【作者】 罗建；

【导师】 徐国禹；

【作者基本信息】 重庆大学 ， 电气工程， 2004， 博士

【摘要】 母线是电力系统厂站最重要的设备之一，母线保护是保障母线安全和可靠运行的保护设备。寻找性能完善、功能强大、可靠性高及智能化程度高的母线保护是母线保护研究的方向。 为实现性能更加优良的母线保护，采用一些新理论和新技术来研究母线保护是必要的。人工神经网络(ANN)是一种具有高度智能化的数学工具，可应用于新型母线保护的研究中。依据ANN具有分类能力的特点，利用电力系统故障数据及特征学习训练ANN，学习训练后的ANN就可用于对电力系统故障的识别和判断，这一ANN分类原理的保护方法在不同保护对象上得到了相当多的研究。但是由于不易获取完整的故障样本，这种ANN保护方法还未在电力系统中得到广泛实际应用。 针对基于ANN分类能力的ANN保护方法的不足，本文提出了基于ANN函数逼近能力的ANN母线保护方法。函数逼近能力是ANN具有的重要性能之一，依据ANN具有的函数逼近能力，可用ANN模型去替代一个确定的物理对象。母线保护的物理对象是一个输入输出关系确定的函数对象，可用一个ANN模型去替代，或用一个ANN数学模型去逼近母线保护物理对象的输入输出之间的函数关系。通过这个在无故障运行时学习训练出来的母线保护对象的ANN数学模型，就能判断区分母线保护对象的区内和区外故障。 在研究基于ANN函数逼近能力的母线保护原理的过程中，首先就母线保护物理对象的输入输出之间的函数关系进行了研究，将物理对象的输入定为母线上各回路电流互感器(TA)传变后的同一时刻二次检测电流，输出定为母线上各回路的一次侧同一时刻流过的电流之和。然后对母线保护的ANN模型进行了研究和讨论，就激活函数为线性函数、分段线性函数和非线性函数的三种情况，分别构建了母线保护的ANN模型，给出了ANN模型的学习训练算法，讨论了训练算法的收敛性，并通过DSP程序实现得出了对母线ANN模型的训练实例。最后通过对三种ANN模型的研究分析结果的比较，选定了分段线性激活函数ANN模型和线性激活函数ANN模型作为母线保护的ANN模型，考虑到实现条件的限制，本文就基于线性激活函数ANN模型的母线保护方法进行了研究和实现。 为研究和实现基于ANN模型的母线保护方法，本文依据母线保护的ANN模型输出来判断区分母线的区内区外故障，采用了ANN模型输出变大与故障发生时间是否同步的原理来消除TA饱和的影响，采用了比例制动方法来保证母线保护在区外故障时的不误动，采用了故障电流方向来判断处理在母线出现区外故障转区内故障时和TA断线时的情况。本文还利用自适应模型的预测和电流瞬时幅值的增量实重庆大学博士学位论文现了母线保护的故障预启动和故障启动，灵敏故障预启动是实现抗TA饱和母线保护方法的基础，可靠故障启动是母线保护出口的必要条件。 本文就信号采集、信号传输、信号处理等母线保护的实现环节进行了深入的研究，解决了实现环节中的数据同步采集、数据同步传输、信号滤波处理等技术关键问题。特别是就故障时的瞬变信号提出了正弦逼近处理方法，该方法的理论基础是瞬变信号的瞬时幅值和相位表示方法，这种表示方法与传统表示方法在结果上是一致的。通过实例分析结果得出，在故障发生后的半个周期内，特别在头一个1/4周期内，依据正弦逼近处理方法求出的瞬时幅值和瞬时相位能正确表征故障的特征，这一特点是正弦逼近处理方法所独特的。 基于ANN模型的母线保护在硬件上采用了分布式硬件结构，有间隔单元和中心单元两层单元，间隔单元实现了信号采集、信号处理、故障启动等功能，中心单元主要实现了ANN模型训练和故障判断的功能，两个单元之间依据通信进行数据传送。基于ANN模型的母线保护在软件上采用了实时多任务操作系统来编写各个功能软件程序，以提高保护软件的可靠性和实时性。 本文就母线在区内和区外故障状况下的不同类型故障对基于ANN模型的母线保护进行了物理模拟实验，实验结果表明基于ANN模型的母线保护能在母线区内故障时可靠出口，在母线区外故障时可靠不出口。关键词:母线保护，人工神经网络，分布式保护，信号处理，函数逼近更多还原

【Abstract】 As one of the most important devices of substation, bus and its protection play a big role in reliable and secure operation of power system. The main target of bus protection is the pursuit of perfect performance, almighty function, high reliability and intelligence.For improvement of bus protection, it is necessary to introduce new theory and technology. As a mathematic tool with high intelligence, ANN (Artificial Neural Network) is suitable for the research on new protection. For a long while, the application of ANN to relay protection is based on classification ability. ANN is trained by sample data, which can characterize fault. The various faults of power system can be distinguished and judged by the trained ANN. Unfortunately, lacking of sample data restricts the abroad application of ANN protection to power system.In order to overcome above shortcomings, innovative principle for ANN bus protection based on function approximation ability is presented in this paper. As one of the most important capability of ANN, function approximation ability can be used to design ANN model, which can characterize certain physics object. The protected object of bus protection is a function with certain input and output, which can be represented by ANN model, or approximated by ANN mathematical model. By use of ANN model trained under normal circumstance, inner and outer bus fault can be distinguished clearly.For the research on bus protection theory based on function approximation ability of ANN, it is crucial to establish the functional relation between the input and the output of bus protection object. The inputs are synchronous currents on second side of each CT, and the output is current sum on primary side of CT. ANN model with different activation functions, like linear, PWL and Sigmoid one, are built, training algorithm and its convergence of three models are discussed. As the models are realized with DSP, training examples and their comparison are given also. Linear function is chosen as the activation function finally.Inner and outer fault are distinguished by outputs of bus protection ANN model. Whether output of ANN model increases synchronously with fault time is the method of overcoming CT saturation for bus protection. Under outer fault circumstance, proportional halt mechanism based on ANN model is applied. Under transition from outer to inner fault with CT corruption and CT disconnection, fault power flow direction judgment is used to insure correct operation of bus protection. By use of prediction ofadaptive model and instantaneous increment of current amplitude, fault pre-triggering and triggering mechanism for bus protection are realized in this paper, and sensitivity and reliability are improved consequently. Sensitive pre-triggering is the foundation of CT anti-saturation, and reliable fault triggering is the prerequisite of protection exit.Some key problems on the synchronous data sampling, the synchronous data transmitting and signal processing are solved. Especially, SAIS(Sinusoidal Approximation of Instantaneous Signal)is presented for instantaneous signal processing under fault circumstance, whose foundation is instantaneous amplitude and phase. Examples show that in half period after fault, especially first quarter period, instantaneous amplitude and phase, which are gotten with sinusoidal approximation, can characterize fault clearly.Distributed hardware structure is adopted for bus protection based on ANN model, which is constituted with bay unit and central unit. Bay unit is specialized in data collection, data processing and fault triggering. Central unit is specialized in ANN model training and fault judgment. Real-time multi-task OS is chosen for programming, which is core of software structure.Physical simulation of bus protection based on ANN model is conducted under different locations and types of faults, whose results are promising.更多还原