线性回归模型的多离群点检测方法及节能应用  

Multiple Outlier Detection Method for Linear Regression Model and Its Energy Conservation Application

作　　者：刘丹丹[1,2] 陈启军[1] 森一之 

机构地区：[1]同济大学电子与信息工程学院,上海201804 [2]上海电力学院计算机与信息工程学院,上海200090 [3]日本三菱电机先端技术研究所

出　　处：《信息与控制》2013年第6期765-771,共7页Information and Control

基　　金：上海市科委国际合作项目(13510711100);国家自然科学基金资助项目(91120308)

摘　　要：办公设备能耗可使用线性回归模型进行描述,为检测该类能耗数据的异常点,提出了基于最小截平方和估计的单链接层次聚类多离群点检测算法.首先将该方法在不同类型的典型数据集中进行验证,证明算法具有优良的性能.而后将算法应用于办公设备能耗数据集,通过实验表明算法处理淹没问题及掩盖问题的能力优于其它算法.其不但能够正确确定异常能耗数据,还能够直观地给出各个异常能耗数据的离群程度,管理者可据此制定正确合理的能源管理方案,最终达到节能的目的.The energy consumption of the office equipment can be described by linear regression model. For the model, a multiple outlier detection algorithm based on single-link hierarchical clustering and LTS （least trimmed squares） estimator is proposed. This method is validated in different types of typical data sets. And the results prove that it has excellent perfor- mance. Then it is applied to office equipment energy consumption data sets. The experiments show that it has better ability to deal with the masking and swamping problems than other algorithms. Also, the method can not only correctly identify the outliers but also provide abnormal degree of outliers. The managers can develop the reasonable energy management solutions and achieve the purpose of energy saving.

关 键 词：单链接层次聚类算法 线性回归模型 离群点检测 最小截平方和估计 掩盖问题 淹没问题 

分 类 号：TK01[动力工程及工程热物理] TP311[自动化与计算机技术—计算机软件与理论]