基于无监督学习的电力用户异常用电模式检测  

Anomaly Detection for Power Consumption Patterns Based on Unsupervised Learning

作　　者：庄池杰[1] 张斌[2] 胡军[1] 李秋硕 曾嵘[1] 

机构地区：[1]电力系统及发电设备控制和仿真国家重点实验室(清华大学电机系),北京市海淀区100084 [2]国家电网公司西北分部,陕西省西安市7100481 [3]南方电网科学研究院,广东省广州市510080

出　　处：《中国电机工程学报》2016年第2期379-387,共9页Proceedings of the CSEE

摘　　要：检测异常用电模式的主要目的在于降低非技术性损失(non-technical losses,NTL),降低电力公司的运营成本。该文提出了基于无监督学习的异常用电模式检测模型,适用于电力用户数据集缺乏训练样本的情况。该模型包括特征提取、主成分分析、网格处理、计算局部离群因子等模块。首先提取多个表征用户用电模式的特征量,通过主成分分析将每个用户映射到二维平面,实现数据可视化并便于计算局部离群因子。网格处理技术筛选出低密度区域的数据点,显著提升了算法效率。该模型输出所有用户用电行为的异常度及疑似概率排序,研究结果表明利用该排序,只需要检测异常度排序靠前的少数用户即可查出大部分异常用户。The primary purpose of anomaly detection for power consumption patterns is to lower the non-technical losses（NTL）, thus reducing the operating costs for power utility. A model based on unsupervised learning was proposed to detect anomaly consumption patterns. This model is suitable for load dataset without training set. The model includes modules of feature extraction, principal component analysis, grid processing, calculation of local outlier factor（LOF）, etc. Firstly, various features were extracted from load profiles to characterize consumption patterns of the customers. Then PCA was used to map customers to a two-dimensional plane. This mapping procedure is in favor of data visualization and LOF calculation. The grid processing procedure can screen data in low density region and thus lift calculation efficiency. The output of the model is abnormal degree for all customers＇ consumption patterns. The result indicates that with the use of this abnormality sequence, detecting customers with higher LOF rank can find out most abnormal consumption patterns.

关 键 词：用电模式 电力大数据 异常检测 无监督学习 局部离群因子 反窃电技术 

分 类 号：TM715[电气工程—电力系统及自动化]