【作者】 黄凯；

【导师】 吕德伟； 陈丰秋；

【作者基本信息】 浙江大学 ， 化学工程， 2002， 博士

【摘要】 随着社会的进步，人们对资源利用的要求越来越高，同时为可持续发展，对环境保护也越来越关注。催化剂在资源利用和环境保护方面因而日益发挥着巨大的作用，而能否设计出高活性、高选择性的催化剂配方，与催化剂配方设计方法的好坏有密切的关系，特别是对于日益重要的多组分催化剂体系而言。 传统催化剂设计方法的缺点很难在较短时间设计出优秀的配方，且实验工作量大。而基于专家系统的计算机辅助催化剂设计方法需要建立催化知识库和推理机，从效率上难以满足催化剂设计的要求。针对这些催化剂设计方法的缺点，在前人开发的人工神经网络辅助催化剂设计方法的基础上，参照前人对人工神经网络结构(包括隐含层个数和隐含层节点个数)的研究，确定了用于催化剂配方建模的人工神经网络隐含节点数选择的基本原则；选用Sigmoid函数作为用于建模的人工神经网络各层的激活函数；通过对传统BP算法、矩方法改进的BP算法和Levenberg-Marquardt方法改进的BP算法等三种神经网络学习算法的比较，认为Levenberg-Marquardt方法的收敛速度较前两种算法有明显的优势，且对初始权值的依赖较前两种算法要小，因此确定在催化剂配方建模中采用基于Levenberg-Marquardt方法的BP算法人工神经网络的学习算法；为了提高神经网络模型的精确度，确定了多轮次逐步训练的策略，使催化剂模型朝正确的方向逼近；将SWIFT方法应用于网络模型局部最优点的求解，应用计算证明这种方法的寻优速度很快，适用于有多个局部最优的问题，并将此方法得到的催化剂配方作为下一轮进一步训练的样本；针对标准遗传算法收敛速度慢、局部寻优能力差的缺点，首次将SWIFT方法和遗传算法结合构成混合遗传算法(Hybrid GA)，提高优化的速度以设计出最优的催化剂配方。基于上述各点，首次提出了一种基于人工神经网络和混合遗传算法的催化剂配方设计方法。该催化剂设计方法通用性较强，无须过多初始样本，建模效率较高，特别适合多组分催化剂配方的设计。 同时为了进一步验证和完善所建立的计算机辅助催化剂配方设计方法，并考虑到甲烷氧化偶联作为天然气综合利用的重要途径之一，而关键的催化剂开发却难有突破，故选择多组分甲烷氧化偶联配方优选作为上述设计方法的首个应用对象。而通过考察前人甲烷氧化偶联催化剂研究的成果，认为组分较少的催化体系难以胜任甲烷氧化偶联反应，而由多种过渡金属相互协同、用相应元素助催化、并用碱金属离子修饰的多组分催化剂有可能获得良好的催化性能。因此选择Zr和Mn元素作为所开发催化剂的主要组分；同时为了抑制Zr和Mn元素的深度氧化特性，用S、P和W元素作为助催化剂，并用碱金属元素进行修饰以提高反应中C2烃的选择性。由此首次设计了一种用于甲烷氧化偶联反应的六组分催化剂。 进而，为了给基于人工神经网络和混合遗传算法的催化剂配方设计方法提供I 有效的初始样友隼，针对该六组分催化剂，应用正交设计的方法设计了25个配1 方，并在于惰件气体稀释的条件下，分别进行了考评。从此25个催化剂中找到ICfu：O，一3：l，辰应渴度为 1069 K条件下，Cfu的转化率达到21．38％，C，选【择。旺达到82．56O（C，烃收率为17．65％｝，且在10 k内活性基本不变。通过对反应 发现Mll和Zr元素的含量对Cth转化率影响较大。同时证明多组分催化剂有可 能获得较好的甲烷氧化偶联催化性能。 随后，将基于人工神经网络和混合遗传算法的催化剂配方设计方法应用于此a 六组分甲烷氧化偶联催化剂配方的优选。以催化剂六个组分的摩尔含量为输入袁 层，第一和第二隐含层的节点分别为24和9，并以催化性能（包括甲烷转化率和nMC。烃选择性购输出层构成了6—24—9—2型网络用于该催化剂配方建模。g 基于所建立的催化剂配方模型，通过六轮的训练－优化－再训练过程，不儿 仅使C。烃收率的实验结果与网络预测结果的误差控制怀p3％范围内（绝对数值卜pn沉 而且设计出了多种优秀的多组分甲烷氧化偶联催化剂配方，其中催化性能最为突；出的两种催化剂在前述的反应条件下，CH4转化率分别达到3 6．91％和37．79％，ge。烃选择性分别达到v】soo，o和刀．soo，K。烃单程收率分别为刀．oso，o和刀．vss卜及忐 在没有用情性气体稀释的同等条件下。该结果已高于目前文献报道的最好结果。＄此外，针对直接神经网络建模的一些缺点，并进一步提高建模的效率，探索兰 了将PLS方法与神经网络相结合卿NNPLS方法），并应用于建立多组分甲烷氧甲愧 化偶联催化剂的模型。与直接神经网络建模相比，NNPLS方法压缩分解了变量，＄减少了计算量，同时使模型的推广能力得到提高，有效地改善了直接神经网络建8 模过程中催化剂模型泛化能力较差的缺点。与线性PLS相比，N’’a－－LS方法的模归 型精度较高，有利于以后的催化剂雌。这方面的工作是对神经网络催化剂建模s 过程的延续和改进，有待进一步深入的研究。R 最后，基于人工神经网络的他化剂配方模型均是黑箱模型o mdel，对t＋中 反应过程中催化剂表面的特更多还原

