【作者】 韩晓霞；

【导师】 谢克明；

【作者基本信息】 太原理工大学 ， 电路与系统， 2010， 博士

【摘要】 建模与优化是工程技术研究的中心问题,预测是科学决策和规划的重要前提,预测的可靠性往往是衡量技术成功程度的重要指标,然而预测却是所有技术的薄弱环节,对于多相催化领域来讲亦是如此。随着国民经济的发展,尤其新能源需求、新材料的不断涌现,多相催化科学与技术面临新的挑战,既要从经济、安全、多功能等方面寻求新型、高效的催化剂,又要从化学进程全局出发全面考察催化剂的性能。基于历史数据的机器学习和数据深度挖掘已成为目前化工信息化领域中急需解决的重要问题。作为催化领域的难点和迫切需要解决的三个关键问题:催化剂的动力学关系模型、催化剂的活性关系模型、催化剂的优化设计,这无论是对于催化剂对象的特性研究,还是实际化工生产过程控制、优化、模拟等都具有重要的现实意义。本文结合混沌与支持向量机,围绕多相催化建模与优化问题展开工作,研究的主要内容包括三个方面:自适应混沌粒子群优化和支持向量机结合的模型预测、基于自适应混沌粒子群优化和支持向量机的最优化设计框架以及相空间重构和支持向量机结合的混沌时间序列模型预测。本文的研究内容属于信息科学、自动化科学、化学科学等学科的交叉领域。本文的主要创新性工作包括:(1)提出一种基于自适应混沌粒子群优化和支持向量机结合(Adaptive Chaos Particle Swarm Optimization-Support Vector Regression,ACPSO-SVR)的预测算法,引入ACPSO启发式寻优机制对SVR模型的超参数进行自动选取,在超参数取值范围变化较大的情况下,效果明显优于网格式搜索算法。选取UCI机器学习数据库中的Forest fires标准数据集进行测试,实验结果表明该方法具有较高的精度和良好的泛化能力,对于解决多变量的回归预测问题是一种有效的方法。将所提出的建模方法用于Cu-Zn-Al-Zr二甲醚合成催化剂建模,在反应动力学模型未知的情况下,同时获得了催化剂组份模型和动力学模型,取得了良好的预测效果。(2)提出一种基于多目标混沌粒子群优化和支持向量机相结合的最优化设计框架。针对多相催化反应具有多阶段、强相关性,获取催化剂性能指标非常困难的特点,将训练好的SVR模型作为最优化设计框架中的适应度评价近似模型,通过多目标ACPSO算法同时优化输入变量的空间,寻求具有全局最优催化性能的催化剂。将该策略用于Cu-Zn-Al-Zr二甲醚合成催化剂的研发,实验结果表明,经ACPSO-SVR最优化设计方法给出的两组新型催化剂的性能指标与实验测试值误差很小,可以缩短催化剂研发的时间,节约资金和时间消耗,不失为一种可行的、有效的实验室设计催化剂新方法。(3)提出一种基于相空间重构和支持向量机结合(Phase Space Reconstruction- Support Vector Regression,PSR-SVR)的非线性时间序列预测建模方法。针对多相催化剂在非定态下的复杂失活机理及活性受多种因素影响,获取催化剂失活过程的时间序列数据非常有限而降低建模效率和预测精度的情况,提出以高维相空间重构的数据表示方法重构失活数据序列,达到改写数据的规律性进而探讨催化剂失活内在复杂本质特征的目的。将该建模方法应用于甲醇氧化羰基化反应中Cu-Si-Al碳酸二甲酯合成催化剂失活过程建模,仿真结果表明催化剂失活模型的预测误差在满意的范围之内,给出的碳酸二甲酯时空收率的预测值可以为反应器的正确设计和操作、反应过程的优化提供有效信息。更多还原

【Abstract】 Modeling and optimization are ones of the central issues in engineering and technological research, which can be employed as a significant prerequisite for scientific decision-making and planning. However, prediction is the weakness for all technologies including heterogeneous catalysis. With the development of economy, particularly the requirements for new energy and materials, the heterogeneous catalysis faces new challenges. During the development of new efficient catalysts, besides economic, safety and multi-function, the study of catalysts properties from a whole chemical process is also required. Machine learning and data depth mining based on history data has become the most important problem to be solved in the area of chemical engineering information. The difficulty and three major problems to be solved in the area of catalysis are the kinetic model, the structure-activity relation model, as well as design and optimization of catalysts. It is of great practical importance not only in the investigation of catalyst characters, but in the whole procedure of the control, optimization and modeling of chemical production.In this study, chaos and support vector regression (SVR) were employed to the modeling and optimization of heterogeneous catalysis. The major research contents includes modeling prediction based on the adaptive chaos particle swarm optimization and support vector regression (ACPSO-SVR), the design of optimal optimization algorithms based on ACPSO-SVR, and the modeling prediction of chaotic time series based on phase space reconstruction and support vector regression (PSR-SVR). The present study crosses the boundaries of numerous scientific fields of information, automation and chemical engineering etc.The main contents and basic findings are as follows:(1) An effective relevance prediction algorithm based on ACPSO-SVR was presented. A heuristic optimization method was introduced to automatic selection of hyper-parameters in SVR. The forest fires standard data set of UCI machine learning database was selected to test. The experimental results showed that the new method has high relatively precision and good generalization ability with a wide range of parameter values, better than that of mesh searching algorithm. It could be used as an effective method to solve the problems of multivariate regression predication. The method was applied to modeling of the Cu-Zn-Al-Zr based catalysts for the synthesis of dimethyl ester (DME). Under the condition of with unknown kinetic model, the catalyst composition model and the kinetic model were obtained and gave good prediction results was produced.(2) A new optimization framework was built by combining multi-objective chaotic particle swarm optimization algorithm with SVR. Considering that it is very difficult to obtain the catalytic performance due to mutli-stage process and strong relevance in heterogeneous catalysis, the trained SVR Model was used as the fitness approximation model, through using multi-object ACPSO algorithm optimizing the space of input variable for searching catalyst with optimal catalytic performance. The optimum strategy was employed in the exploring of the Cu-Zn-Al-Zr based catalysts for DME synthesis. The two new types of catalysts obtained through ACPSO-SVR optimized scheme show catalytic performance close to the experimental value. In conclusion, this optimization method shorts time and reduces cost in catalyst exploration, and is exactly an effective method for catalyst development in laboratory.(3) A novel method for nonlinear time series forecasting of the catalyst deactivation based on PSR-SVR was presented. Considering the complicated mechanisms of catalysts deactivation multi-factors influencing the catalyst performance; moreover, the limitation of getting time series data during the deactivation process of catalysts leads to the modeling efficiency and prediction precision. The data presentation method of high dimensions phase space reconstruction was used to assessment of the deactivation data and the regularity of data recomposition is attained. Finally, the aim of disclosing the nature of complicated character of catalyst performance was realized. The novel method was applied to predict the deactivation process of the Cu-Si-Al based catalysts for the synthesis of dimethyl carbonate (DMC). The simulation results showed that the prediction error of catalyst deactivation model is in a range of tolerance. The prediction space time yield (STY) value of DMC could provide important information for the design and operation of reactors as well as the optimization of the reaction conditions.更多还原