摘要
识别海量居民用户的用电行为模式并进行合理分类,可为需求侧精益化管理提供辅助决策。该文提出一种基于卷积神经网络自动编码器与层次聚类多任务联合模型的居民用电模式分类方法。首先,提出基于同时刻量测数据均值的缺失值填补方法和基于季节性极端学生化偏差检验的异常点检测方法,对海量且高维的用电数据进行数据清洗与修正;其次,利用卷积神经网络自动编码器对居民用电数据进行特征提取,获取可表征用户用电行为的特征向量;然后,结合层次聚类算法以及轮廓系数指标确定用户聚类个数以及聚类中心向量,并利用聚类中心向量初始化神经网络聚类层,进行用户聚类,将特征提取过程与用户聚类过程进行联合,组成多任务学习神经网络,实现端到端的用电模式分类;最后,结合环境温度和电价影响因素,在实际数据集进行验证。
Identifying the electricity consumption behavior patterns of massive residential users and then making a reasonable classification,can provide auxiliary decision-making for demand-side lean management.This paper proposes a method of residential electricity consumption pattern classification based on a multi-task joint model of convolutional neural network auto-encoder(CNN-AE)and hierarchical clustering.Firstly,a method for filling missing values based on the mean value of simultaneous measurement data and an outlier detection method based on seasonal hybrid extreme studentized deviate test,were proposed to clean and correct massive and high-dimensional electricity data.Secondly,the CNN-AE was used to extract the features of the residential electricity consumption data,and obtained the feature vector which could characterize the residents'electricity consumption behavior.Then,combining the hierarchical clustering algorithm and silhouette coefficient to determine the number of users'cluster and each cluster centers'vector,initialized the neural network layer for user clustering with cluster centers'vector;and joined the feature extraction process and user clustering process to form a multi-task learning neural network.This network was used to achieve end-to-end classification of residential electricity consumption patterns.Finally,considering environmental temperature and electricity price factors,the proposed method was verified on actual dataset.
作者
徐明杰
赵健
王小宇
宣羿
陈伯建
Xu Mingjie;Zhao Jian;Wang Xiaoyu;Xuan Yi;Chen Bojian(College of Electrical Engineering Shanghai University of Electric Power,Shanghai 200090 China;Hangzhou Power Supply Company State Grid Zhejiang Electric Power Co.Ltd Hangzhou,310016 China;Power Science Research Institute of State Grid Fujian Electric Power Co.Ltd,Fuzhou 350000 China)
出处
《电工技术学报》
EI
CSCD
北大核心
2022年第21期5490-5502,共13页
Transactions of China Electrotechnical Society
基金
国家重点专项(2020YFB1506804)
国家自然科学基金(51907114)
上海市教育发展基金会晨光计划(19CG61)资助项目。
关键词
居民负荷
负荷聚类
卷积神经网络
自动编码器
联合模型
Residential load
load clustering
convolutional neural network
auto-encoder
joint model
作者简介
徐明杰,男,1997年生,硕士研究生,研究方向为电力大数据。E-mail:xmj36@foxmail.com;通信作者:赵健,男,1990年生,副教授,研究方向为中压配电网精益化管理,图像处理、自然语言处理技术在电力系统运营管理中的应用等。E-mail:zhaojianee@foxmail.com。
