Abstract: Based on power customers which reach a very large amount and the feature of presence of outlier,and limitations of Partitioning Around Medoid(PAM) algorithm in handling large amounts of data and predefining the number of clusters,a new hybrid clustering algorithm called SOM-DB-PAM that is suitable for fast clustering of large number of electricity customers,is proposed.In the proposed algorithm,the Self-Organizing Map(SOM) neural network is used to train input data to ind prototype vectors that represents patterns of the input data set but far less than the number of it,and the prototype vectors are clustered by the PAM algorithm and to ensure the validity of clustering,the Davies-Bouldin(DB) indexis calculated for SOM prototype vectors to solve optimal number of clusters.Experimental results show that,compared with traditional clustering algorithms,the accuracy of classification is enhanced and when the amount of electricity customers is large,the proposed algorithm can achieve a fast and effective clustering.In addition,the blindness and subjectivity of predefining the number of clusters artificially is decreased.

Key words: power customer segmentation, Partitioning Around Medoid(PAM), Self-Organizing Map(SOM), hybrid clustering algorith, clustering analysis

摘要： 针对电力客户具有客户数量大、存在孤立点等特点,提出一种适用于对大量电力客户进行快速聚类的SOM-DB-PAM混合聚类算法。该算法利用自组织映射神经网络训练输入数据,以获取代表输入模式且数据量远小于输入数据量的原型向量,使用围绕中心点的切分(PAM)对该原 型向量聚类并用Davies-Bouldin指标判定最优聚类个数以保证聚类效果。实验结果表明,与传统聚类算法相比,该算法具有更高的分类正确率,当客户数量较大时,能实现对客户的快速、有效聚类,并减少人为指定聚类个数的盲目性和主观性。

关键词: 电力客户细分, 围绕中心点的划分, 自组织映射, 混合聚类算法, 聚类分析

CLC Number: 

• TP391