Identification of Abnormal Power Consumption Behavior Based on Dimensionality Reduction of Characteristic Index and Improved Density Peak Clustering Algorithm
Purposes In order to accurately identify the abnormal power consumption behavior of power users and reduce the non-technical loss of power grid, an abnormal power consumption behavior identification model based on linear discriminant analysis (LDA) and improved density peak clustering (IDPC) algorithm is proposed. Methods First, a feature set was constructed from the power consumption data to reflect the power consumption behavior of users. Second, LDA was used to reduce the dimension of the extracted feature set. Then, IDPC algorithm was used to cluster the feature set after dimensionality reduction, and the users with different electricity behavior characteristics were clustered before abnormality identification. Finally, the outlier index was defined to describe the degree of user outlier in the feature set after dimensionality reduction, and the final model outputs the order of the degree of abnormality of all users’ electricity behavior. In this paper, in view of the deficiency that the truncated distance in the density peak clustering (DPC) algorithm needs to be set manually, the Bonferroni index was defined, and the improved whale optimization algorithm was applied to the optimization process of DPC truncated distance parameters. Results Experimental results show that the proposed model can detect most abnormal users only by detecting a small number of users with large degree of abnormality, and has higher recall rate and accuracy than those of other models.
在实际的异常检测问题中,异常样本通常较正常样本比例小,导致了数据类别的不平衡性[12]。这样的不平衡分布会使得分类模型在训练过程中倾向于优化对多数类别的分类,使得少数类别的辨识准确率较低。而回归模型面对海量智能电表时,为每个电表建立预测模型将变得非常耗时,还会面临巨大的计算和存储资源挑战[13]。因此,本文重点研究不需要训练过程的无监督学习模型。基于聚类的异常辨识方法属于最典型的无监督学习,聚类模型首先对数据进行聚类操作,将数据样本划分为不同的群集,然后检测不属于任何群集或属于小群集的数据点作为异常值。K-means作为一种简单高效的聚类算法,其用于异常检测的效果很大程度上取决于聚类中心的选取。文献[14]通过均匀选择数据紧密区域并避免离群点区域,对K-means聚类算法的初始聚类中心选择过程进行了优化。近几年,基于密度聚类的方法逐渐成为异常检测问题的研究热点,文献[15]采用基于密度的带噪声空间聚类方法(density-based spatial clustering of applications with noise,DBSCAN)对三次指数平滑模型的预测电量与真实电量之间的残差进行聚类,从而检测出有异常的用电量值,实验结果表明其检测率和误报率等评价指标均优于LOF算法和支持向量机算法。然而DBSCAN方法的检测结果对参数取值十分敏感,而且需要确定邻域半径Eps和最小邻居数目minPts双参数,现阶段DBSCAN的参数主要根据人工经验进行选取,缺乏相应的参数选取理论。文献[16]采用LOF算法进行电力用户异常用电行为的挖掘,并采用网格技术提高算法效率,然而LOF算法并不适合检测包含多种用电行为的电力负荷。综上所述,考虑到用电行为的多样性和算法参数选取的便捷性,文章采用密度峰值聚类(density peak clustering,DPC)作为异常辨识模型的核心算法,基于改进鲸鱼优化算法对其参数进行寻优,利用离群异常指数判别数据中的离群点。文章按照特征集构造、特征降维、聚类分析、离群点检测的流程进行异常用电行为的辨识。
由于构造的特征数量较多且某些特征之间包含重叠信息,具有很强的相关性。为了在低维空间直观地展示用户的用电行为并且提升算法效率,有必要对特征集进行降维处理。常用的降维方法包括主成分分析(principal component analysis, PCA)、因子分析(factor analysis, FA)和线性判别分析(linear discriminant analysis, LDA)。就模型的监督性而言,其中PCA和FA属于无监督学习模型,即不需要数据标签来学习。PCA主要关注数据的方差分布,使得投影后的数据具有最大的方差,由于不需要类别标签,在某些情况下可能无法很好地区分不同类别。FA专注于揭示观测数据背后的潜在结构,它假设数据是由潜在因素和随机误差组成的,其数学模型可以表示为,其中 X 为原始数据矩阵, A 为因子载荷矩阵,F为公因子, AF为潜在因子,为残差矩阵。FA的目标是估计 A 、F以及,使得 X 能够被最好地重建。与PCA和FA不同,LDA是一种有监督学习模型,即需要明确数据的类别信息标签,需要明确指出的是,这里的标签是数据固有的类别标签,而非异常辨识的异常用户标签。LDA可以最大化不同类别之间的间距,同时最小化同一类别内部的差异。就模型可保留新特征的最大维度而言,PCA和FA降维后的新特征维度可以达到原数据的维度,而LDA降维后新特征维度最多为类别标签数减1。
YORUKOGLUS, NASIBOVF, MUNGANM,et al.The effect of the types of network topologies on nontechnical losses in secondary electricity distribution systems[J].IEEE Transactions on Industry Applications,2016,52(5):3631-3643.
