基于数据驱动的配电网状态估计及拓扑辨识研究

【作者】 罗杰；

【导师】 刘敏； 李海林；

【作者基本信息】 贵州大学 ， 电气工程， 2025， 硕士

【摘要】 随着能源转型的不断推进,清洁能源的比例逐步提高,配电网面临着更加复杂的挑战。尤其是分布式能源的接入和不可控负荷的增加,使得配电网的运行不确定性显著提升。为了确保电力供应的高质量和高可靠性,传统的配电网规划、调度和运维方式已难以充分满足现代电力系统的需求。因此,如何准确实时地估计配电网的状态、识别其拓扑结构并进行优化管理,成为当前电力系统面临的重要课题。这一背景下,线损计算、配电网状态估计(Distribution System State Estimation,DSSE)与拓扑识别的有机结合显得尤为重要。线损计算能为状态估计提供基础数据,状态估计为拓扑识别提供动态信息,而拓扑识别则为线损计算和状态估计提供结构化的电网模型。通过三者的协同工作,能够实现对配电网运行状态的全面感知、精准分析与智能优化,提升电网的管理效率和运行水平,从而有效保障电网的稳定性与高效运行。然而,由于隐私等原因的限制,中低压配电网的研究通常缺乏真实开放的数据和测试模型,这对研究工作的推进构成了较大的挑战。因此研究中低压配电网测试网络模型法、精确的线损计算、状态估计与拓扑识别方法具有重要意义。本文首先全面回顾了国内外关于低压测试网络模型、台区线损计算、状态估计以及台区拓扑识别的研究现状,剖析了现有研究的局限性,并在此基础上提出了本文的创新性解决方案。然后,针对中低压两级网络因缺乏开放细节数据和测试模型而限制技术开发与升级的问题,本文系统地提出了一种基于智能电表数据的两级优化网络模型构建方法。通过特征工程和基于综合相对重要性(CRiteria Importance Through Intercriteria Correlation,CRITIC)赋权的模糊C均值(Fuzzy C-Means,FCM)聚类方法有效区分智能电表用户的用电特性,并结合禁忌搜索(Tabu Search,TS)改进的二进制粒子群(Binary Particle Swarm Optimization,BPSO)优化算法,求解考虑功率约束和用户类别约束的优化模型,为网络模型提供了一种高效且灵活的负荷建模解决方案,为后续基于数据驱动的研究奠定了数据基础。最后,结合IEEE33节点算例,通过仿真分析验证了本文所提方法的有效性。其次,本文考虑将用户侧智能电表的实时用电数据作为台区动态线损的特征量进行建模,利用FT-Transformer(Feature Tokenization-Transformer)实现了高效精准的实时线损计算。随后,对模型计算结果进行了过拟合分析,并与门控循环单元(Gated Recurrent Unit,GRU)、长短期记忆网络(Long Short-Term Memory,LSTM)等多种模型进行了实时线损计算误差指标的对比。结果表明,FT-Transformer表现最佳,相对计算误差最小,能够实现高精度和高粒度的台区线损实时计算,为台区运行管理提供了有力支持。在此基础上,对于中压网络,提出了利用实时线损数据和智能电表聚合有功功率进行伪量测建模来进行DSSE,从而提升DSSE精度的方法。通过在IEEE33节点进行仿真对比分析,验证了该方法对于提升DSSE电压幅值和相角结果精度的有效性。最后,实现了台区拓扑中对户变和相别关系的自动识别。通过采用户节点一周的电压日冻结曲线进行特征工程,有效避免了因短时间内数据波动产生的偶然性误差。随后使用均匀流形逼近与投影(Uniform Manifold Approximation and Projection,UMAP)降维算法对特征进行降维处理,既保留了电压波动的基本特性,又提高了聚类效率。采用基于层次密度的空间应用聚类算法(Hierarchical Density-Based Spatial Clustering of Applications with Noise,HDBSCAN),无需预先指定簇数,减少了人为干预的同时提高了拓扑识别的精度和鲁棒性,实现了户变和相别关系的有效识别。更多还原

