摘要
对异常用电行为进行自动分析是电力部门十分关心的问题,故提出一种基于L0稀疏超图半监督学习的用户窃电行为识别算法。与经典图模型中两个顶点的边连接方式不同,超图模型将具有相似特征的多个样本组建为超边,有利于表示用户用电数据间的复杂关联关系。然而常用的K近邻构建超边方法存在超边大小固定,无法有效匹配用户用电数据非均匀分布的缺点。为解决该问题,建立基于L0稀疏重构的超图模型,模型针对每一用户数据建立一个超边,通过L0范数约束的稀疏分解自适应选择与当前用户紧密关联的多个样本,更有利于匹配用户数据的分布结构。继而构建超图拉普拉斯正则约束的半监督分类模型,利用少量的标定样本数据,判别用户用电行为是否存在异常。选取某城市8 900余居民300余天的实际用电数据作为测试样本集,实验结果验证了方法的有效性。
It is a very important problem for the electric power department to carry out the automatic analysis of the abnormal power consumption behavior. This paper proposed a semi-supervised classification algorithm based on L0 sparse hypergraph learning. Different from the pair-wise link in traditional graph model,hyperedge in hypergraph is a subset of data points sharing with same attributes,which is beneficial for representing the complex correlations between the power load profiles of different users. In practices,the size of hyperedges constructed by common K-nearest-neighbor method is same for all vertexes,which make it hard to capture the non-uniform distributed characteristics of the power load dataset.In order to solve this problem,this thesis established a hypergraph model based on L0 sparse reconstruction. The model created a hyperedge for each user's data. It was more advantageous to match the distribution of user data by adaptively selecting multiple samples closely related to the current user through sparse decomposition of L0 norm constraints. Then,a semi-supervised classification model of the hypergraph Laplacian regular constraint was constructed,and a small amount of calibration sample data was used to determine whether the user 's electrical behavior was abnormal. We selected the actual electricity consumption data of more than 300 days in a city of more than 8 900 residents as the test sample set. Experimental results verified the effectiveness of the proposed method.
出处
《计算机应用与软件》
北大核心
2018年第2期54-59,共6页
Computer Applications and Software
基金
国家自然科学基金项目(61300162)
国家电网公司2016年科技项目
关键词
用电行为
分析L0稀疏
超图半
监督学习
Electricity behavior analysis
L0 sparsity
Hypergraph
Semi-supervised learning
