ANN for hybrid modelling of batch and fed-batch chemical reactors

An unconventional modelling methodology based on artificial neural networks is proposed to rapidly develop a model from data obtained during different batch experiments. The objective of the global model is to predict time evolution of concentrations of all species present in the reaction medium. For this, different recurrent neural networks are elaborated to estimate a particular species as a function of operating parameters and concentrations of all species and then assembled in a complex global model. To validate the approach, the esterification reaction of methanol by acetic acid, which presents equilibrium, has been chosen. The kinetic evolution of the chemical species during experiments conducted in batch mode are correctly represented whatever the operating conditions. Finally, the global model based on neural networks is integrated in a hybrid model. This permits to transpose the reaction to a semi-batch chemical reactor which has not been considered during the learning phase. (c) 2021 Elsevier Ltd. All rights reserved.

Automated summary

A novel modelling methodology using artificial neural networks is proposed to develop a global model for predicting the time evolution of concentrations of all species in the reaction medium. Different recurrent neural networks are utilized to estimate specific species as a function of operating parameters and concentrations, which are then assembled into a complex global model. This approach successfully represents the kinetic evolution of chemical species under varied operating conditions, and the neural network-based global model is integrated into a hybrid model for transposing reactions to a semi-batch chemical reactor.