裴宇婷，秦 超，余贻鑫
AuthorsHTML:裴宇婷，秦 超，余贻鑫
AuthorsListE:Pei Yuting，Qin Chao，Yu Yixin
AuthorsHTMLE:Pei Yuting，Qin Chao，Yu Yixin
Unit:天津大学智能电网教育部重点实验室，天津 300072
Unit_EngLish:Key Laboratory of Smart Grid of Ministry of Education，Tianjin University，Tianjin 300072，China
Abstract_Chinese:为提高电网的安全运行水平和经济性，灵活的可重构的网络拓扑结构是未来智能配电网的基本特征．而配电管理系统(DMS)中的大部分功能，如状态估计，潮流计算和电压控制等，都基于网络当前的拓扑结构．因此，拓扑辨识是DMS的基础功能之一，研究更为高效和准确的智能配电网拓扑辨识方法具有重要意义．结合配电网的结构和运行特点，建立了基于机器学习的智能配电网拓扑辨识框架，并提出了基于LightGBM和深度神经网络(DNN)的配电网在线拓扑辨识方法．该方法借助LightGBM实现特征选择，筛选出对配电网拓扑辨识最有效的少量量测，以深度神经网络实现配电网运行断面量测数据与其拓扑结构间的映射．考虑实际应用中可能存在量测数据丢失的情况，提出了基于最小方差的缺失值填补方法．同样利用样本间的最小方差进行未知拓扑的甄别，并借助增量学习机制，通过增量训练DNN模型实现拓扑知识库的更新．与现有方法相比，本文提出的拓扑辨识方法仅需要配电网中少量节点的运行断面量测数据，同时适用于辐射状和弱环网结构，计算效率可支持在线拓扑辨识．通过IEEE 33节点配电网和PG&E 69节点配电网验证了所提方法的有效性与优越性，并分析了对于不同噪声水平情况、量测特征值缺失和存在未知拓扑的适应性．
Abstract_English:To improve the security and economy of the power grid，future smart distribution grids should have a flexible topology reconfiguration as a fundamental characteristic. Most functions of the distribution management system (DMS)，such as state estimation，power flow calculation，and voltage control，require the current topology of the grids. Therefore，topology identification is a key function of the DMS. Exploiting more accurate and efficient topology identification approaches is thus of great significance. Considering the topologies and operational characteristics of distribution grids，a topology identification framework of smart distribution grids under a machine learning scheme was developed in this study. An online topology identification method based on LightGBM and deep neural networks(DNNs)was also presented. A feature selection algorithm based on LightGBM was adopted to select the most effective measurements for topology identification task，and DNNs were built to model the mapping relationship between measurement data snapshots and topologies. Considering the possibility of missing measurement data in practice，an imputation method based on the minimum variance was proposed. The minimum variance was also used to identify unknown topologies that are not included in the training set. An incremental learning mechanism was then added to adjust the DNN and update the knowledge base of topologies. Compared with existing methods，the proposed method simply requires a small portion of nodal measurement snapshots；and is able to identify radial and mesh networks. Its computational efficiency meets the requirement of online applications. Finally，the proposed method was validated through IEEE 33-node and PG&E 69-node distribution grids using simulation data. The sensitivity to different noise levels，missing data，and unknown topologies was further analyzed.
Keyword_Chinese:智能配电网；拓扑辨识；机器学习；LightGBM；深度神经网络
Keywords_English:smart distribution grids；topology identification；machine learning；LightGBM；deep neural networks

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6514