为了提高电力负荷曲线聚类精度，文中提出了一种基于核主成分分析 (KPCA)和改进 K-means 算法的电力负荷曲线聚类方法。该方法首先在划分聚类算法 K-means 基础上融入密度聚类思想，提出了融合密度思想的 K-means 算法 (DK-means 算法)，并在电力负荷曲线实验集上对比分析其聚类效果; 接着在实验集上比较各种降维算法的降维聚类精度和降维速度; 最后分析 KPCA + DK-means 组合算法的降维聚类能力。结果表明，戴维森堡丁指数 (DBI) 更适合作为电力负荷曲线聚类评价指标; 以 DBI 为评价指标，与 K-means、BIRCH、DBSCAN 和 EnsClust 4 种聚类算法相比，DK-means 的聚类精度更高; 与 LLE、MDS、ISOMAP 3 种非线性降维算法相比，KPCA 的降维速度更快;KPCA + DK-means 组合算法有良好的降维聚类能力，较 DK-means 在聚类精度和聚类效率上均有提升。KPCA + DK-means 组合算法可以实现电力负荷曲线的高效降维、精确聚类，对用电行为模式的准确提取起关键技术支持作用。

A clustering method of power load profiles based on kernel principal component analysis (KPCA) and improved K-means algorithm was proposed to improve clustering accuracy of power load profiles． Firstly，a K-means algorithm based on density idea，namely，density K-means (DK-means) was proposed by combining the density clustering method with partitioning clustering algorithm K-means． And the clustering effect was comparatively ana-lyzed on the experimental set of power load profiles． Then the dimensionality reduction accuracy and speed of vari-ous dimensionality reduction algorithms were compared on the experimental set． Finally，the dimensionality reduc-tion and clustering ability of the KPCA + DK-means combination algorithm was analyzed． The results show that，firstly，Davies-Bouldin Index (DBI) is more suitable for the evaluation index of power load profiles clustering; se-condly，taking DBI as the evaluation index，the clustering accuracy of DK-means is higher than that of K-means，BIRCH，DBSCAN and EnsClust algorithms; thirdly，compared with LLE，MDS and ISOMAP，the dimensionality reduction speed of KPCA is faster; finally，the KPCA + DK-means combination algorithm has good dimensionality reduction and clustering ability，and its clustering accuracy and clustering efficiency are better than those of DK-means． In short，the KPCA + DK-means combination algorithm can achieve efficient dimensionality reduction and accurate clustering of power load profiles，thus plays a key technical role in accurately extracting information of electricity consumption behavior．