(345t) Process Structure-Based Recurrent Neural Network Modeling for State Estimation and Predictive Control of Nonlinear Processes

Neural networks have proven their capability, as machine learning (ML) techniques, in providing good modeling approximations of complex, highly nonlinear and uncertain chemical processes that have been used for control and state estimation of nonlinear processes over the last four decades [1,2,3]. Specifically, recurrent neural networks (RNN) have been broadly employed for modelling a general class of dynamical systems for control and state estimation purposes. Model predictive control (MPC) based on RNN was proposed in [3] assuming that all the states are measurable. As a matter of fact, full state measurements may not be available in numerous chemical process applications, thus it would be challenging to apply ML based control strategy to such processes [4]. Therefore, to address this problem, an RNN-based state observer and predictive control framework for nonlinear systems was developed in [5] utilizing fully connected RNN models. Although the approximation accuracy of the fully connected RNN developed as a black box was satisfying, there is still significant room for improvement of RNN model approximation by developing the RNN structure using information about the interdependence of process state variables.

This work utilizes two different RNN building methods described in [6] to develop RNN models accounting explicitly for process structure information. The first method determines the RNN structure based on a priori physical knowledge of a given process, and it is implemented via a partially connected RNN modelling method. The second one is the weight constrained RNN model, which is developed by employing weight constraints on the RNN training process optimization problem to remove network links that do not significantly contribute to the network prediction accuracy. The two developed RNN models-separately- are utilized in an RNN-based state observer to estimate the unmeasured states. Furthermore, a model predictive controller is developed that uses the RNN models as the prediction models to calculate the optimal control action. A chemical process example is used in order to demonstrate improvements in approximation performance and its impact in state estimation and control using physics-based RNN and weight constrained RNN models compared to the fully connected RNN model.

[1] Sontag, E. D. "Neural nets as systems models and controllers." Proc. Seventh Yale Workshop on Adaptive and Learning Systems, 1992.

[2] Mohanty, S. Artificial neural network based system identification and model predictive control of a flotation column. J. Process Control, 19, 991−999, 2009.

[3] Wu, Z., A. Tran, D. Rincon and P. D. Christofides, "Machine Learning-Based Predictive Control of Nonlinear Processes. Part I: Theory,'' AIChE J., 65, e16729, 2019

[4] Bequette, B.W., Nonlinear control of chemical processes: A review. Industrial & Engineering Chemistry Research, 30, 1391-1413, 1991.

[5] Alhajeri, M. S., Z. Wu, D. Rincon, F. Albalawi, and P. D. Christofides, Machine-learning-based state estimation and predictive control of nonlinear processes. Chemical Engineering Research and Design, 167, 268-280, 2021.

[6] Wu, Z., D. Rincon and P. D. Christofides, Process structure-based recurrent neural network modeling for model predictive control of nonlinear processes. Journal of Process Control, 89, 74-84, 2020.