【作者】 王晓霞；

【导师】 王兵树； 马良玉；

【作者基本信息】 华北电力大学 ， 控制理论与控制工程， 2014， 博士

【摘要】 为满足电网负荷需求,火电机组常需按照电网调度指令改变机组负荷而处于变负荷动态过渡工况运行。以往针对机组稳定工况的故障诊断研究无法适应这一新情况,严重影响故障诊断系统在实际机组中的应用效果。深入开展变负荷动态过渡过程故障诊断研究,对提高机组可用率、保障机组安全、经济运行具有重要的现实意义。为此,本文基于机组监控系统提供的丰富的热力系统运行参数,深入研究变负荷动态过渡过程热力系统故障的多智能融合诊断方法,并以超临界机组高压给水加热器系统故障诊断为例对算法进行深入仿真试验研究。文中首先对混合智能诊断算法中涉及的粒子群优化算法、Elman神经网络和近邻分类法的基本概念和原理进行了论述,并结合具体实例阐明其在不同领域应用的有效性。针对变工况动态过渡过程中故障特征参数正常参考值难于获取的困惑,考虑Elman神经网络辨识非线性动态系统的优势和粒子群算法优良的优化性能,建立了一种基于进化Elman神经网络的动态过程特征参数参考值预测模型。该模型采用离散粒子群算法优化Elman网络的结构,改进粒子群算法(IPSO)调整Elman网络的参数,结合两种算法的优势同时进化Elman网络,以自动产生一个高泛化性能和低结构复杂度的网络。IPSO算法中采用新的学习策略并引入变异操作以增强种群多样性,从而提高算法的寻优精度和收敛速度。2个时间序列预测的实例验证了该方法的可行性和有效性。将该方法应用于高加系统变工况动态过渡过程特征参数的预测,取得了满意的结果。以变工况动态过程故障的准确诊断为目的,提出了一种基于粒子群和最近邻的热力系统故障诊断新方法。针对故障类别多样、机组运行工况多变导致的故障样本集过于庞大的问题,采用IPSO算法优化典型故障原型,有效减少样本数量,并确保产生的精简原型集能够全面地表示不同故障类的典型特征分布。结合进化Elman神经网络实时预测的特征参数参考值进行故障征兆计算,并依据最近邻分类规则对故障进行诊断。8个UCI标准数据集的分类结果验证了IPSO算法优化原型的有效性。借助某600MW超临界机组全工况仿真系统,在100%至80%额定负荷范围内,对高加给水系统故障进行详细的仿真诊断试验,结果表明该方法对不同稳态下高加给水系统程度迥异故障以及变负荷动态过渡过程中发生的高加故障均能给出满意的诊断结果。为进一步提高变负荷动态过渡过程热力系统故障诊断的准确率,利用神经网络优良的非线性逼近能力,提出了一种融合多个基分类器的集成分类器诊断模型。该模型使用原始的训练集建立基分类器以保证故障信息的完整性,提高基分类器的诊断正确率；采用不同的学习参数训练单个基分类器,以增加基分类器的多样性；运用神经网络融合多个基分类器的诊断结果进行二次诊断,有效提高集成分类器整体的诊断正确率。借助600MW超临界机组全工况仿真系统,在100%至70%额定负荷范围之间进行基于集成分类器的高加故障诊断仿真试验,表明集成分类器诊断模型对70%额定负荷稳态工况以及变负荷动态过渡过程高加给水系统程度迥异故障均能给出满意的诊断结果。诊断结果同时验证了神经网络融合方法相比于多数投票和加权平均融合方法的有效性。本文研究成果对提高智能故障诊断方法对热力系统变工况动态过渡过程的适应性,促进热力系统故障诊断从理论方法研究走向工程应用,提高机组自身运行的安全性和经济性具有重要作用。更多还原

