《电气工程毕业设计翻译_小波包神经网络在电力系统继电保护中的应用》.doc
西华大学毕业设计外文资料翻译 译文：小波包神经网络在电力系统继电保护中的应用摘要，本文提出解决继电保护测试仪波形畸变问题的小波包神经网络（WPNN）方法。凭借其良好的时频局部化的逼近能力，WPNN是用来建立一个对继电保护测试仪非线性放大器的识别模型。有待放进工具的错误数据是用来被识别模型的调节功能补偿的，使整个使用仪器系统显示表现得线性，以便使产生的波形失真大大地被限制。模拟结果表明，提出方法具有可行性和有效性，原型已进入实际运行。1 介绍现代电力网络的规模和复杂性不断扩大，它要求配置较高可靠性的电力系统继电保护，错误数据在进入继电保护测试仪运算之前被放大器放大是改善他们性能的有效方式12。传统的继电保护测试仪器能够实现这样的测试功能，但他们曾经采用模拟放大器，这是一个典型的非线性系统，实现功率放大。因此，输出波形的非线性失真不可避免地成为继电器保护测试的严重问题。在本文中，WPNN方法提出了解决这一问题的方法。WPNN是小波包理论和神经网络，它不仅具有小波包的良好的局部化性质和特征提取能力，但也继承神经网络的许多优点，如自适应性和最高兼容性等34。它选择小波基为神经元的激活函数，并规范设计程序和稳固学术基础，所以WPNN已广泛应用在许多技术领域567。在这项研究中，WPNN是采用了建立继电保护测试仪的非线性放大器的识别模型，并通过比较识别模型与理论输出，自适应调整功能是在错误数据被送进仪器之前，这使得整个仪器系统将显示变得线性，使输出波形的失真限制很大。用故障仿真数据进行记录，其结果证明了可行性和WPNN应用对电力系统继电保护测试的有效性，以及所提出的方法样机已投入实际运行。2 建设WPNNWPNN是小波神经网络的升级（WNN）。WNN小波神经网络可以被看作是使用的的正交小波变换多分辨率分析（MRA）的基础上重建的组合空间的小波8910。大家都知道，小波空间可进一步利用小波包分解，使信号可以分解为更多的频段，来比MRA提高频率分辨率。因此，选择最佳小波包基由于网络神经元的激活功能将获得更好的网络时频局部化性质和逼近能力。所以WPNN利用小波包基的输入信号特征提取和神经网络在WPNN识别信息。WPNN可分为两部分：小波包特征提取和神经网络信息识别，这是图1所示。通过本文，Z表示所有整数的集合。让 和表示波基和小波从各个的产生。对WPNN结构设计包括以下三个主要步骤：第1步。规模计算范围：用和来表示和目标系统 时间范围，他们的精力集中地区的频率范围内可被看做实验数据，这是表示为分开的和分开。根据傅里叶变换的性质，随着扩大的小波，频率范围将会扩大到,即的频率扩大为。因此，小波尺度包含一个用于覆盖有限的范围，它可以通过以下计算：和分别表示较小或更大的整数第2步。选择最佳小波包基：Shannon准则引入到计算的尺度范围内的节点组系数由第一步得到。然后，如果子的总和比父节点少，两个子节点取代它下面的节点的直接父节点。在这个方法中，我们可以在最小的基础上设定，可以如下表示：其中E是最佳小波包基数第3步。的节点数目的测定：这一步也是可以被看作是转换因子为每小波尺度的。这是被称为，小波包的时间范围 是不随着N的变化而变化的。所以小波包的时间范围基 可表示为 ，在时间轴上滑动的程度，随着的增加或减少。由于覆盖的时间范围 ，的范围确定 由以上，结构和WPNN第一部分参数的三个步骤可以肯定确定。因此，第二部分可以作为一个简单的三层已知的输入值，其连接的神经网络。整体结构的WPNN因此以如下形式，并在图1所示 3.总体方案的继电保护测试仪所提到的介绍，输出波形的非线性失真是继电保护测试最严重的问题。针对这个问题，一个闭环继电保护测试仪的新计划，提出在图2所示。 双CPU的配置，包括上层控制器和较低的放大器适用于本系统。上层控制器采用高性能便携式计算机或嵌入式计算机为核心，实现了数据采集，故障分析和综合控制。此外，还可以调整采样频率，释放速度或根据测试数据谐波含量的要求输入。数字和模拟测试软件防护服成功嵌入到仪器。它可以模拟前的数字平台，提高了灵活性和可重复性，避免潜在的伤害，测试设备11。较低的放大器，主要包括数字信号处理（DSP）芯片，智能功率模块（IPM）的，反馈电路阵列。 DSP的形式接收数据通过CAN总线上控制器的计算机，并产生PWM（脉宽调制）通过定期抽样方法脉搏，IPM是由PWM脉冲驱动去实现功率放大。反馈电路设计输出信号进行采样，组成闭环配置，这主要是考虑到了幅值和极性转换。为了消除非线性失真，一个使用基于硬件的数字闭环修正算法12，可描述如下：查明实验数据较低的放大器的部分，建立仪器仪表系统的投入产出模型。通过比较模型的识别与理念输出，产生调节作用，引导之前，都要输入到仪器，使输出波形可以最远的理想值的方法使数字区的故障数据自适应调整。很显然，系统准确识别算法具有重要义，WPNN可用于完成此任务，因为其良好的时频局部化性质和逼近能力。4 程序的算法与WPNN该数字闭环与WPNN修改程序显示在图3，可以这样解释：一些在有效范围内随机取样点输入到建议的配置和输出波形记录仪的实际使用反馈电路。该小组的组成由采样数据及其相应的反馈被视为实验数据集。该小组的组成由采样数据及其相应的反馈被视为实验数据集。一个识别模型，建立了作为未知非线性性能的放大器的算法代替实验集的数据。准确的系统识别和调节功能的获得是两个算法的关键点。凭借其良好的时频局部化的逼近能力，WPNN用于建立该系统的辨识模型。选择一个合适的母小波函数和估计的非线性性能，频率域与实验数据集。WPNN的网络结构和神经元数量可确定在第二部分提出的方法，和WPNN连接权可以训练一些优化算法，如反向传播（BP），遗传算法（GA）等。调节功能得到了迭代修正方法，如图3所示。表示数据的故障点，某些数据被输入到仪器和所确定的输出模式扩增。和的不同点和放大价值观点, 是这个观念的放大因素，被用来适应从到的原始数据，然后设置是初始点，重复以上过程直至为记录到调整值的形式，最后将被输入到测试仪，实现故障波形放大。该算法本质上是一种用于放大器，使仪器系统显示在整个线性特性的非线性性能补偿方法，使输出波形的非线性误差可大大减少。5 模拟结果为了验证应用在电力系统继电保护测试WPNN成效，仿真模拟实验中使用了江西省某地区实际过失录得的数据。按照上面，建立一个识别模型的基础上使用WPNN训练数据和相关的补偿值都可以通过闭环修改，这是绘制在图4中提到采样数据计算出来的程序。结果表明，该模型能准确地识别近似模拟非线性性能，其跟踪误差在0.1以内。图5显示了初步的模拟输入数据段和它的补偿值调整的过程。最初的数据是相当前操作系统故障，其最大达到10A电流。而在峰值或谷点，输入数据有较大的更严重，因为非线性衰减补偿值。输出波形的比较和没有在本文提出的方法如图6所示。从波形分析结果表明。1）由于非线性放大器的性能，将不可避免地进入失真的输出波形，继电保护测试可能导致错误的结论通2）过使用系统识别和闭环修改，输出波形均方根误差从2.09降低到0.76。这样的失真限制的输出波形可以准确地模拟电源故障3）补偿功能是最显着，尤其是在接近峰值或谷值点6 结论（1）一种新型的神经网络，WPNN，最佳小波作为神经元的激活函数包的基础是在本文中介绍了，它具有精确的系统结构设计和规范的程序实施的性能。（2）在这项研究中，WPNN求解继电保护问题测试仪器输出波形的变形，仿真结果证明其可行性和有效性，并与该算法的原型现在已经投入实际操作。（3）WPNN已逼近复杂的非线性系统的优异性能，所以它也可以应用到其他模型或在电力系统优化问题模式识别，故障诊断，负荷预测和数据压缩等。参考文献1 Jodice, J.A.:：继电器性能测试：电力系统继电保护委员会出版。有关功率传输12（1997）169-171汇刊2 Sachdev, M.S., Sidhu, T.S., McLaren, P.G.:：问题和数值继电器测试机会。 IEEE电力工程学会夏季会议（2000）1185年至1190年3 Benediktsson, J.A., Sveinsson, J.R., Ersoy, O.K., Swain, P.H.:小波包并行双人工神经网络神经5（1995）5-13工程系统4 Avci, E., Turkoglu, I., Poyraz, M.:：智能目标识别神经网络的小波包。专家系统的应用29（1）（2005）175-1825 Zhou, Z.J., Hu, C.H., Han, X.X., Chen, G.J.:：自适应小波包神经网络的放大器的故障诊断。第二届国际研讨会论文集神经网络（2005）591-596 6 Wang, L., Teo, K.K., Lin, Z.:：预测时间序列与小波包神经网络。