Prediction Of Catalytic Reforming Based On BP Neural Network For CH ₄ -CO ₂

CH4-CO2 reforming reaction can transforms two kinds of greenhouse gases (CH4 and CO2) into synthesis gas (H2 and CO) effectively, which can achieve CO2 emission reduction and high efficient utilization of C1 resources. Considering the environmental protection and utilization of CO2, this research has very important value. The reaction temperature, CO2 and CH4 feed ratio, catalyst type and content, carrier, auxiliary agent, modifier and so on were widely researched presently. Many factors have effect on the reaction system, and their influence is nonlinear, which makes the system very complex. Theoretical simulation is an important supplement to the experiment, it can provide guide and predication for the experiment. Artificial neural network has very strong nonlinear mapping ability among many theoretical simulation methods. So we use the artificial neural network (ANN) to establish prediction models, which can provide reference for the optimization of the reaction conditions, and the selection of the catalyst species. According to the experimental data of CH4-CO2 reforming reaction with three types catalyst, we established three prediction models respectively.In the activated carbon modified by H2O2 catalytic CH4-CO2 reforming prediction model, we selected reaction temperature, the mesoporous area of catalyst, the amount of phenolic hydroxyl group and ether oxygen group as the the network inputs.The network outputs were CH4 initial and stable conversion rate, CO2 initial and stable conversion rate. Here, BP neural network prediction models were established with multiple outputs and single output. By comparison of the mean square error of training samples and validation samples, it indicates that the two models have the similar predictive power and both the average prediction errors are less than 12%. And sensitivity analysis is carried out for the single output prediction model. The analysis results show that the temperature has the largest effect on reaction, and the influence of the physical structure of catalyst, including mesoporous area, phenolic hydroxyl and ether groups, is relatively small.In the Ni/Al2O3 catalytic CH4-CO2 reforming prediction model, we established pediction models of temperature and space velocity. These two models were respectively trained by BP algorithm and improved BP algorithm. In the temperature model, we selected temperature, the content of Ni, the catalyst surface area, pore volume and pore size as the network inputs, and CH4 conversion, CO2 conversion and H2/CO ratio as outputs. For BP model and improved BP model, their mean values of prediction error are less than 1.5%. Sensitivity analysis of the models shows that, temperature and Ni content have large influence, while the physical structure of the catalyst has smaller effect. In the space velocity model, only the temperature was changed into the space velocity in network inputs, the other same with the temperature model. For BP model and improved BP model, their mean values of prediction error are less than 3.7%. Sensitivity analysis of the models shows that, Ni content has the greatest impact on the reaction.The second is space velocity, pore size and specific surface area, which have similar effects. Pore volume has the smallest influence. In both temperature and space velocity prediction model, improved BP model is better than the BP model in view of the stability, convergence speed and prediction accuracy.In the Ni/ZSM-5 catalytic CH4-CO2 reforming prediction model, we selected temperature, Ni content, Ce content and CO2/CH4 feed ratio as the network inputs.The network outputs were CH4 conversion, CO2 conversion, H2 selectivity and H2/CO. Here, three models were established, including BP neural network prediction model, GA-BP prediction model based on genetic algorithm optimization, PSO-BP prediction model based on particle swarm optimization.In these three models, the average relative errors of of the validation sample are less than 9%. For GA-BP model and PSO-BP model, their prediction ability is improved by comparing with BP model. Sensitivity analysis of the three prediction models shows that, CO2/CH4 feed ratio has the biggest influence on the reaction results.The second is the temperature and Ni content.Ce content plays the smallest effect.