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.
出处
《中国电机工程学报》
EI
CSCD
北大核心
2016年第2期379-387,共9页
Proceedings of the CSEE
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
power consumption patterns
power big data
anomaly detection
unsupervised learning
local outlier factor
anti-stealing of power energy