【Abstract】 As the energy transition continues to advance and the proportion of clean energy gradually increases,distribution networks are facing more complex challenges.In particular,the integration of distributed energy and the increase in uncontrollable loads have significantly heightened the operational uncertainty of distribution networks.To ensure high-quality and highly reliable power supply,traditional distribution network planning,dispatching,and operation and maintenance methods are no longer sufficient to meet the needs of modern power systems.Therefore,accurately and real-time estimating the state of the distribution network,identifying its topological structure,and optimizing its management have become crucial issues for current power systems.In this context,the integration of line loss calculation,Distribution System State Estimation(DSSE),and topology identification is of paramount importance.Line loss calculation provides foundational data for state estimation,state estimation provides dynamic information for topology identification,and topology identification offers a structured grid model for both line loss calculation and state estimation.Through the collaboration of these three components,it is possible to achieve comprehensive awareness,precise analysis,and intelligent optimization of the distribution network’s operating state,improving the grid’s management efficiency and operational level,thereby effectively ensuring the stability and high-efficiency operation of the network.However,due to restrictions such as privacy concerns,research on medium and low voltage distribution networks often lacks access to real,open data and test models,which poses significant challenges to advancing the research.Therefore,methods for constructing medium and low voltage distribution network test models,precise line loss calculation,state estimation,and topology identification are of great importance.In this paper,we first comprehensively review the current research status of domestic and international studies on low-voltage test network modeling,line loss calculation,state estimation,and station topology identification,analyze the limitations of the existing studies,and propose innovative solutions in this paper on this basis.Then,to address the limitations faced by the two-level network due to the lack of open data and test models that restrict technological development and upgrades,this paper systematically proposes a two-level optimization network model construction method based on smart meter data.Through feature engineering and CRITIC(CRiteria Importance Through Intercriteria Correlation)-weighted Fuzzy C-Means(FCM)clustering,the method effectively distinguishes smart meter users.By combining the Tabu Search(TS)-improved Binary Particle Swarm Optimization(BPSO)algorithm,an optimization model considering power and user category constraints is solved.This provides an efficient and flexible load modeling solution for the network model and lays a data foundation for subsequent data-driven research.Finally,simulation results based on the IEEE33-node test case validate the effectiveness of the proposed method.Next,this paper considers using real-time electricity consumption data from user-side smart meters as dynamic line loss features for feeder modeling.The FT-Transformer(Feature Tokenization-Transformer)is employed to achieve efficient and precise real-time line loss calculation.Subsequently,the model calculation results were analyzed for overfitting and compared with various models such as Gated Recurrent Unit(GRU)and Long Short-Term Memory(LSTM)for real-time line loss calculation error indicators.The results show that the FT-Transformer performs the best,with the smallest calculation error,achieving high precision and fine-grained real-time line loss calculation,thus providing strong support for feeder operation management.Building on this,for medium-voltage networks,a method is proposed to enhance DSSE accuracy by utilizing real-time line loss data and smart meter-aggregated active power for pseudo-measurement modeling.Simulation comparisons on the IEEE33-node test case demonstrate the effectiveness of this method in improving the accuracy of DSSE voltage magnitude and phase angle results.Finally,an automatic recognition method for user connections and phase relationships within the feeder topology is implemented.Feature engineering is conducted by using a week-long voltage day-freezing curve of user nodes,effectively avoiding incidental errors caused by short-term data fluctuations.Uniform Manifold Approximation and Projection(UMAP)is then applied for dimensionality reduction,which not only retains the essential characteristics of voltage fluctuations but also improves clustering efficiency.The Hierarchical Density-Based Spatial Clustering of Applications with Noise(HDBSCAN)algorithm is used,which does not require pre-specifying the number of clusters,thus reducing human intervention while improving the accuracy and robustness of topology recognition,enabling effective identification of customer and phase relationships.更多还原