Thermodynamics-based neural network and the optimization of ethylbenzene production process

Abstract

Thermodynamics is significantly important for analyzing properties and establishing mechanism models of chemical process. Based on the characteristics of the reaction and distillation process and their similarity with back-propagation process, improved BP neural network models are built for reactor and distillation column. In the reactor model, the chemical potential is used to adjust the errors of output nodes. For distillation column, the errors of output nodes corresponding components are adjusted according to their gas-liquid equilibrium constant. The models have improved predictive capability and convergence. An ethylbenzene unit is simulated and optimized by the proposed method, and its heating utility consumption is reduced by 55.2%.