【作者】 文孝强；

【导师】 徐志明；

【作者基本信息】 华北电力大学 ， 热能工程， 2013， 博士

【摘要】 换热器污垢的形成是在影响因素众多,动量、能量、质量传递甚至有生物活动同时存在的多相、多组分流动过程中进行的,其理论基础除传热传质学外,还涉及到化学动力学、流体力学、胶体化学、热力学与统计物理、微生物学、非线性科学以及界面科学等相关知识,是一个典型的多学科交叉的高度复杂问题。作为20世纪80年代以来污垢研究的基础和三个主要方向之一的污垢预测,旨在通过对污垢形成过程的理论分析和实验研究,建立一个通用、准确而又便于应用的预测模型,为换热设备的设计和运行提供指导。传统的预测研究方法虽取得了一些可喜的进展,但由于污垢的形成过程影响因素众多,加之多学科交叉带来的重重困难,其进展仍是缓慢,离预期目标依然十分遥远。本文基于所搭建的实验系统以及期间所积累的大量污垢数据,尝试利用支持向量机、偏最小二乘算法、模糊数学等智能预测理论与方法,对换热器污垢特性进行建模与预测研究,具体的研究内容如下：作为建模工具,将支持向量机算法引入换热器污垢特性的建模中,并研究了在径向基函数作为核函数的情况下,参数变化对支持向量机模型预测能力的影响,针对传统惩罚系数和核系数寻优过程中所呈现出来的问题,首次提出了“显微镜”理论,实例检验证明,该方法提高了寻优的速度和准确率,加上模拟退火算法的有效配合,为后续的建模和优化工作奠定了基础。以松花江水为冷却介质,实验研究了板式换热器污垢特性,通过所搭建的实验系统有选择性的测量了对污垢形成影响较大的几个水质参数：pH值、电导率、溶解氧、浊度、硬度、碱度、氯离子、化学需氧量、铁离子浓度、细菌总数,以及运行工况、污垢热阻等参数,获得一组典型水质的污垢数据。以该水质参数为自变量,以污垢热阻为因变量,分别基于偏最小二乘算法、支持向量回归机,对板式换热器污垢特性进行了预测建模,并分析了各水质参数对模型预测精度的影响。研究结果表明：两种方法预测精度都能控制在10%以内,满足工程要求,由此证明,从循环冷却水水质角度来预测换热器污垢特性是合理可行的,从而也为今后在已知水质条件下设计冷却水系统提前预知污垢特性提供了一种有效的新方法；预测结果的对比表明,SVR (support vector machine)方法优于PLS (partial least squares algorithm)方法,建议采用SVR方法对板式换热器污垢特性建模和预测研究；通过逐一删除水质参数项的方式,讨论了各水质参数对预测模型的影响,结果表明,部分水质参数的删除既在一定程度上提高了模型预测精度,也降低了测量成本。以人工配置的硬水为冷却介质,来模拟析晶垢,实验研究了光管换热器析晶污垢特性,测得温度、污垢热阻等参数,获得了一组同一实验管的两个运行周期的污垢数据。以出、入口温度,壁温等为自变量,以污垢热阻为因变量,分别搭建了PLS预测方程和SVR预测模型。预测结果表明：两种方法预测精度皆满足工程要求,皆可用于光管析晶垢的预测研究；相对来说,环境温度等参数的获得比较容易,而且节省人力和物力,由温度等参数推测污垢热阻值可实现换热设备污垢热阻的在线监测。同时,采用完全相同的两根不锈钢弧线管,通过向工质中加入MgO微粒的方式来模拟颗粒污垢,分流速恒定、可变两种情况,实验研究了弧线管换热器颗粒污垢特性,并分别搭建了SVR预测模型。通过对比,结果表明：当流速等影响换热器污垢热阻的主要因素由常量变成变量,随时间而变化时,应该对预测模型做出修改,以提高模型的预测精度。将类心向量理论引入燃煤结渣特性预测研究,该方法在准确预测混煤的结渣倾向性的同时,还可以有效解决了混煤掺烧比例的问题；基于模糊集理论,提出了模糊关联系数,构造了模糊相对权重,在此基础上提出了换热设备结渣特性模式识别算法,预测结果表明,所提出的方法是可行的、有效的,为换热设备结渣特性模式识别理论提供了一种新的研究方法,是对传统模式识别理论的发展与完善；将Vague集理论引入燃煤锅炉结渣评判中,同时,采用一种新的计算相似度的方法—距离意义下相似度量—计算Vague集的相似度,评判结果表明此方法是可行的,不但如此,此方法所得数据结果能够使现场运行人员比较容易地得出当前运行锅炉的结渣状况,从而消除了干扰因素的影响；基于RBF (radial basis function)网络建立了锅炉结渣预测模型,预测结果表明,所建RBF模型的评判准确率高于常规的BP网络,而且避免了局部极小点问题；利用非线性支持向量回归机方法对燃煤结渣特性进行了有效预测,该方法不但预测精度高,而且该方法最为突出的优点是能够利用小样本进行训练学习,解决了多维向量空间下的模式识别问题。为了对各种预测方法进行比较,在采用相同的已知样本训练后,对同组测试样本进行预测,结果表明：评判准确率最高的是RBF、SVR、FRW (fuzzy relative weight)及Vague集模型,其次是PLS方法,准确率最低的为类心向量法。基于煤灰的化学分析成分及Elman网络搭建了煤灰软化温度预测模型,所搭建的Elman灰熔点预测模型能够较好的完成对某热电厂煤灰熔点的预测,而且较BP (Back-Propagation network)网络方法,精确度更高。通过对模型进行分析,找到对煤灰灰熔点起主要作用的八种灰成分更多还原

【Abstract】 The formation of heat exchanger fouling is a multiphase, multi-component flow process affected by many factors, momentum, energy, and mass transfer, even the existence of biological activities. Its theoretical basis involves not only heat and mass transfer, but chemical kinetics, fluid mechanics, colloid chemistry, thermodynamics and statistical physics, microbiology, nonlinear science and interface science, etc. It is a typical multi-disciplinary highly complex problem. Fouling prediction, as the foundation of fouling study since the1980s and the three main directions, is aiming at establishing a common, accurate, and easily applied prediction models though theoretical analysis and experimental study of the formation of fouling, to provide guidance for the design and operation of the heat exchanger. The traditional prediction methods have made some encouraging progress, however, because of the impact of many factors on fouling formation process and difficulties brought by multidisciplinary, research progress is still slow, even far away from the target. Based on the experimental system built and fouling data accumulated during the PhD, this paper attempts to model construction and prediction of the heat exchanger fouling