(91b) Partially-Connected Recurrent Neural Network Modeling for Predictive Control Using Noisy Data

Over the last four decades, neural networks have demonstrated their ability to provide adequate modeling approximations of complex, highly nonlinear, and stochastic chemical processes when utilized as machine learning (ML) techniques for control and state prediction of nonlinear processes [1,2,3]. Recurrent neural networks (RNN) have been widely used for modeling a broad class of dynamical systems with objectives of control and state estimation. In our previous work [4], a model predictive control (MPC) that uses a process structure-based RNN model, termed partially-connected RNN, was proposed and achieved improvements in closed-loop performance and computational time relatively to the standard (i.e. fully-connected) RNN models. However, in the development of the partially connected RNN models it is assumed that both the given system and the data are noise-free. In fact, state measurements in a variety of chemical industrial operations are corrupted by noise. Therefore, it is necessary to study the effect of noise when utilizing the underlying modeling approach.

This work investigates the effect of both gaussian and non-gaussian noisy data on the performance of partially-connected RNN model approximation of class on multi-input-multi-outputs systems. Furthermore, two different RNN building methods, specifically Monte Carlo dropout rate and co-teaching methods, are utilized in development of partially connected RNN-LSTM. These two techniques are employed to improve the open-loop accuracy as well as the closed loop performance under Lyapunov based MPC. Aspen Plus Dynamics, which is a well-known high-fidelity software, is used to develop a large-scale chemical process example in order to demonstrate the anticipated improvements in both open-loop approximation and controller performance in two cases: gaussian noise, and non-gaussian noise. In the presence of each noise type, both open-loop and closed-loop simulation will be carried out, analyzed, and discussed.

References:

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

[2] Mohanty, Swati. "Artificial neural network based system identification and model predictive control of a flotation column." Journal of Process Control 19.6 (2009): 991-999.

[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] Alhajeri, M., J. Luo, Z. Wu, F. Albalawi and P. D. Christofides, "Process Structure-Based Recurrent Neural Network Modeling for Predictive Control: A Comparative Study,'' Chem. Eng. Res. & Des., 179, 77-89, 2022.