基于极限学习机的水泥熟料游离氧化钙含量预报方法研究

发布时间：2018-06-03 07:02

本文选题：极限学习机 + 游离氧化钙　；参考：《湖南大学》2015年硕士论文

【摘要】：游离氧化钙(f-CaO)是指水泥熟料中以游离形式存在而没有和其他物质结合的氧化钙,其含量是判断水泥熟料质量的重要指标之一。目前大部分水泥厂采用离线方法在实验室对f-CaO含量进行人工测量,且每小时只检测一次,影响中控室操作工人对熟料煅烧工况的及时判断和调节,因此,急需研究一种f-CaO含量的自动预报方法。但是,由于水泥生产过程机理复杂,且具有大滞后、非线性、强耦合和时变等特点,因此很难建立准确的机理模型对水泥熟料中的f-CaO含量进行预报。极限学习机是一种基于数据驱动的建模方法,是一种改进的单隐层前馈神经网络,具有参数设置简单、学习速度快、泛化性能好的优点,在复杂过程建模领域得到了越来越多的运用。本文依托国家自然科学基金项目及湖南省自然科学基金项目,以江西某水泥厂熟料煅烧过程为具体对象,研究基于极限学习机的水泥熟料f-CaO含量预报方法,具有较高的理论价值和工程意义。论文完成的主要工作和结论如下:(1)根据水泥熟料煅烧过程中物料的运动、气流的运动和燃料的运动对水泥熟料煅烧工艺进行了详细分析,并介绍了熟料煅烧过程的重要过程设备以及f-CaO含量的人工检测方法。(2)基于水泥熟料煅烧工艺机理,结合现场的实际经验,选取了主机电流、分解炉温度和篦冷机二室风压强三个易测量的过程参数作为极限学习机的输入变量,并设计了f-CaO含量预报极限学习机模型;针对水泥熟料煅烧过程时滞性特点,重点研究了极限学习机输入输出变量之间时间匹配的问题。(3)在江西某水泥厂采集了一个多月的现场数据,包括生料进入预热器到熟料出篦冷机之间所有过程参数、化验室每个小时记录的f-CaO含量等重要工况数据。对采集的原始数据进行了预处理,并且剔除了由于设备故障等原因产生的异常数据,获得720组可靠的训练和测试样本集。(4)用MATLAB语言编程实现了极限学习机对水泥熟料f-CaO含量的预报算法。分别采用训练样本及测试样本对建立的模型进行验证,预报的均方误差分别为0.247和0.196,低于已有文献报导中其他算法(BP神经网络和支持向量机)的均方误差(0.5左右),效果良好。(5)分析研究了预测结果的影响因素,包括隐层节点个数的选取、不同的数据滤波处理算法以及输入输出变量之间时间间隔的选取。分析结果表明,隐层节点个数设置为310,对原始数据进行5分钟均值滤波,时间匹配选取方案C得到的预测结果更好。(6)将ELM模型预测结果与支持向量机预测结果进行了对比,结果表明,ELM模型预测结果在最大绝对误差、平均绝对误差和均方误差指标上均优于支持向量机。
[Abstract]:Free calcium oxide (f-CaO) is a calcium oxide in cement clinker, which is free in the form of free form and is not combined with other substances. Its content is one of the important indexes to judge the quality of cement clinker. At present, most cement plants use off-line method in the laboratory to measure the content of f-CaO in the laboratory, and only once per hour, affecting the control room operation. For workers to judge and adjust the calcining conditions of clinker in time, it is urgent to study an automatic prediction method of f-CaO content. However, it is difficult to establish an accurate mechanism model to predict the content of f-CaO in cement clinker because of the complex mechanism of the cement production process and the characteristics of large lag, nonlinear, strong coupling and time-varying characteristics. The limit learning machine is a data driven modeling method. It is an improved single hidden layer feedforward neural network. It has the advantages of simple parameter setting, fast learning speed and good generalization performance. It has been used more and more in the field of complex process modeling. This paper relies on the National Natural Science Foundation Project and the natural science foundation of Hunan province. The gold project, taking the calcining process of clinker in a cement plant in Jiangxi as the specific object, studies the prediction method of f-CaO content of cement clinker based on extreme learning machine, which has high theoretical value and engineering significance. The main work and conclusion are as follows: (1) according to the movement of material, the movement of air flow and the transportation of fuel in the process of calcining of cement clinker. The process of calcining of cement clinker is analyzed in detail, and the important process equipment of clinker calcination process and the artificial detection method of f-CaO content are introduced. (2) based on the process mechanism of cement clinker calcining and the actual experience of the field, three easy measurement processes are selected, the main current of the host, the temperature of the calciner and the strong pressure of the two chamber air pressure of the grate cooler. The parameter is the input variable of the limit learning machine, and the f-CaO content prediction limit learning machine model is designed. According to the time delay characteristic of the calcining process of cement clinker, the problem of time matching between the input and output variables of the limit learning machine is studied. (3) a month's field data is collected in a cement plant in Jiangxi, including the entry of raw material. The data of all process parameters between the heater to the clinker grate cooler and the f-CaO content recorded in the laboratory per hour. Preprocessing the collected raw data and eliminating the abnormal data caused by equipment failure. 720 groups of reliable training and test samples are obtained. (4) programming by MATLAB language The prediction algorithm for the f-CaO content of cement clinker by the limit learning machine is verified by training samples and test samples respectively. The mean square error of the prediction is 0.247 and 0.196 respectively, which is lower than the mean square error (about 0.5) of other algorithms in the literature (BP neural network and support vector machine), and the effect is good. (5) analysis and research. The factors affecting the prediction results, including the selection of the number of hidden layer nodes, the different data filtering algorithms and the selection of the time interval between the input and output variables, show that the number of nodes of the hidden layer is set to 310, the mean value of the original data is 5 minutes, and the time matching selection scheme C is better. (6 Compared with the prediction results of the ELM model and the support vector machine, the results show that the prediction results of the ELM model are better than the support vector machines in the maximum absolute error, the mean absolute error and the mean square error index.
【学位授予单位】：湖南大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TQ172.6;TP181

【参考文献】