Process structure-based recurrent neural network modeling for predictive control: A comparative study

Abstract

Recurrent neural networks (RNN) have demonstrated their ability in providing a remarkably accurate modeling approximation to describe the dynamic evolution of complex, nonlinear chemical processes in several applications. Although conventional fully-connected RNN models have been successfully utilized in model predictive control (MPC) to regulate chemical processes with desired approximation accuracy, the development of RNN models in terms of model structure can be further improved by incorporating physical knowledge to achieve better accuracy and computational efficiency. This work investigates the performance of MPC based on two different RNN structures. Specifically, a fully-connected RNN model, and a partially-connected RNN model developed using a prior physical knowledge, are considered. This study uses an example of a large-scale complex chemical process simulated by Aspen Plus Dynamics to demonstrate improvements in the RNN model and an RNN-based MPC performance, when the prior knowledge of the process is taken into account.