【Abstract】 With exhaust of crude oil, it becomes a main topic to find another energy. Because of larger reserves, nature gas would be replace oil to become main chemical raw material in the next years.Undoubtedly, by using methane in natural gas as reactant, it will become a strategic topic to synthesize ethylene, which is the most important raw material in chemical industry at the present time. But it should be known that design of catalyst is key to develop the chemical process. By reviewing a great deal of catalysts developed by the former researchers, it shows that the catalytic system contained 2-3 chemical elements could not get better catalytic performance, and the catalyst, which contains some interactional transition metal elements, is assistant-catalyzed by corresponding elements arM is modified by alkali metal ions, could show better performance in oxidative coupling of methane.By reviewing the main components in catalysts of Syngas, it could be found that transition metal elements could activate methane molecule efficiently, so. Zr and Mn were chosen as main components of the developing catalyst. At the same time, S, P and W were chosen as assistant-catalyst components and the catalyst was modified by some alkali metal ions in order to suppress deep oxidation of C2. A multi-component catalyst is designed based on the above considerations.25 multi-element catalysts for oxidative coupling of methane were designed by orthography-design method. And 25 catalysts were tested on a fixed-bed micro-reactor without inert gases as thinner. A better catalyst was found in these catalysts. Experimental results showed that, the Cl]4 conversion would be 21.3 8% and C2 selectivity would be 82.56% (C2 yield be 17.65%) when GHSV was 33313 rnFg%f’, Cl]4 02 was 3 1 and reaction temperature was 1069 K, the catalytic active kept constant in 10 hr. A maximum CH4 conversion could be found when temperature increased gradually on some catalysts, which reason could be increasing of the parallel side-reactions. The influence of different element on catalytic performance was researched primarily.In order to design a best catalyst, artificial neural network and genetic algoritlun are applied to design a multi-component catalyst for oxidative coupling of methane. Especially for multi-component catalyst, the relations of different components m catalyst could be complex and high nonlinear, and could not be expressed by a certain analytic function. Artificial neural network is a high-nonlinear system and have stronger map function. If the complex internal relations of catalyst components couldbe expressed by a certain neural network, the relations between catalyst components and catalytic reaction results (such as conversion of reactant and selectivity of product) would be established by the neural network which have a certain structure, such as the number of hidden layer(s) and the number of neurons. The relations expressed by neural network can be seen as a robust catalytic reaction model. Then SWIFT (Sequential Weight Increasing Factor Technique) method which could find local maximum, and a hybrid genetic algorithm which could get global maximum, would be applied to design some better catalysts. Since the number of catalysts in the training group is not larger enough to obtain the best generalization ability for the neural network at the beginning of design, it is unavoidable to adjust the weight matrix of the trained neural network. The results of first optimization should be added to the training group, and the neural network should be trained again by using the weight matrix obtained in first training as initial weight matrix. Based on the adjusted model, some other better catalysts would be designed by SWIFT method and hybrid genetic algorithm. Rest may be deduced by analogy until the predicted results of neural network approach the experimental results of the optimal catalyst.Back-propagation algorithm, which is a sort of neural network, is widely applied in many fields, and also used in this paper. Bu更多还原