国际神经网络联席会议3（2001） 1593年至1597年 7 Schuck Jr., A., Guimaraes, L.V., Wisbeck, J.O.:：Dysphonic语音分类中的应用小波包变换与人工神经网络。一年一度的国际会议的IEEE工程医学和生物学3（2003）2958年至2961年8 Zhang, Q.: Benveniste, A.:：小波网络。电机及电子学工程师联合会神经网络3（6）（1992）889-8989 Gao, X.P.:：一个基于小波神经网络的比较研究。的第九届国际神经信息处理会议（2002年）1699至1703年 10 Zhao, X.Z., Ye, B.Y.:：对电机振动噪声的自适应小波神经网络的信号识别。第三次国际研讨会神经网络2（2006）727-734 11 Meliopoulos, A.P.S., Cokkinides, G.J.:一种继电保护评估测试虚拟环境系统会刊（2004）104-11112 Sun, X.M., Du, X.W., Liu D.C., Cai, X.:：障碍复发的扩音设备，基于数字闭环改性技术。电气自动化电力系统28（4）（2004）49-53原文：Application of Wavelet Packet Neural Network on Relay Protection Testing of Power SystemAbstract. The paper presents a wavelet packet neural network (WPNN) approach for solving the waveform distortion problem of protective relaying testing instrument. With its excellent time-frequency localization property and approximation ability, WPNN is used to establish an identification model of the non-linear amplifier of the protective relaying testing instrument. The fault data to be put into the instrument is compensated by an adjusting function getting from the identification model, which makes the whole instrumentation system show linear performance so that the distortion of the output waveform is constrained greatly. Simulation results indicate the feasibility and validity of the proposed approach, and a prototype has been put into practical operation.1 IntroductionThe continuous expansion of the modern electric networks scale and complication of its configuration requires higher reliability of protection relays in power system, and testing protection relays with fault recoding data amplified by instrument before putting into operation is an effective way for improving their performance 1, 2. Traditional protective relaying testing instruments could realize such testing function, but they used to adopt analog amplifier, which is a typical non-linear system, to realize power amplifying. So the non-linear distortion of output waveform inevitably be comes a serious problem for the relay protection testing. In the paper, a WPNN ap proach is presented for resolving this problem. WPNN is a combination of wavelet packet theory and conventional neural network, which not only possesses good localization property and feature extraction ability of wavelet packet, but also inherits most merits of neural network such as selfstudy, adaptability and high fault-tolerant 3, 4. It selects wavelet packet basis as its neurons activation function and has normative design procedures and solid academic foundation, so WPNN has been widely applied in many technical fields 5, 6, 7. In this study, WPNN is adopted to establish an identification model of the nonlinear amplifier of the protective relaying testing instrument. And by comparing the identification models output with idea output, an adjusting function is generated to guide adaptive adjustment of fault data before to be put into the instrument, which makes the whole instrumentation system show linear performance so that the distortion of the output waveform is constrained greatly. A simulation using fault recording data is carried out, whose results demonstrate the feasibility and validity of application of WPNN on relay protection testing of power system, and a prototype with the proposed approach has been put into practical operation.2 Construction of WPNNWPNN is the development of wavelet neural network (WNN). WNN can be viewed as the combination of reconstructions using wavelet basis of orthogonal wavelet spaces of based on multi-resolution analysis (MRA) 8, 9, 10. As everyone knows, wavelet space can be decomposed further using wavelet