【作者】 庞清乐；

【导师】 孙同景；

【作者基本信息】 山东大学 ， 控制理论与控制工程， 2007， 博士

【摘要】 国内外中压配电网中性点广泛采用小电流接地(包括不接地、经消弧线圈接地和经高电阻接地)方式，以避免发生单相接地故障(又称为小电流接地故障)时跳闸造成供电中断。对于小电流接地故障，由于故障电流微弱、电弧不稳定和随机因素影响等原因，接地故障选线比较困难，一直缺乏可靠的故障选线方法和高准确度的小电流接地故障选线装置，至今许多变电站仍然使用人工拉路方法查找故障线路。随着人们对配电网自动化水平要求的提高，小电流接地故障自动选线问题更加突出，迫切需要从根本上予以解决。因此，研究高准确度、高可靠性的自动选线技术和研制相应自动选线装置，对于提高供电可靠性、减少停电损失和提高配电自动化水平具有重要的意义。本文在分析小电流接地系统单相接地故障特征基础上，分别对基于故障暂态特征信息的选线和基于故障暂态和稳态特征信息的融合选线进行了深入的研究。首先，利用故障暂态特征信息，提出了基于粗糙集理论和小波包分析的故障选线方法。其次，利用故障暂态和稳态特征信息，提出了基于粗糙集理论的融合选线方法。再次，由于基于粗糙集理论的融合选线方法的选线准确度较低，提出了基于神经网络的融合选线方法。然后，针对基于神经网络的融合选线模型训练时间较长的问题，提出了基于粗糙集理论的样本归一化方法。最后，以基于神经网络的融合选线为判据，设计出了小电流接地故障智能选线装置的方案。主要的研究工作如下：(1)分析了小电流接地系统单相接地故障的特征。通过对小电流接地系统发生单相接地故障时零序电流的稳态特征、暂态特征以及谐波特征的分析，确定了在进行小电流接地故障选线时应选取零序电流的暂态特征、有功分量特征、五次谐波特征和基波特征作为故障特征。(2)分析了基于小波包分析的小电流接地故障选线方法存在的问题，在此基础上，提出了基于粗糙集理论和小波包分析故障选线新方法。基于小波包分析的小电流接地故障选线方法以暂态零序电流为故障特征，因为暂态分量的幅值比稳态分量大很多，所以该方法具有较高的选线准确度。由于受硬件电路的限制，对暂态信号的采样频率不可能太高。经过低频采样后的暂态零序电流信号的幅值会发生不同程度的衰减，当采样后故障线路的暂态零序电流信号幅值衰减严重时，基于小波包分析的故障选线方法可能出现选线错误。所以需要对衰减的信号进行增强处理，衰减越严重、越可能是故障线路的信号，需要增强的比例就要越大。本文提出了采用粗糙集理论对信号增强的小波包选线新方法。首先，分别对暂态零序电流信号进行低频采样和短时间的高频采样，将从高频采样和低频采样的零序电流信号中提取的稳态工频分量幅值、采样前后信号衰减比例和信号首波头极性作为故障特征。然后，将这些故障特征作为条件属性，信号需要增强的比例系数作为决策属性，构建决策表。通过属性约简和值约简得到最小决策规则，根据最小决策规则实现对低频采样信号的增强。最后，对增强的低频采样信号利用小波包故障选线方法进行故障选线。仿真结果表明，该方法无论是在能量衰减严重或首波头极性检测错误时，还是在相电压过零时的故障、母线故障和高阻接地故障，均能实现正确选线，有效地提高了故障选线的准确度。(3)定义了小电流接地故障时各种故障特征的故障测度，在此基础上，提出了基于粗糙集理论的融合选线方法。分别采用小波包分析法、有功分量法、五次谐波法和基波幅值法从零序电流信号中提取暂态分量、有功分量、五次谐波分量和基波分量作为故障特征，根据每种故障特征的特点分别定义了它们的故障测度。将这四种故障特征的故障测度作为决策系统的条件属性，线路的故障状态作为决策系统的决策属性，提出了基于粗糙集理论的融合选线方法。通过属性约简和值约简后，得到最小决策规则，根据该规则实现故障选线。该方法消去了冗余的基波分量故障特征，融合了对选线结果影响较大的暂态分量、有功分量和五次谐波分量故障特征，实现了融合选线的目的。利用仿真和现场数据对该选线方法进行测试的结果表明，该方法的选线准确度高于基于单一故障特征的选线方法。(4)针对基于粗糙集理论的融合选线方法选线准确度不高的缺点，将经粗糙集理论约简得到的故障特征的故障测度作为神经网络的输入，将融合后的故障测度作为神经网络的输出，提出了基于神经网络的融合选线方法。利用仿真和现场数据对该选线方法进行测试的结果表明，基于神经网络的融合选线方法的选线准确度高于基于粗糙集理论的融合选线方法。(5)针对神经网络的训练样本中不同类样本距离较近时神经网络训练时间较长的问题，提出了基于粗糙集理论的神经网络样本归一化方法，并将其应用于小电流接地故障选线。首先，将神经网络的输入作为条件属性，输出作为决策属性，构建决策表。然后，计算决策表中不同类样本间的最小距离，根据该最小距离确定原样本需要伸缩的比例，最小距离越小，伸缩的比例越大。各样本根据各自的伸缩比例进行伸缩处理，然后再将样本归一化。最后，利用归一化处理后的样本训练和测试神经网络。以基于神经网络的小电流接地故障选线为例，对该算法进行了验证，测试结果表明，采用该样本归一化方法的神经网络的训练时间明显缩短。(6)利用电磁仿真软件EMTP-ATP建立了小电流接地系统的仿真模型。根据各种故障情况做了大量的仿真试验，仿真数据可以用于各种故障选线方法的分析。(7)根据基于神经网络的融合选线判据，设计出了小电流接地故障选线装置的总体结构方案、主要的硬件电路设计和主程序软件流程。最后对全文进行了总结，并对下一步的研究工作进行了展望。更多还原

【Abstract】 Many countries, including China, adopt non-effectively grounding (isolated, Pe-tersen coil or high resistance earthed) neutral in medium voltage distribution networks to reduce outages caused by single-phase-to-earth fault (small current earth fault). When single-phase-to-earth fault occurs, it is difficult to detect the line with an earth fault due to weak current, unstable fault arc and stochastic factors. To date electric utilities still use manual switching to detect an earthed line in the absence of a more reliable fault line detection method and a more accurate device for detecting earth fault line. The problem of fault line detection for single-phase-to-earth fault becomes more and more prominent and it is urgent to resolve this problem thoroughly in order to meet the requirements of distribution automation. Therefore, the study of the high accuracy and high reliable automatic fault line detection technology and the corresponding device is significant to the improvement of the power supply reliability, the reduction in the outage loss and the improvement of distribution automation system.Based on analyzing the fault characteristics for single-phase-to-earth fault in non-effectively grounding neutral system, the paper studies the properties of the fault line detection based on fault transient information, and the integrated fault line detection based on fault transient information and fault steady information, respectively. First, applying the fault transient information, the fault