基于密度聚类和Fréchet判别分析的电价执行稽查方法

A Method to Inspect the Implementation of Electricity Price Based on Density Clustering Analysis and Fréchet Discriminant Analysis

针对传统的电价执行稽查方法存在着一定的人为因素和随意性的不足,提出了一种基于密度聚类分析和Fréchet判别分析的电价执行稽查方法。首先,从计量营销一体化系统中提取电力用户用电数据,利用数据预处理模块对其进行预处理;其次,利用密度聚类分析技术,根据电力用户月用电量邻域内的户数密度,选取高户数密度样本作为k-means聚类算法的最优初始聚类中心并根据组合聚类评价指标确定最佳聚类数,从而构建不同营业区域、不同用电类别的典型电力用户月用电量轨迹曲线;然后,采用基于Fréchet距离判别的分析方法对待稽查用户进行辨别分析,包括:1)计算新样本与典型月用电量轨迹的Fréchet距离,设定合理的距离判别阈值;2)计算每个用户用电异常嫌疑系数和判别吻合系数,分别设定阈值,确定电价异常用户。通过某供电企业的应用实例,文中提出的电价执行稽查方法的稽查准确率为83.33%,优于传统稽查方法的稽查结果(28.57%);异常用户嫌疑系数普遍较高,判别吻合系数相对较低。实例验证了所提稽查方法的准确性和有效性。

There are a certain anthropogenic factors and arbitrariness when executing traditional marketing inspection. For this reason, a methodology of inspecting implementation of electricity price based on density clustering analysis and Fréchet discrimination analysis is proposed. Firstly, electric power user data from Metering and Marketing Integration System is extracted and data pre-processing performed. Secondly, initialcluster centers of k-means algorithm are selected with Density k-means Cluster Analysis（DKCA） based on monthly electricity consumption of each electricity userdetermined according to user density in neighborhood of monthly electricity consumption; multi clustering validity index is applied to determine optimal number of clusters. In this way, typical electric power consumption curves in different business place and different type of consumers are established. Then Fréchet Distance Discriminant Analysis（FDDA） is adopted to discriminate new consumers, including： 1） calculate distance between new power users and typical monthly electricity consumption curves and set proper distance threshold; 2） calculate suspect coefficient and discriminant fitting coefficient in each user and assign them with rational threshold respectively. Therefore, abnormal users are determined. A case study based on certain power supply is analyzed. Simulation results show that the proposed methodology is obviously advantageous in accuracy with value 83.33%, compared with traditional one with 28.57%. Suspect coefficients are larger while discriminant fitting coefficients are smaller in each user generally. Results demonstrate accuracy and effectiveness of the proposed method.