packet, so signals can be decomposed in more frequency bands to increase frequency resolution than by MRA. Therefore, selecting best wavelet packet basis to be network neurons activation function will obtain better time-frequency localization property and approximation ability for the network. So WPNN utilizes wavelet packet basis extracting feature of input signal and neural network in WPNN takes charge of information identification, i.e., WPNN can be divided into two parts: wavelet packet feature extraction and neural network information identification, which is shown in Fig.1.Throughout the paper, Z denotes the set of all integers. Let 和denote wavelet basis and wavelet packet generated from respectively. The structure design of WPNN consists of following three primary steps:Step 1. Calculating scale range: Using and to denote the time extent of and the goal system , their energy concentrating areas of frequency extent can be estimated with training data, which are expressed as and separately. According to the properties of Fourier transform, with the increase of the wavelet scale j, frequency extent will expand by , i.e., frequency extent of is . Therefore the wavelet scale j contains a finite range for covering ,and it can be calculated by below:Where and denote choosing smaller or bigger integer value nearby respectively.Step 2. Selecting best wavelet packet basis: Shannon entropy criterion is introduced to calculating the entropies for the set of coefficients of each node in scale range getting in step1. Then, replace the parent nodes by the two children nodes directly below it if the sum of childrens entropies is less than that of parent. In this method, we can uncover the set of minimum entropy basis, which can be denoted as follows:Where E is the number of best wavelet packet basis.Step 3. Determination of number of nodes: This step is also can be seen as determination of translation factor k for each wavelet scale j. It is known as that the time extent of wavelet packet ( ) is invariable with n changes, so the time extent of wavelet packet basis can be expressed as .With the increase or decrease of k, the extent slides on the time axis. For covering the time area of , range of k is determined as:By the three steps above, the structure and parameters of first part of WPNN (feature extraction) can be definitely determined. So the second part (information identification) can be viewed as a simple three-layered neural network with known input value, whose connection rights w( n , j , k )are also that of WPNN. The whole structure of WPNN is thus of the following form, and is illustrated in Fig.1.3 Overall Scheme of Relay Protection Testing InstrumentAs referred in introduction, the non-linear distortion of output waveform is the most serious problem for relay protection testing. Aiming at this problem, a new scheme of closed-loop relay protection testing instrument is proposed as shown in Fig.2. Double CPUs configuration including upper-controller and lower-amplifier is applied in this system.Upper-controller adopts high-performance portable computer or embedded computer as its core, which realizes data acquisition, fault analysis and integrated control.Besides, it can also adjust sampling frequency, value, releasing speed or harmonic content of the input data according to the requirements of testing. And a suit of protection testing digital simulation software is successfully embedded into uppercontroller of the instrument. It can simulate the testing process before analog testing on the digital platform, which improves the flexibility and repeatability and avoids potential harm to the tested equipments 