特征选择及数据质量对负荷识别算法的影响研究  

Research on the Influence of Feature Selection and Data Quality on Load Identification Algorithm

作　　者：延菲 张瑞祥 孙耀杰[1,2,3] 康巍 张健[5] 孙洁 李琦芬 YAN Fei;Zhang Ruixiang;SUN Yaojie;KANG Wei;ZHANG Jian;SUN Jie;LI Qifen(School of Information Science and Engineering, Fudan University, Shanghai 200433, China;Institute for Six-sector Economy, Fudan University, Shanghai 200433, China;Shanghai Engineering Research Center for Artificial Intelligence Integrated Energy System, Shanghai 200433, China;New Energy Department, China Quality Certification Centre, Beijing 100070, China;Shanghai Minghua Electric Power Science and Technology Co., Ltd, Shanghai 200090, China;Shaanxi Province Energy Administration, Xi'an, Shanxi 710006, China;College of Energy and Mechanical Engineering, Shanghai University of Electric Power, Shanghai 200090, China)

机构地区：[1]复旦大学信息科学与工程学院,上海200433 [2]复旦大学六次产业研究院,上海200433 [3]上海综合能源系统人工智能工程技术研究中心,上海200433 [4]中国质量认证中心新能源风电部,北京100070 [5]上海明华电力科技有限公司,上海200090 [6]陕西省能源局,陕西西安710006 [7]上海电力大学能源与机械工程学院,上海200090

出　　处：《复旦学报（自然科学版）》2021年第6期811-816,共6页Journal of Fudan University：Natural Science

基　　金：国家重点研发计划(2019YFB2103200,2018YFB1500904);2019年度上海市工程技术研究中心建设计划(19DZ2252000);2020年第一批上海市信息化发展专项资金(智慧城市建设和大数据发展)(202001015)。

摘　　要：非侵入式负荷监测(Non-Intrusive Load Monitoring,NILM)是电网对需求侧进行精细化管理的关键技术之一,NILM通过实时监控用户用电设备的运行状态与能耗状况,为电网侧制定调度策略及用户侧制定节能计划提供了重要依据.在家庭NILM系统中,识别精度、实时性和实现成本是评价负荷识别算法的重要指标,选择唯一性强、维度低的负荷特征可提升负荷识别模型的精度和实时性,在不牺牲精度的前提下,降低数据采集频率,有助于控制硬件成本.本文提取了用电设备稳态运行时的形状和幅值特征,测试了改进kNN算法对不同特征组合的识别效果,筛选出唯一性最强的负荷特征为形状特征与幅值特征的组合.然后在此基础上,以采样频率和V-I轨迹分辨率为自变量,通过实验研究了数据质量与识别效果的对应关系,找出导致识别效果下降的采样频率和分辨率的临界点.Non-Intrusive Load Monitoring(NILM)is one of the key techniques for precise management of demand side.NILM monitors the operation state and energy consumption of users'electrical equipment in real time,which provides an important basis for the grid side to formulate scheduling strategies and the user side to formulate energy conservation plans.In the home NILM system,recognition accuracy,real-time performance,and implementation cost are important indicators for evaluating load recognition algorithms.Selecting unique and low-dimensional load features can improve the accuracy and real-time performance of the load recognition model.Without sacrificing accuracy,reducing the frequency of data acquisition helps control hardware costs.This paper extracts the shape and amplitude characteristics of electrical equipment during steady-state operation,tests the recognition effect of the improved kNN algorithm on different combinations of characteristics,and selects the most unique load feature as the combination of shape and amplitude.Then,on this basis,taking the sampling frequency and the resolution of V-I trajectory as independent variables,the corresponding relationship between the data quality and the recognition effect is studied through experiments,and the critical points of the frequency and resolution that lead to the reduction of the recognition effect are found.

关 键 词：负荷识别 KNN 分辨率 采样频率 负荷特征组合 

分 类 号：TM714[电气工程—电力系统及自动化]