Abstract:

By screening the auxiliary components and preparation methods, a Fe₃O₄ composite oxide catalyst was prepared for production of hydrogen from methane. Based on the properties of neural network, an improved back-propagation network model was developed to simulate the relationship between components of the catalyst and catalytic performances of the catalyst life and the formation rate of hydrogen. The model network structure and the computer-aided design procedures were established after investigation of the structural organization, the training method, the activation function, and the generalization ability of artificial neural network. Upon used the Levenberg-Marquardt training method, the network convergence was improved significantly and a formulation model of artificial neural network was achieved with strong generalization capability. To further enhance efficiency of catalyst design, a hybrid genetic algorithm was employed for global optimization of design. After six cycles of design optimization, a series formulation of Fe₃O₄ composite oxide catalysts for production of hydrogen from methane was developed. When one of the optimized catalysts was applied in hydrogen production, the catalyst life and formation rate of hydrogen were 4.46 h and 1.16 mmol·min^-1·(g Fe)^-1, respectively, which were better than those of previously reported catalysts.

Key words: methane, hydrogen production, catalyst, artificial neural network, genetic algorithm

摘要：

通过筛选辅助组分和制备方法，制备了一种用于甲烷直接制氢的Fe₃O₄复合氧化物催化剂。应用人工神经网络建立了该催化剂的配方模型，对人工神经网络模型学习算法、激活函数以及网络结构进行了考察，确定了该催化剂辅助设计的步骤及模型的网络结构，将Levenberg-Marquardt方法用于网络的训练，改进了网络的收敛特性，最终获得了泛化能力较强的人工神经网络配方模型。以建立的模型为目标函数，采用改进的混合遗传算法作为优化方法，经过6轮优化，获得了一系列较优的甲烷直接制氢的Fe₃O₄复合氧化物催化剂配方。选用其中一种优化获得的配方进行甲烷制氢反应，催化剂寿命和氢气生成速率分别达到4.46 h和1.16 mmol·min^-1·（g Fe）^-1，优于以往报道的催化剂。

关键词: 甲烷, 制氢, 催化剂, 人工神经网络, 遗传算法