【作者】 曾九孙；

【导师】 刘祥官；

【作者基本信息】 浙江大学 ， 运筹学与控制论， 2009， 博士

【摘要】 钢铁工业是国民经济的重要支柱产业,也是大量消耗能源的产业。而作为钢铁生产体系中能耗最大的环节,高炉炼铁的每一个技术进步都将带来巨大的经济和社会效益。本文以内蒙古包钢6#高炉在线采集的冶炼过程数据为基础,针对高炉的实际情况及炼铁生产中工长关注的实际问题,对高炉冶炼过程的辨识、预测及控制进行研究,从而为高炉冶炼过程的闭环控制提供了一种新的思路,具有一定的理论意义和应用价值。高炉冶炼过程自动控制的核心难题是炉温的预测和控制。其复杂性来源于几种复杂动力学的交叉。一个良好的预测模型对实际生产有重要指导价值,而如果能够得到一个有效的控制模型,其意义则更为显著。因此,高炉冶炼过程的预测控制问题成为本篇论文研究的核心问题。论文的第二章简要介绍了高炉生产流程以及高炉专家系统的最新设计思路,通过炼铁工艺机理模型、炉况诊断推理模型、系统优化的数理模型以及炉温预测控制模型的智能化集成,为实现高炉冶炼过程的闭环控制奠定基础。论文的第三章针对高炉实际数据中经常出现的缺失值和异常值分别进行处理,对于数据中的缺失值,通过将离散属性的数据进行分类从而利用C4.5决策树算法进行填充,取得了较好的填充效果。对于数据中的异常值,基于马氏距离等4种多变量异常值检测方法来判定异常值并将其从原始数据中删除。对异常值和缺失值的合理处置大大降低了数据的波动性。在此基础上,论文第四章介绍了一种较新的系统辨识方法—子空间方法。这种方法基于输入输出数据,对模型结构的先验知识需求较少,其算法实现依赖于一些简单可靠的线性代数工具因而在数值计算中具有一定的鲁棒性并且能够较好地适用于高炉冶炼过程等多变量系统。利用子空间方法对第三章的数据处理效果进行验证,结果表明数据处理能够对高炉冶炼的辨识和预测产生较好的效果。为进一步提高辨识效果,论文第五章基于子空间方法分别对高炉冶炼过程采用Wiener模型和Hammerstein-Wiener模型结构进行辨识。对于Wiener模型辨识,通过对非线性模块的逆函数进行逼近从而显式地得到系统的非线性特性,将多输入单输出非线性问题转化成为一个多输入多输出的模型进行辨识,利用辨识得到的模型对包钢数据进行预测,预测命中率达到81%。对于Hammerstein-Wiener模型,在辨识过程中利用径向基函数良好的逼近能力及其对多变量系统的方便处理分别对Hammerstein非线性和Wiener非线性部分进行逼近,并利用BFGS最优化方法对参数进行优化。利用四种径向基函数:Laplacian径向基函数、Logistic径向基函数、Gaussian径向基函数和薄板样条径向基函数进行建模,结果表明利用薄板样条径向基函数对高炉冶炼过程进行辨识和预测得到的效果更佳,其命中率达到85%。在探讨了高炉冶炼过程的辨识和预测问题之后,论文的第六章对高炉冶炼过程的控制做进一步的研究。在现行的生产实践中,高炉工长对冶炼过程的控制主要是以专业知识和实践经验为基础的人工控制。由于高炉工长控制水平各不相同,造成了高炉冶炼过程的不稳定。本章针对这个问题,基于子空间方法构造了一种适合于高炉冶炼过程的自适应预测控制方法,提出了一种更为简便的约束处理方法,同时在方法中对各个输入变量调整的优先级进行考虑。结合高炉的实际情况,从实际数据的计算出发得到了保持高炉高产低耗的最佳炉温水平。仿真结果表明,设计的预测控制模型能够有效降低高炉冶炼过程的波动性,而各个输入变量的调整均没有超出约束,这表明预测控制算法的有效性和可行性。此外,这种思路克服了之前预测模型的“预测—控制”模式所带来的“模型悖论问题”,即工长利用了炉温预测信息对高炉行程进行调控后,炉温的发展改变了轨迹,从而与初始预测值产生了偏差的问题。而自适应预测控制方法将静态预测-控制-动态预测结合在一起,整体考虑炉温预测控制问题,因而能达到更好的应用效果。论文第七章对全文的研究内容以及创新点做了归纳,并对课题的后续研究做了展望。更多还原

