【作者】 常雪；

【导师】 娄杰；

【作者基本信息】 山东大学 ， 电气工程（专业学位）， 2023， 硕士

【摘要】 近年来,电力电缆凭借其供电安全、可靠性高等优点被广泛应用于城市配电网中。然而,中压配电电缆在运行过程中,会因自身缺陷、环境影响及外力破坏等导致其局部绝缘性能下降,并可能会在该部位引发多次时间极短且具有自恢复性的电弧故障,本文称此类故障为电缆潜伏性故障,其会加速电缆绝缘劣化,且往往会最终发展为永久性接地故障。但因其故障持续时间短,特征微弱,继电保护装置难以检测到该类故障,常将其视为暂态扰动,这给电缆潜伏性故障的检测识别造成了极大的困难。在此背景下,围绕中压配电电缆潜伏性故障的辨识方法开展研究,主要工作如下:(1)分析电缆潜伏性故障的发展机理及故障特性,在PSCAD/EMTDC平台中搭建了典型的10kV小电阻接地电缆配电网模型,基于Kizilcay电弧模型建立了适用于小电阻接地配电系统电缆潜伏性故障的等效模型,仿真获得的电气量波形特征与电弧故障特征吻合,验证了所提出的电缆潜伏性故障模型的有效性;通过控制变量实验研究了三种模型参数对潜伏性故障特性的影响,为样本数据的生成确立参数依据;搭建了恒阻抗接地故障、电容器投切和负荷突变扰动模块为后续的故障辨识研究提供数据基础。(2)提出了一种基于多维多域特征提取和优化的电缆潜伏性故障的数据驱动辨识模型。对故障电流样本进行时域、频域和时频域的多维特征分析提取,对提取的特征进行有效性验证和主成分分析,以实现故障特征向量的优化;建立了基于极限学习机(Extreme Learning Machine,ELM)的电缆潜伏性故障智能辨识模型,将特征向量作为模型输入,并利用粒子群优化算法对ELM算法的随机初始参数进行优化;与基于其他分类模型的对比证明了本文所建立的模型在分类准确率、训练时间等方面的优势。(3)提出了一种基于数据-知识联合驱动的电缆潜伏性故障辨识模型。介绍了数据-知识联合驱动模型的理论知识,提取了面向电缆潜伏性故障辨识的经验知识;将其与前文建立的基于ELM的数据驱动模型进行结合,利用经验知识对数据驱动算法的规律挖掘提供指导,提出冲突函数的概念并定义其计算规则,将其与损失函数结合构建知识函数,建立了数据-知识联合驱动的电缆潜伏性故障辨识模型;利用粒子群算法对模型参数进行优化,建立随机因子和信任度参数以调节领域经验知识在融合模型中的参与度,降低经验知识出现错误时对联合驱动模型训练产生的不利影响;通过将其与纯数据驱动模型在两种典型运行场景下的多次随机试验中的多个指标进行对比,证明了经验知识的引入可有效提升模型的泛化性、稳定性和安全性。更多还原

【Abstract】 In recent years,power cables have been widely used in urban distribution networks due to their advantages of power supply safety and high reliability.However,during the operation of the medium-voltage distribution cable,its local insulation performance will decrease due to its own defects,environmental impact and external damage,and may cause multiple arc faults with extremely short time and self-recovery in this part.This paper calls such faults as cable latent faults,which will accelerate the deterioration of cable insulation and often eventually develop into permanent ground faults.However,due to its short duration and weak characteristics,it is difficult for relay protection devices to detect such faults,which are often regarded as transient disturbances,which makes it difficult to detect and identify cable latent faults.In this context,the research is carried out around the identification method of latent faults in medium voltage distribution cables,and the main work is as follows.(1)The development mechanism and fault characteristics of cable latent fault are analyzed.A typical 10 kV small resistance grounding cable distribution network model is built in PSCAD/EMTDC platform.Based on Kizilcay arc model,an equivalent model suitable for cable latent fault of small resistance grounding distribution system is established.The waveform of electrical quantity obtained by simulation conforms to the waveform characteristics of arc fault,which verifies the effectiveness of the cable latent fault model built in this paper.The influence of three model parameters on latent fault characteristics is studied by control variable experiments,which establishes the parameter basis for the generation of sample data.A constant impedance grounding fault,capacitor switching and load change disturbance module are built to provide data basis for subsequent fault identification research.(2)A data-driven identification model for cable latent faults based on multi-dimensional multi-domain feature extraction and optimization is proposed.The multi-dimensional feature analysis and extraction of fault current samples in time domain,frequency domain and time-frequency domain are carried out,and the validity verification and principal component analysis of the extracted features are carried out to optimize the fault feature vector.An intelligent identification model of cable latent fault based on Extreme Learning Machine(ELM)is established.The feature vector is used as the model input,and the random initial parameters of ELM algorithm are optimized by particle swarm optimization algorithm.The comparison with other classification models proves the advantages of the model established in this paper in terms of classification accuracy and training time.(3)A cable latent fault identification model based on data-knowledge joint-driven is proposed.The theoretical knowledge of the data-knowledge joint-driven model is introduced,and the empirical knowledge for cable latent fault identification is extracted.Combining it with the data-driven model based on ELM established above,using empirical knowledge to provide guidance for the rule mining of data-driven algorithms,the concept of conflict function is proposed and its calculation rules are defined.Combining it with the loss function to construct a knowledge function,a data-knowledge joint-driven cable latent fault identification model is established.The particle swarm optimization algorithm is used to optimize the model parameters,and the random factor and trust degree parameters are established to adjust the participation of domain experience knowledge in the fusion model,so as to reduce the adverse effect on the joint drive model training when the experience knowledge is wrong.By comparing it with multiple indicators of the pure data-driven model in multiple random tests under two typical operating scenarios,it is proved that the introduction of empirical knowledge can effectively improve the generalization,stability and security of the model.更多还原