【作者】 戴斌；

【导师】 杜林；

【作者基本信息】 重庆大学 ， 电气工程， 2010， 硕士

【摘要】 电力系统过电压直接影响到电力系统的安全运行,是发展超高压和特高压电网所必须研究的重要课题。电力系统中的过电压发生类型多种多样,发生机理不尽相同,波形、幅值、持续时间也不相同。过电压信号携带着丰富的电力系统运行状态信息,利用目前监测到的过电压信号进行特征提取以及识别算法研究,实现过电压类型的自动识别和诊断,对保证电网安全运行具有十分重要的意义。基于目前学界对于过电压类型的划分,本文在充分考虑不同类型过电压之间的从属层次关系后,提出了过电压分层识别的思路。与传统的单层整体识别方式所不同,分层识别算法采用基于不同数学方法的独立分类器,实现对过电压的逐层依次细分。由于各个分类器是相互独立的,可以根据该分类器所需要识别的过电压类型,采取多种数学方式,优化提取特征量及识别算法,从而能有效提高系统的针对性及计算效率。同时,由于采用了模块化思想来设计分类系统,整个分层识别系统易于进行修改,系统扩展性好。本文通过对电力系统中雷电过电压,工频电压升高,谐振过电压及操作过电压的发生机理,发展过程,波形特点的分析,阐述了上述几种过电压在本质特征方面的区别。文中介绍了小波理论,并结合不同过电压信号在时频空间能量分布差异,提出了基于小波时频理论的过电压特征提取方式。本文还介绍了S变换理路和矩阵奇异值分解理论,并从理论上分析了S变换和奇异值分解理论在降低信号随机扰动方面的独特优势,将S变换和奇异值分解理论结合,提出了基于S变换奇异值分解理论的过电压特征提取方式。根据感应,反击,绕击等三种雷电过电压的波形特点,提出了采用输电线路电流行波的时域参数来辨识三种雷电过电压的方法。该方法避开了雷电流行波折反射干扰,具有计算量小,识别方式简单等优点。在提出了多种过电压特征量提取方法后,本文针对性的考虑了分层识别系统各层分类器的不同识别任务,综合考虑各种过电压特征提取方式的优点与实用性,为过电压分层识别系统的各个分类器选择了合理的特征量。最后,本文介绍了支持向量机和粒子群优化算法的基本原理以及特性。针对目前实测过电压数据较少,训练样本不多的情况,提出了采用基于样本风险最小化思想的支持向量机作为识别算法,对过电压进行辨识。系统研究了采用分层结构后,对过电压识别率造成的影响,提出需要对每一个分类器优化其识别算法以保证最终的识别率。针对目前支持向量机参数选取缺乏理论依据的问题,提出了利用粒子群算法来优化选取每层分类器的支持向量机参数。由于粒子群算法对自身初始参数设定不敏感,收敛性好,因此十分适合于优化支持向量机参数。测试计算表明,本文提出的粒子群优化支持向量机算法能有效提高支持向量机的识别率,避免分层识别结构对识别率造成的影响,在同样的样本训练数情况下,识别结果优于单层识别结构系统。更多还原

【Abstract】 Overvoltge is one of the key factors for safety operations of power system. There are varieties of overvoltage types which are caused by different reasons happened in power system ever year. Since the overvoltage signal contains plenty of power system information. Utilizing the data from overvoltage on-line monitoring system to identify the types of the overvoltage automatically is quite beneficial for the self-diagnose and safety design or operation of power system.Based on the currently consensus about overvoltage types classification, a new overvoltage multi-level recognition structure is proposed by this paper by considering the multi-level subordination relationship of different overvoltage types. Unlike the traditional single level recognition structure, the multi-level overvoltage recognition algorithm uses different independent classifier to subdivision the overvoltage type gradually. Since each classifier is independent from each other, different mathematic methods can be used to abstract characteristics parameters and construct classifier algorithm. As a result, each classifier can classify overvoltage pertinently with high efficient and the whole multi-level recognition algorithm is easy to be modified.The happening reason, developing process and waveform characteristics of lightning overvoltage, power frequency overvoltage, operation overvoltage and resonance overvoltage are introduced in this paper. Wavelet theory is adopted to abstract the characteristics of overvoltage energy distribution in time-frequency space. Also, due to the unique advantage of S transform theory and singular value theory in reducing the signal random disturbance, the two theories above are combined together to abstract overvoltage characteristics in this paper. For lightning overvoltage, shield failure overvoltage and back flashover recognition, this paper suggests use front part of transmission line current wave as research object because the front part of current wave can avoid the disturbance of wave reflection. The time domain characteristic parameters of transmission current wave are raised to identify the three kind of lightning overvoltage above. By comprehensively considering the mathematic method above and the task of each classifier, this paper chooses different characteristic parameters abstraction mathematic methods for each classifier pertinently.At last, the basic principles of support vector machine and particle swarm optimism algorithm are introduced. Support vector machine, which is suitable for classify data in finite sample set condition, is adopted as classify algorithm cause there is no too much field overvoltage data. The effect of multi-level recognition structure to classify rate is discussed in the paper and it is found that in order to maintain high classify rate, suitable initial value for the parameters of support vector machine is necessary. The particle swarm optimism algorithm is introduced to improve the support vector machine. The particle swarm optimism takes the classify rate as fitness function and the improved support vector machine is used to classify the overvoltage type. The testing calculation demonstrates that since the particle swarm optimism algorithm is quite insensitive to itself initial parameters value and has good performance in convergence, it is quite suitable to improve the support vector machine. The compare of three kinds of classify algorithm shows that the particle swarm optimism algorithm optimized support vector machine can greatly remedy the disadvantage of multi-level structure and has better performance than common support vector machine and BP artificial networks under same condition.更多还原