【作者】 靳涛；

【导师】 杨勇平； 付忠广；

【作者基本信息】 华北电力大学（北京） ， 热能工程， 2011， 博士

【摘要】 火电机组热力系统在运行过程中具有非线性、时变、多变量和多回路耦合以及响应速度不同等各种动态过程交织在一起的特点,其建模方法一直是富于挑战性的研究课题之一。火电机组热力系统建模方法是解决机组优化运行、控制、性能评估、故障诊断、仿真等关键问题的基础。针对现有建模方法在实际应用中存在的精度不高,不能准确反映过程特点的问题,本文深入研究了一种新的建模理论——反向建模,以期提高火电机组热力系统的建模效率和建模精度。本文首次提出了火电机组反向建模方法的基本定义、一般建模模式、要求、意义和作用,形成了较完整的理论和方法。深入研究反向建模过程中的实时数据验证、特征变量提取和神经网络算法、偏最小二乘算法、最小二乘支持向量机算法和遗传算法等具体反向建模算法以及反向建模的可行性和模型的准确性验证。确保建模数据的正确性是反向建模的基础,本文研究实时运行数据的验证方法。根据火电厂不同测量数据布置的测点数目不同,把数据分两类验证。对于单、双测点数据,研究相邻数据变化率的方法和曲线拟合残差的方法。为了克服这两种方法对火电机组某些特殊运行工况下产生的突变数据的误判,提出用另外与它有密切关联关系的测点的变化量作为参考进行二次判断。对于多测点参数的数据,提出用格拉布斯准则进行实时验证。通过对实时运行数据的验证表明,所用方法可以有效剔除和校正实时数据中的显著误差。特征变量是反向建模中的重要参数,对建模精度有很大影响,因此研究特征变量的提取方法是提高反向建模精度的关键。本文通过对多种特征变量提取方法的研究,提出在反向建模过程中用灰关联分析法提取特征变量,运用少量输入变量达到改善模型精度和提高模型速度的目的,并在主蒸汽流量的反向建模中得到了验证。在确立了反向建模理论和方法后,进行反向建模的可行性研究和模型的准确性验证。对火电厂运行数据进行的反向建模分析表明,其运行数据之间的关系主要表现为两种形式：一种是自变量之间有良好的线性相关性,且自变量和因变量之间也有良好的线性相关性；另一种是自变量之间有良好的线性相关性,但自变量和因变量之间的相关性很弱。对第一种情况本文提出以偏最小二乘算法建模；对第二种情况,提出以神经网络算法和最小二乘支持向量机算法建模,并用遗传算法优化最小二乘支持向量机的参数。本文选取火电机组中具有代表性的参数为例,运用反向建模的方法分别建立了火电厂主蒸汽流量、汽轮机排汽焓、飞灰含碳量、亚临界和超(超)临界机组的高温受热面管壁温度、中间点温度等的数学模型,结果表明所建模型可以解决目前火电机组在线性能计算和关键参数的冗余分析中遇到的难题。所建模型的精度验证了反向建模方法的可行性,为火电机组建模提供了新途径。更多还原

【Abstract】 Due to the characteristics of nonlinear, time-varying, multi-variable, multi-loop coupling and different responding speed, the thermal system modeling method of power unit is always one of the challenging projects. The thermal system modeling method is a basic project in the field of the system operation optimization, control, performance evaluation, fault diagnosis, stimulation and etc. To solve the problems of the thermal system model in the application, a new modeling method called reversed modeling method was developed in the paper which aims at overcoming the lower efficiency and lower accuracy of the thermal system model.The basic definition, general modeling mode, requirements, significance and the applications of inversed modeling method were presented in the paper first time. A more integrated method and theory were preliminarily built. Real-time data validation, feature variables extraction, modeling algorithm including neural network, partial least squares algorithm, support vector machines and genetic algorithm, reversed modeling feasibility and model accuracy validation were studied in detail in the paper.Real-time data validation methods were studied in the paper because to ensure the creditability of modeling data is the basis for reversed modeling method. Different measurment parameters have different measuing point numbers, so data were divided into two categories to test. For single and double measuring point data, two real-time data validation methods are studied in the paper. One is based on the rate of adjacent data change and the other is based on curve fitting residual. To avoid the misjudgment of data validation on some special conditions with above two methods, the change rate of a measuring point data associated with the testing data could be taken as the reference data so as to improve the accuracy of model. For multi-point measurement data, Grubbs was recommended for real-time data validation. The results show that the methods can removed and regulate gross error effectively.Feature variables extraction is a very important issue for reversed modeling method, because feature variables have a great impact on the modeling accuracy. According to the exploration on the feature variables extraction methods, gray relational analysis was recommended for feature variables extraction in the paper. Useing a small amount of input variables can improve the model accuracy and speed. The example of main steam flow model demonstrates the effectiveness and availability of the variety selection method suggested in the paper for independent variables.After reversed modeling theory and method are established, reversed modeling accuracy and feasibility are studied in the paper. There are two type of relationship between power plant operation data. The first one is that the relationship between the independent variables and relationship between independent variables and the dependent variable are linear correlation; another one is that the relationships between the independent variables are linear correlation, but the relationship between the independent variables and the dependent variable are less correlation. For the first case, partial least squares algorithm was adopted for modeling. For the second case, neural network and support vector machines algorithm were developed for modeling in the paper. In the paper, the reversed modeling method was applied to build the mathematic model of main stream flow, steam turbine exhaust enthalpy, unburned carbon in fly ash, the high-temperature surface metal temperature and boiler intermediate point temperature of supercritical unit. The research results show that the models built in the paper are useful for solving the problem of the on-line performance computing and the redundancy analysis of key parameters. The modeling examples indicated that reversed modeling method was feasible with high accuracy. Reversed modeling method is a new way for thermal power unit modeling.更多还原