characteristics, by using intelligent forecasting theory and methods, such as support vector machines, partial least squares algorithm, fuzzy mathematics, etc. The specific contents are as follows:As a tool for model building, support vector machine (SVR) was introduced into modeling of the heat exchanger fouling characteristics. The affection of the parameters on the SVM model was studied under the condition that RBF function was selected as kernel function. To solve the problem encountered in the optimization process of penalty coefficient and nuclear factor, the "microscope" theory was put forward for the first time. It was proved by practice tests that the method could improve the speed and accuracy of the optimization and became the basis of subsequent works coupled with effective co-ordination of the simulated annealing algorithm.Fouling characteristic of plate heat exchanger was studied through the experimental system, with the Songhua River water as working fluid. Several water quality parameters: pH value, conductivity, dissolved oxygen, turbidity, hardness, alkalinity, chloride ion, iron ion concentration, chemical oxygen demand, total bacterial count, which had great influence on the formation of fouling, as well as running condition, fouling resistance and other parameters were measured through the experimental system built. A group of fouling data of the typical water quality was obtained. Two prediction models of fouling characteristics of the plate heat exchanger were built based on partial least squares algorithm (PLS) and support vector regression machine (SVR) with water quality parameters as independent variables and fouling resistance as dependent variable, and the impact of water quality parameter on predicting accuracy was analyzed. Research results showed that:the prediction accuracy of two methods could be controlled within10%and meet the requirements of the project, which proved that it was feasible to predict heat exchanger fouling characteristics by water quality of circulating cooling water, and put forward an effective new method to forecast fouling characteristic under the condition of known water quality parameters in the process of designing the cooling water system. Through the comparison of the prediction results, it was proved that the SVR method was better than the method of PLS, and it was recommended modeling and predicting of the fouling characteristic of plate heat exchanger based on SVR. The impact of the water quality parameters on prediction model was discussed by the means of removing the water quality parameters one by one. The results showed that deletion of part of water quality parameters could both improve the prediction accuracy of the model to some extent, but also reduce the cost of measurement.Fouling characteristic of plain tube was studied through the experimental system with the man-made hardness water as cooling medium to simulate the crystallization fouling. The parameters of temperature, fouling resistance et al. were measured and a group of fouling data of the same test tube was obtained, which was across the two operation cycle. Two predicting models of fouling characteristics of the tube were built based on PLS and SVR with outlet temperature, inlet temperature, wall temperature et al. as independent variables and fouling resistance as dependent variable. Research results showed that:the prediction accuracy of two methods could meet the requirements of the project and be used