| 引用本文: | 汪颖,杨维,肖先勇,张姝.基于去噪自编码器网络特征降维与改进小批优化K均值算法的海量用户用电行为聚类及分析[J].电力自动化设备,2022,42(6): |
| WANG Ying,YANG Wei,XIAO Xianyong,ZHANG Shu.Clustering and analysis of electricity consumption behavior of massive users based on network feature dimension reduction of denoising autoencoder and improved mini-batch K-means algorithm[J].Electric Power Automation Equipment,2022,42(6): |
|
| 摘要: |
| 海量用户用电特性的挖掘与分析对实现电网与用户间的双向互动具有十分重要的意义。提出一种适用于海量用户用电行为聚类及分析的算法，以降低算法时间复杂度，提升海量用户负荷数据分析效率。提取用户用电行为特征，构建多层去噪自编码网络，实现多维特征的降维；利用小批优化K均值算法进行聚类分析，并对算法进行初始聚类质心优化与超参数优化的改进以提升算法收敛速度与效果，其中超参数优化利用基于高斯过程的贝叶斯优化算法进行；利用类间分离度和类内内聚度的相关指标对聚类效果进行评价；通过互信息筛选有效聚类特征，实现用户画像。算例结果表明，所提方法在特征优化、聚类效果与收敛速度上均有较好的表现。 |
| 关键词: 用电行为 特征降维 聚类分析 互信息 小批优化K均值算法 超参数优化 贝叶斯优化 |
| DOI：10.16081/j.epae.202203017 |
| 分类号:TM73 |
| 基金项目:国家自然科学基金资助项目(52077145,52007126) |
|
| Clustering and analysis of electricity consumption behavior of massive users based on network feature dimension reduction of denoising autoencoder and improved mini-batch K-means algorithm |
|
WANG Ying, YANG Wei, XIAO Xianyong, ZHANG Shu
|
|
College of Electrical Engineering, Sichuan University, Chengdu 610065, China
|
| Abstract: |
| The mining and analysis of electricity consumption features of massive users is of great significance for realizing bi-directional interaction between power grid and users. An algorithm suitable for clustering and analysis of electricity consumption behavior of massive users is proposed to reduce the time complexity of the algorithm and improve the load data analysis efficiency of massive users. The electricity consumption behavior features of the users are extracted, and a multi-layer denoising autoencoder network is built to realize dimension reduction of multi-dimensional features. The mini-batch K-means algorithm is used for clustering analysis, and the improvements of initial clustering centroid optimization and hyperparameter optimization on the algorithm are carried out to improve the convergence speed and effect of the algorithm, in which, the hyperparameter optimization is carried out with Bayesian optimization algorithm based on Gaussian process. The related indexes of separation degree between the clusters and the cohesion degree within the clusters are used to evaluate the clustering effect. The effective clustering features are screened through mutual information to realize user portraits. The case results show that the proposed method has good performance in feature optimization, clustering effect and convergence speed. |
| Key words: electricity consumption behavior feature dimension reduction clustering analysis mutual information mini-batch K-means algorithm hyperparameter optimization Bayesian optimization |