line detection based on rough set theory and wavelet packet analysis is proposed. Next, applying fault transient information and fault steady information, the paper proposes the integrated fault line detection based on rough set theory. Next, the integrated fault line detection based on neural networks is presented because of the lower accuracy of the integrated fault line detection based on rough set theory. Next, in order to avoid the longtime training of the fault line detection model based on neural networks, a novel sample normalization algorithm based on rough set theory is proposed. Finally, applying the integrated fault line detection criterion based on neural networks, the design scheme of the device for detecting grounded line in non-effectively grounding neutral system is developed. The main researches are as follows:(1) The fault characteristics for single-phase-to-earth in non-effectively grounding neutral system are analyzed systematically. Through analyzing the steady characteristics, transient characteristics and harmonic characteristics of zero sequence current, the paper reveals that the transient characteristics, active component characteristics, the fifth harmonic characteristics and fundamental characteristics of zero sequence current should be selected as fault features to perform fault line detection.(2) The problems of the fault line detection method based on wavelet packet analysis for single-phase-to-earth fault in non-effectively grounding neutral system are analyzed. Therefore, a novel fault line detection method based on rough set theory and wavelet packets analysis is proposed. The fault line detection based on wavelet packet analysis, which selects transient component of zero sequence current as fault feature, provides high accuracy of fault line detection due to the magnitude of transient component of zero sequence current is much bigger than the steady component. The sampling frequency rate cannot be very high due to the limitation of the hardware circuit. The magnitude of transient signal sampled by lower sampling frequency rate will decay in different degree. When the magnitude of the sampled transient zero sequence current is attenuated rigorously, the result of fault line detection may be wrong. So the sampling signal must be enhanced. The more rigorously the transient signal is attenuated and the larger the possibility of fault is, the bigger the transient signal enhancement ratio is. The paper proposes a novel fault line detection method based on rough set theory and wavelet packet analysis. Firstly, the transient zero sequence current signals are sampled in lower sampling frequency rate and in higher sampling frequency rate for a short time, respectively. The amplitude of fundamental component, decaying ratio of amplitude between before and after sampling and polarity of initial wavefronts are extracted from the zero sequence current signal sampled in lower sampling frequency rate and in higher sampling frequency rate. Then, the fault features extracted are selected as condition attributes and the signal enhancement ratios are selected as decision attributions and an information system is constructed. Af- ter attribute reduction and attribute value reduction, the minimum solution of decision rules can be obtained. These rules can be used to enhance the low frequency sampling signals. Finally, the enhanced low frequency sampling signals are decomposed by wavelet packet to realize fault line detection. The simulation results show that the proposed method can detect the fault line when the magnitude based criterion or the polarity based criterion is invalid, or when the phase voltage is close to its crossover point, or when the bus grounding fault or the high earth resistance grounding fault occurs. So the method improves markedly the accuracy of fault line detection.