【Abstract】 The power unit changes the unit load according to the dispatching commands to meet the load demand of the power grid, so it often runs under variable load dynamic transient conditions. Many successful works have been done in this field, but most of them focus on fault diagnosis under steady state load conditions. Therefore, it is of great practical significance to carry out fault diagnosis study for the thermal system under load-varying dynamic transient conditions with potential advantages of improving the availability, safety and economy of a power unit. In order to take advantage of the diversity and complementarity of individual intelligent diagnosis methods, we fuse these methods to improve accuracy of a diagnosis system. The fault diagnosis method for the thermal system under variable load dynamic transient conditions is investigated. The high-pressure feedwater heater system of a600-MW coal-fired power unit is taken as a target system and typical faults are simulated and diagnosed.The basic conceptions and principles of particle swarm optimization algorithm, Elman neural networks and nearest neighbor classification are introduced in this dissertation. An illustration for each algorithm is provided to show the validity for variable application fields.Aiming at the puzzled problem that the reference values of feature parameters are not easy to be accurately predicted under load-varying dynamic conditions, an evolutionary Elman neural network(ENN) prediction model is proposed. A novel hybrid optimization algorithm is presented for simultaneous optimization of ENN’s structure and its parameters. When training an ENN, a discrete PSO algorithm is adopted to determine the structure of the ENN and an improved PSO (IPSO) algorithm is employed to adjust the parameters. The hybrid algorithm combines the advantages of DPSO and IPSO to automatically generate a network with low complexity and high generalization performance. In the IPSO algorithm, each particle adjusts its trajectory towards its personal best position, the global best position achieved so far by the whole swarm and the best particle in the current swarm. A mutation operation is integrated into the IPSO algorithm. These improvements can enhance convergence and increase diversity of the swarm. Two time series prediction examples are illustrated to thoroughly verify the feasibility and efficiency of the proposed method for structure and parameter learning of ENN. Further application of the proposed method to feature parameters prediction of the feedwater heater system in a600-MW coal-fired power generating unit demonstrates the validity of the method.With the purpose of correct and rapid fault diagnosis under load-varying dynamic conditions, a novel fault diagnosis approach is proposed based on particle swarm optimization algorithm and nearest neighbor classifier. Due to multiple fault types and varying operation conditions, the fault sample set becomes too large. To solve this puzzle, the IPSO algorithm is employed to generate a sufficient small set of the typical fault prototypes which can represent efficiently the distributions of the fault classes and discriminate well when used to classify. Moreover, the evolutionary Elman neural network is used to predict the reference values of fault feature parameters. Real-time fault symptoms are calculated with the predicted values and the real values of the feature parameters. Fault diagnosis is implemented by calculating the similarity between the real-time fault symptoms and the generated fault prototypes with the nearest neighbor classifier. The effectiveness of the proposed prototype optimization method based on IPSO algorithm is evaluated on eight real world classification problems. Finally, the proposed diagnosis method is applied to diagnose faults of a high-pressure feedwater heater system in a600-MW coal-fired power unit. The obtained results show the proposed method can recognize varying degrees of faults under steady load conditions and achieve good diagnosis results under load-varying dynamic transient conditions.In order to further improve the accuracy of fault diagnosis of the thermal system under load-varying dynamic conditions, a new fault diagnosis model based on ensemble classifier is proposed which employs a neural network to fuse the results of multiple base classifiers. The original data set is used to train the base classifiers to ensure the integrity of the faulty information, different learning parameters are employed to improve the diversity of base classifiers and a neural network is used to fuse the results of multiple base classifiers to improve the accuracy of fault diagnosis of the whole ensemble classifier. The proposed ensemble classifier can obtain satisfactory diagnosis results for the faults of varying degrees in a power plant thermal system under both steady load conditions and load-varying dynamic conditions when the scope of loading conditions is expanded to70%rated-load.It is of great practical significance for improving the diagnosis accuracy of intelligent diagnosis methods, promoting thermal system fault diagnosis from theoretical research to engineering application and improving the safety and economy of the power unit.更多还原