2024年基于数据驱动的配电网拓扑识别方法研究报告

基于数据驱动的配电网拓扑识别方法研究 马丽副研究员 中国科学院电工研究所 2024年9月21日西安 团队介绍 中国科学院电工研究所电网技术研究部，是国内最早从事分布式能源和微网研究方向的团队之一，现任研究部主任为裴研究员。现有人员50余人，研究方向包括交直流配电网、新型储能系统、新能源电力系统稳定与控制、人工智能在电力系统中的应用、分布式能源市场、综合能源运行优化等。 团队主持了科技部863项目、973课题、国家重点研发专项项目及课题10余项，自然科学基金重点及面上项目20余项，省部级及企业项目60余项，正在牵头中国科学院先导B“新能源主导的电力系统稳定控制新理论新技术”专项。牵头获得了北京市科技进步一等奖，中国仿真学会技术发明一等奖、中国电力科学技术进步一等奖、电力科技创新一等奖等奖项。 个人基本情况 教育背景 ：2012-2017，博士，华北电力大学，电力系统及其自动化（导师：张建华/刘念教授）·2008-2011，硕士，华北电力大学，电力系统及其自动化（导师：张建华/刘念教授·2004-2008,学士，华北电力大学，电气工程及其自动化 工作经历 ■研究方向 ：2021.10－，副研究员，中国科学院电工研究所（裴玮研究员团队：2020.9-2021.8，博士后，斯蒂文斯理工学院（合作导师：吴磊教授IEEEFelloW）：2018.9-2020.8，博士后，威斯康辛大学密尔沃基分校（合作导师：王凌峰教授）：2011.4-2018.8，研发工程师，中国电力科学研究院配电研究所 配电网规划运行优化：拓扑识别、多主体博弃、需求侧管理、本地能源交易等。 社会兼职 长时规模储能重点实验室副主任IEEESeniorMember；中国电机工程学会高级会员；中国电工技术学会会员担任多个国内外期刊审稿人，多次获得IEEETrans期刊年度最佳审稿人 Outline 1、 Background 2. Framework and Method 3、 Case Study 4. Extended Application 5. Conclusion Background Why the observability of topology is weak in distribution network? Large scale &Low monitoring rate Frequent topologychanges Largescaleoffeeders,switchchangescausedbynormaloperation or faults in distribution network;Sensors'coverage rate (like FTU、DTU)are not highThe others are manually maintained, completeness anaccuracy of topology information cannot be guaranteed;ThedevelopmentofDERs(ES,DG,controllableloadEV...) has exacerbated the frequency of switch changes. Topology identification (TI) is conducted to detect topology changes due to faultevents and for some control and optimizationfunctions in normal operation.?中国科学院电工研究所 Background Background https://pcmp.springeropen.com/articles/10.1186/s41601-020-0153-1https://zhuhaiyujian.en.made-in-china.com/product/HslxrohzElkV/China-Advanced-Metering-Infrastructure-AMl-andSystem.htmlhttps://www.giscloud.com/blog/use-case-mapping-and-planning-of-electric-distribution-network/®中国科学院电工研究所 Background Existing studies in topology identification of distribution network Mathematical optimization methods Data-driven methods : Markov Random Field algorithmBayesian-based learning methodsDeep learningLinear regression model Mixed-integer quadratic programmingMixed-integer linear programDistributed sub-gradient algorithm Dalavi F, Golshan M E H, Hatziargyriou N D. A review on topology identification methods and applications indistribution networks[J].Electric Power Systems Research,2024, 234: 110538 Background Open problems The number of topology categories in a batch data sample is difficult toobtain. Synchronous phasor measurement (e.g. μ-PMU) is not widely available The quality of data set is not high due to untimely, incomplete andinaccurate data collection. Computation burden in large scale topology needs to be reduced Outline 1、Background 2. Framework and Method 3、 Case Study 4. Extended Application 5. Conclusion Framework and Method The two-stage topology identification framework StageI:HistoricalDataIdentification Atwo-stagetopologyidentificationframeworkforPDNsisproposedtorecognizethemixedtopologies in historical batch data and predict the real-time topology based on the available nodalmeasurements.P中国科学院电工研究所 Framework and Method Data structure and source Historical Labeled Data:Sample 1: [S11, J11, Tu].....[Sin, Yin, Tn]Sample P: [Spi, Yp1, Tpi].....[Spm ypm, Tpm] Historical Unlabeled Data:Sample 1: [Si, Y'n]......[Sin, J'in]Sample P: [Sp, J'pi]......[Spm, ypm] For each record [sij, yujl, it is the nodal measurements of allnodes in the PDN at a point in time, and Sii includes thevoltageamplitudesandphaseangleswhile vector yijincludesthe activeand reactivepowerinjectionsThetopologycategoryiisdetermined afterthetopologyidentificationprocedure is performed, which is added to the original recordto generate the labeled data [sij, Yij, Tuj] Smart meters FrameworkandMethod Basic math models The parameters in the historical topology identification problem:0 =((T m ,α m),m E [1,2...,M J); Thelogarithmic likelihood function of sample X (including N records [x 1...,x n D) under 0 can be expressed as. p(xi;)is theprobability of xiwithin parameterp(x,Tj; 0) is the probability of x; belonging to T, withinparameter 0, and p(xilTj; 0) is the probability of xi givenT,withinparameter.Qisalsoknownastheconditionaldistributionofthej-thtopologyforthei-threcordinthesample Framework and Method Expectation-maximization (EM) model The EM iteration alternates between the expectation step (E-step)and maximization step (M-step). > The E-step calculates Q(t + 1) (with element Q,(t +1)in the matrix)based onthe current estimateof theparameters o(t). The conditional distribution Q'(t + 1)can also be regarded as theposteriorprobabilityof the i-threcord belongingto thej-th topology category. > The M-step maximizes L*(X, 0; Q(t + 1)) to update theestimateoftheparameter(t+1): Frameworkand Method Split-EM method TheEM-based topology identificationmodelis suitableforthecasewithknownnumberoftopology categories. The number of topology categories inahistorical