(3) The fault measures of various fault features for single-phase-to-earth in non-effectively grounding neutral system are defined. Therefore, the integrated fault line detection method based on rough set theory is presented. The fault features of transient component, active component, the fifth harmonic component and fundamental component are extracted from zero sequence current through wavelet packet analysis method, zero sequence current active component method, the fifth harmonic current method and fundamental current component amplitude comparison method, respectively. Then they are transformed into fault measures according to their characteristics. The integrated fault line detection method based on rough set theory, in which the condition attributes are these fault measures and the decision attribute is the integrated fault measure, is proposed. Having performed attribute reduction and attribute value reduction, the minimum solution of decision rules can be obtained. These rules can be used to realize fault line detection. This method reduces these redundant attributes and integrates the fault features of transient component, active component and the fifth harmonic component, which have more effect on fault detection result, and realizes the integrated fault line detection. The fault line detection method is verified by the simulation data and field data for single-phase-to-earth fault and the testing results show that the method can provide the higher accuracy of fault detection than the fault line detection method based on single fault characteristic.(4) To overcome the problem that the accuracy of the fault line detection based on rough set is low, the integrated fault line detection method based on neural network is proposed, in which the inputs are these fault measures reduced by rough set theory and the output is the integrated fault measure. The fault line detection method is verified by the simulation data and field data for single-phase-to-earth fault. The testing results show that the method can provide higher accuracy of fault line detection than the fault line detection method based on rough set theory.(5) A novel sample normalization algorithm based on rough set theory is proposed to avoid the longtime training of neural networks classifier caused by the smaller distances between samples of different classes. After the inputs of neural networks are selected as the condition attributes and the outputs of neural networks are selected as decision attribute, the decision table is constructed. Then the minimal distances between different samples of different classes in the decision table are calculated. According to these minimal distances, the samples are extended or contracted. The smaller the minimal distance is, the bigger the radio of extension or contraction is. After all the original samples are extended or contracted, they are normalized. Finally, the normalized samples are used to train and verify the neural network. The method is analyzed with an example of fault line detection in non-effectively grounding neutral system. The simulation results show that the training time of neural networks with the sample normalization method is shortened markedly.(6) Using Electromagnetic Transients Program EMTP-ATP, a lot of experiments of single-phase-to-earth fault for all kinds of fault types are carried out. The simulation data can be used to analyze every fault line detection method.(7) According to the integrated fault line detection criterion based on neural networks, the design scheme of the device for detecting grounded line in non-effectively grounding neutral system is introduced. Meanwhile, the block diagram of its main hardware and the flow of its main software are designed.更多还原