Methanol production reactor simulation and optimization under kinetic parameter uncertainty conditions

Abstract

A catalytic reaction process for producing methanol from carbon dioxide and hydrogen gases has been suggested and simulated. However, there can exist parametric uncertainties on the process model such as reaction kinetics. A reactor model considering parametric uncertainty results in a distributional process output and it can give more informative data compared to the conventional modeling methods which use a single parameter set. However, the distributional model needs a lot of computational loads because of the Monte Carlo simulation and iterative calculations for convergence. In order to alleviate the heavy computational load and reflect the skewness of the distributional data, generalized extreme value distribution and neural network technique are utilized. The formation parameters of generalized extreme value distribution are learned by shallow and deep structured neural network and as a result distributional reactor model in an explicit formulation is proposed. Compared to the result using shallow structured neural network for learning the formulation parameters, that using deep neural network shows improved predictive performance especially adjacent to the boundary layers of process inputs. The proposed model can be utilized to real-time stochastic model based approaches in optimization and control with less computational load because of its explicit and distributional formulation.