11. Lower-amplifier mainly consists of Digital Signal Processing (DSP) chip, array of Intelligent Power Modules (IPM), and feedback circuit. DSP receives data form upper-controller computer through CAN bus and generates PWM (Pulse WidthModulation) pulse by regular sampling method, and IPM is drove by the PWM pulse to realize power amplification. Feedback circuit is designed to sample the output signals to compose closed-loop configuration, which mainly takes charge of the transformation of amplitude and polarity.For eliminating non-linear distortion, an algorithm of digital closed-loop modification is used based on the proposed hardware 12, which can be described as follow:Identify the lower-amplifier part with training data and establish an input-output model for the instrumentation system. By comparing the identification models output with idea output, an adjusting function is generated to guide adaptive adjustment of fault data in numeric area before being to be input to the instrument, so that the output waveform can furthest approach to ideal value. It is clear that accurate identification of system is of great importance in the algorithm, and WPNN can be applied to complete this task because of its excellent time-frequency localization property and approximation ability.4 Procedure of the Algorithm with WPNNThe procedure of digital closed-loop modification with WPNN is shown in Fig.3,which can be explained like that: Some random sampling points within the effective range are input to the actual instrument with proposed configuration and the output waveform is recorded using the feedback circuit. The group composing by the sampling data and their corresponding feedback is regarded as training data set. An identification model is established by the training data set as substitute of unknown non-linear performance of instruments amplifier in the algorithm. And then compare the output of to the idea output, and construct an adjusting function to compensate the initial to be put into the instrumentation system for realizing the goal of constraining distortion of output waveform greatly.Accurate system identification and acquirement of adjusting function are two the key points of the algorithm. With its excellent time-frequency localization property and approximation ability, WPNN is used to establish the identification model for the system. Select a suitable mother wavelet function and estimate the frequency domain of the non-linear performance with training data set. Network structure and neurons number of WPNN can be determined by the method proposed in the second section, and the connection weights of WPNN can be trained by some optimization algorithm, e.g.,back propagation (BP), genetic algorithm (GA), and etc. And the adjusting function is obtained by the method of iterative modification. As shown in Fig.3, denotes a certain data point of the fault data to be input to the instrument and is its output amplified by the identified model . The difference of and idea amplifying value , where A is the idea amplification factor, is used to adjust the original data to. And then setting as initial point, repeat the process above until meets the precision requirement. The last is recorded into adjusting value form and the last will be input to the testing instrument to realize fault waveform amplification.This algorithm is essentially a compensating method for the non-linear performance of the amplifier, which makes the instrumentation system show linear characteristics on the whole, so that the non-linear error of output waveform can be greatly reduced.5 Simulation ResultsTo testify the effectiveness of