to predict the crystallization fouling characteristic of plain tube. Relatively speaking, environment temperature et al. gets easier, and saves manpower and material resources, so, it could realize on-line monitoring of fouling resistance of heat exchangers to predict the fouling resistance by temperature et al. At the same time, fouling characteristic of arc tube was studied through the experimental system, in which there were the same two stainless steel test tube, with the MgO particle being added into working liquid to simulate particle solution. The SVR model was selected to predict the fouling characteristics of arc tube. By comparison, the results showed that:the predicting model should be modified to improve the precision of the model, when the flow velocity or other main parameters were no longer constant, changing along with time.The theory of class centroid vector was introduced into predicting the slagging characteristics of coal-fired. The method could not only accurately predict the slagging tendency of mixed coal, but also effectively solve the problem of mixing ratio of the coals. It was based on fuzzy theory that the fuzzy correlation coefficient was put forward, fuzzy relative weight was constructed, and the pattern recognition algorithm of slagging characteristics of heat exchanger was proposed. The predicting results showed that the method proposed was feasible and effective. It provided a new research method of predicting the slagging characteristics of heat exchanger, which was a development and perfection of traditional pattern recognition theory. Vague sets theory was introduced into the prediction of the slagging characteristics of coal-fired boilers, and at the same time, a new formula of similarity degree, which was based on the sense of distance, was proposed to calculate the similarity between vague sets. The results showed that not only this method was feasible, but also the operators could easily predict the slagging state of the coal-fired boilers based on this method, and thus eliminate the influence of interference factors. The prediction model of slagging state of coal-fired boilers was built based on RBF network. The prediction results showed that, the RBF model was higher in prediction accuracy than that of the conventional BP network, and avoided the local minima problem. The slagging characteristics of coal-fired boilers were predicted effectively based on nonlinear support vector machine for regression. The method was not only high in prediction accuracy, but also had the most prominent advantage of small samples learning, which solved the problem of pattern recognition in multi-dimensional vector space. In order to compare the predicting methods above, the same known samples were used for training and testing, the comparing results showed that:RBF method, SVR method, FRW method and Vague sets method had the highest predicting accuracy rate, followed by PLS method, and the lowest was the method of class centroid vector.The prediction model of softening temperature of coal ash was built based on the chemical analysis component and Elman network. The model could not only be able to accurately predict the softening temperature of coal ash from the thermal power plan, but also was higher in prediction accuracy than that of traditional BP network. Through analysis of the model, the eight kinds of ash composition were found out, which played the major role in softening temperature of coal ash.更多还原