【Abstract】 As the pillar industry of national economy, steel industry is among the most energy intensive, while for all the sub-processes in steel industry, blast furnace (BF) ironmaking consumes the largest part of energy. Thus every technological progress in BF ironmaking will bring enormous economic and social benefits. By investigation into the actual situation of BF ironmaking and BF operators? concerns, the current study discussed the identification and predictive control of blast furnace ironmaking with data collected from No. 6 BF at Baotou Steel. The study introduces a new idea for closed-loop control of BF ironmaking and is meaningful both in theory and practice.In the efforts to achieve closed loop control of BF ironmaking, the crucial problem is prediction and control of silicon content in hot metal, whose complexity is the result of interaction between chemical reaction dynamics and kinetics. An accurate predictive model for silicon content will greatly enhance BF ironmaking, while an effective controller will bring significant benefits. These two problems become the main research efforts in this dissertation. Chapter 2 gave a brief introduction to the blast furnace ironmaking process and some latest development on design of BF expert system. By integration of 4 kinds of models, e.g. mechanism models, inferential models, optimization models and predictive control models a closed loop controller for BF ironmaking is constructed. Chapter 3 dealt with the problem of missing values and outliers in the raw data. For missing values, a C4.5 decision tree based algorithm was adopted by discretizing the missing attribute and good results are achieved. While for the outliers, 4 multivariate outlier detection methods were used and the detected outliers are deleted from the original dataset. The fluctuation of data decreased significantly after processing of missing values and outliers. Chapter 4 introduced a novel method in the theory of system identification- subspace methods and tested the method on data before and after processing. Subspace methods identify system model from input and output data, its algorithm uses some simple and reliable linear algebra methods and is robust and efficient. Its good capacity to deal with multivariate system is very suitable for identification of blast furnace ironmaking process. Simulation experiments were carried out to test the method based on data before and after processing of missing values and outliers. It is shown that processing of missing values and outliers enhances the identification of blast furnace ironmaking process. To get more accurate models, Chapter 5 identifies the BF ironmaking process based on Wiener model and Hammerstein-Wiener model using subspace methods. For identification of Wiener model, a polynomial function was used to approximate inversion of the nonlinear part of Wiener model. By this technique the multivariate input single output (MISO) nonlinear problem was converted to a multivariate input multivariate output (MIMO) linear system and the identification became simpler. The identified model was then test on data from Baotou Steel and a hit-rate of 81% was achieved. As for the Hammerstein-Wiener model, radial basis functions (RBF) were used to approximate both the Hammerstein and Wiener nonlinearity, a BFGS quasi-Newton method was used to optimize the parameters. Four kinds of RBFs, e.g. Laplacian, Logistic, Gaussian and Thin-plate Spline RBF were tested. Simulation results shown identification using thin-plate spline RBF is the most accurate with the hit-rate of 85%.After discussion on the identification and prediction problem, Chapter 6 made further research on predictive control of BF ironmaking. Currently, control of BF ironmaking heavily relies on expert experience of operators. However, since different operators have different habits, control of BF ironmaking becomes inconsistent and unstable. To solve this problem, an adaptive predictive control method was constructed based on subspace method. The predictive control method adopts a simpler constraint handling method and priorities of input variables were also taken into consideration. The best level of silicon content was computed from practical data. Simulation results shown that the designed predictive control method can effectively handle the fluctuation of BF ironmaking problem, while the constraints of input variables were not violated. Thus the effectiveness and feasibility of the proposed method was proved. Besides, the adaptive method can overcome the problem of previous predictive models. In practice, the BF operators make control movements based on future information of silicon content obtained through predictive models, which results in deviation between actual and predicted silicon content. The adaptive predictive control method is a combination of static prediction, control and dynamic prediction so that it is capable of handling the overall condition of the ironmaking process. Finally, Chapter 7 gave the conclusion and theoretical innovations in the paper, issues for further research were also investigated.更多还原