基于高斯核函数改进的电力用户用电数据离群点检测方法

Improved Outlier Detection Method of Power Consumer Data Based on Gaussian Kernel Function

针对智能配用电大数据背景下用电数据离群点检测方法的适用性以及实际数据集中异常用电样本获取成本较高的问题,提出一种基于高斯核函数改进的电力用户用电数据离群点检测方法。首先通过模糊聚类的方法将用户分类;然后提取每一类用户的用电行为特征量,采用主成分分析法对特征集进行降维;最后利用高斯核函数改进局部离群因子算法,提出高斯核密度局部离群因子（Gaussian kernel densitybased local outlier factor,GKLOF）算法,通过理论推导与仿真实验相结合的方式分析了GKLOF算法的特性。选取了5000个用户真实的用电数据进行实验分析,实验结果表明,该方法具有较高的检测准确率以及较为稳定的判定阈值,并且受局部数据分布的影响较小,更加适用于用户用电行为复杂多样以及实际数据集中所有用户用电行为类型信息未知情况下的离群点检测。

In allusion to applicability of power consumer data outlier detection method in context of big data in smart power distribution and consumption systems, and high cost of obtaining abnormal samples for power consumption in actual data sets, an improved outlier detection method of power consumer data based on Gaussian kernel function was proposed. Firstly, the users were classified with fuzzy clustering method. Then various features of each type of users were extracted and PCA（principal components analysis） was used to reduce the dimension of feature vectors. Finally, Gaussian kernel function was used to improve local outlier factor（LOF） algorithm, and Gaussian kernel density local outlier factor（GKLOF） algorithm was proposed. Effectiveness of GKLOF algorithm was verified by combination of theoretical analysis and simulation. 5000 users’ real power data were selected to perform the simulation, and simulation results proved that the proposed method had high detection accuracy and stable decision threshold. In addition, local data distribution had minor impact on this method. Therefor it is more suitable for outlier detection in the case that power consumption behavior is complex and type of power consumption behavior is unknown.