[2]
YIPS C, WONGK, HEWW P,et al.Detection of energy theft and defective smart meters in smart grids using linear regression[J].International Journal of Electrical Power and Energy Systems,2017,91:230-240.
ZHUT Y, AIQ, HEX,et al.An overview of data driven electricity consumption behavior analysis method and application[J].Power System Technology,2020,44(9):3497-3507.
ZHANGZ P, LIS, LIUW X,et al.Outlier detection algorithm based on fast density peak clustering outlier factor[J].Journal on Communications,2022,43(10):186-195.
LINZ A, LIUS Y, JINW C,et al.Detection method of hidden danger for power utilization of low-voltage users based on local outlier factor-boosting[J].Automation of Electric Power Systems,2022,46(1):130-138.
LIG C, LUJ, WANGY,et al.Isolated-forest electricity theft detection algorithm based on Bagging secondary weighted ensemble[J].Automation of Electric Power Systems,2022,46(2):92-100.
[11]
LIY C, QIUR X, JINGS T.Intrusion detection system using Online Sequence Extreme Learning Machine (OS-ELM) in advanced metering infrastructure of smart grid[J].PloS One,2018,13(2):e0192216.
WUY K, LIW, NIM Y,et al.Anomaly detection model based on one-class support vector machine fused deep auto-encode[J].Computer Science,2022,49(3):144-151.
[14]
ZOUD X, XIANGY J, ZHOUT,et al.Outlier detection and data filling based on KNN and LOF for power transformer operation data classification[J].Energy Reports,2023,9(S7):9698-9711.
WANGW T, ZHAOJ, WANGX Y.Abnormal electricity consumption detection based on hedge-antagonism mechanism and graph attention network[J].Automation of Electric Power Systems,2022,46(22):120-128.
[17]
BIANJ H, WANGL, SCHERERR,et al.Abnormal detection of electricity consumption of user based on particle swarm optimization and long short term memory with the attention mechanism[J].IEEE Access,2021,9:47252-47265.
WANGJ Y, ZHANGS F.Anomaly detection for power consumption patterns based on linear discriminant analysis and density peak clustering[J].Automation of Electric Power Systems,2022,46(5):87-98.
[20]
TIANY, SEHOVACL, GROLINGERK.Similarity-based chained transfer learning for energy forecasting with big data [J].IEEE Access,2019,7:139895-139908.
XIAOY, ZHENGK H, YUZ Z,et al.Power data anomaly detection based on Holt-winters model and DBSCAN clustering[J].Power System Technology,2020,44(3):1099-1104.
ZHUANGC J, ZHANGB, HUJ,et al.Anomaly detection for power consumption patterns based on unsupervised learning[J].Proceedings of the CSEE,2016,36(2):379-387.
SUNY, LIS H, CUIC,et al.Improved outlier detection method of power consumer data based on gaussian kernel function[J].Power System Technology,2018,42(5):1595-1606.
DENGMB, XUZ M, DENGZ F,et al.Study on stealing recognition model based on multi feature fusion[J].Computer & Digital Engineering,2017,45(12):2398-2401,2414.
LIJ H, ZHANGJ H, MUG,et al.Day-ahead optimal scheduling strategy of peak regulation for energy storage considering peak and valley characteristics of load[J] Electric Power Automation Equipment,2020,40(7):128-136,140.