Architectures for neural networks as surrogates for dynamic systems in chemical engineering

Abstract

Surrogate models for dynamic systems in chemical engineering are increasingly of interest. Neural networks have already been applied in research, but it remains unclear which types of neural network architectures are actually required for practical systems. The focus here lies on recurrent neural networks of type Jordan, Elman, and LSTM layers. These are investigated for different types of data sets as training basis: batch trajectories, data of a proper excitation of a continuous process, and a typical operation trajectory of a large chemical plant. To ensure a rigorous investigation, hyperparameter tuning by Bayesian and Bandit optimization is included. As a first, a dynamic surrogate model using LSTM layers for a batch distillation system is presented, which is valid from start-up until shutdown. The evaluation shows further need for adjustments in data preparation and objective/loss function compared to the state of the art.