提出了一种基于DBSCAN的自适应伪标签提取算法,用于提取电力数据的伪标签.READ模块采用双LSTM自编码器提取多级时序特征,并得到重构误差,后利用伪标签训练逻辑回归分类器,对重构误差进行分类,进行异常判定.通过某地区的电网运行数据验证了所提方法的有效性,实验结果表明,所提方法优于其他无监督异常检测方法.
Abstract
The energy system is vulnerable to various abnormalities,faults and abnormal energy consumption be-havior.Anomaly detection can improve the reliability of the energy system.Aiming at the problem of difficulty in threshold selection and low detection accuracy ofexisting anomaly detection models,an unsupervised anomaly detection framework pseudo-label-based anomaly detection(PLAD)is proposed in this paper,consisting of a pseudo label ex-tractor(PLE)and a reconstruction-error-based anomaly detector(READ).An adaptive pseudo label extraction al-gorithm based on density clustering(DBSCAN)was proposed in PLE module to extract pseudo labels of energy data.The READ module integrated a dual long short-term memory(LSTM)Autoencoder to extract the multi-level temporal features and obtain the reconstruction error.Then the pseudo label was used to train the logistic regression classifier to classify the reconstruction error and judge the anomaly.The effectiveness of the proposed method has been verified by the grid operation data of a certain area,and the experimental results show that our framework is superior to other unsupervised detection methods.
关键词
异常检测/伪标签/无监督/双长短时自编码器/密度聚类Key words
Anomaly detection/Pseudolabel/Unsupervised/Dual LSTM Autoencode/Density clustering引用本文复制引用
基金项目
国家自然科学基金(61362030)
国家自然科学基金(61201429)
中国博士后科学基金(2015M581720)
中国博士后科学基金(2016M600360)
出版年
2024
NETL
NSTL
万方数据