Abstract:

A prediction model between the content of promoter and the activity of catalyst was established by BP neural network, with which the promoter of Fe_1-x O ammonia synthesis catalyst was optimized. Firstly, the preliminary experimental data were summarized into five types of catalysts including three, four, five, six and seven promoters. With the content of the promoters (volume fraction) as the input model variable and the ammonia concentration (reactivity) at the outlet of the reactor at 425℃ as the output one, the formula of the promoter was optimized. The results showed that maximum mean square error of fitting values of BP neural network prediction model was 0.2784, while that of the predicted values was 0.1592, indicating the accuracy of the BP neural network model was high. On the basis of this model, multiple population genetic algorithm was used to search the extreme value, and the optimal catalyst formula was obtained and verified by experiments. The maximum relative error between the measured values of 5 samples prepared according to the optimized formula and the predicted ones was 2.88%. The highest activity was 18.83% for the catalyst containing seven promoters, 1.31% higher than the average reactivity value of the original sample (17.52%), and a relative increase of 7.48%.

Key words: Fe_1-x O, catalyst, promoters, neural networks, genetic algorithm, optimization

摘要：

运用BP神经网络建立了助催化剂含量与催化剂活性之间的预测模型，对Fe_1-x O基氨合成催化剂的助催化剂进行优化。首先将前期实验数据整理归纳为含有3、4、5、6和7个助催化剂等5类催化剂，以助催化剂含量（体积分数）为输入变量，以425℃反应器出口氨浓度（活性）为输出变量，对助催化剂进行优化。结果表明，BP神经网络预测模型拟合值均方误差最高为0.2784，预测值均方误差最高为0.1592，构建的BP神经网络模型准确度较高。在该模型的基础上，运用多种群遗传算法进行极值寻优，求解最优的催化剂配方，并进行实验验证。结果表明，根据优化结果制备5个样品的实验测定值与预测值的相对误差最高为2.88%，优化结果较为准确；含有7个助催化剂的催化剂活性最高为18.83%，比原样本的统计平均活性值（17.52%）高1.31%，相对提高7.48%，助催化剂含量优化取得满意的结果。

关键词: Fe_1-x O, 催化剂, 助催化剂, 神经网络, 遗传算法, 优化