采用欧式形态距离的负荷曲线近邻传播聚类方法

Affinity Propagation Clustering Method of Typical Load Curve with Euclidean Morphological Distance

为提取电力负荷数据的有效信息,改善传统聚类方法在电力负荷数据中相似度衡量方式单一及聚类效果较差的问题,提出一种采用欧式形态距离的负荷曲线近邻传播(AP)聚类方法。使用五分位法将用电负荷曲线重表达为曲线形态变化特征序列,使用改进最长公共子序列算法衡量不同特征序列之间的模式匹配度,以此作为曲线之间的差异度;构造一种兼顾曲线整体分布特征和曲线形态变化特征的双尺度相似性度量方法,使用熵权法对两种特征进行自适应配比;将所提相似度衡量方法应用到AP聚类方法中,改进相似度矩阵计算方法,对用户典型日用电负荷曲线进行聚类。在标准合成时间序列数据集上进行了实验对比,结果表明:欧式形态距离度量方法能够有效区分负荷曲线的变化特征;所提方法具有较高的聚类质量和稳健性,相比其他相似度量方法,调整兰德系数提高了9.0%~43.8%,DB指标与标准集相差0.0143,在电力实测数据集上能对用户进行合理划分。

An improved affinity propagation clustering algorithm with Euclidean morphological distance was proposed to extract effective information from load curve and improve the poor clustering effect of the traditional clustering algorithms.The quantile method was used to extract and re-express morphological characteristic of the original load curve,the improved longest common subsequence algorithm was introduced to measure the pattern matching distance between different user load consumption behavior sequences,and the measurement was taken as the degree of difference between the curves.Then a dual-scale similarity measurement considering the overall distribution of the curve and the change pattern of the curve shape was conducted,and the entropy weight method was used to adaptively match the weight between the two characteristics.The proposed similarity measurement was applied to the affinity propagation clustering algorithm to improve the similarity matrix calculation scheme,and the typical daily load curve of users was clustered.Taking the standard synthetic time series data as an example,clustering results indicate that the Euclidean morphological distance measurement method can effectively distinguish the dynamic change characteristics of load curves.The proposed method has higher clustering quality and algorithm robustness,increases the adjusted Rand index from 9.0%to 43.8%,differs from the standard set in DB index by 0.0143,and can distinguish users on the measured data set appropriately.