(172f) Distributed Model Predictive Control of Nonlinear Processes Using Machine Learning Models

As machine learning techniques have demonstrated their effectiveness in traditional engineering applications, its usage in model identification in process control and operations has garnered increasing interest in recent years [2,3]. Particularly, recurrent neural networks (RNN) have been widely utilized to model nonlinear dynamical systems [4,5]. Furthermore, given the increased dimension and complexity of the control network with both networked and point-to-point sensor communications, distributed control systems have been developed to carry out calculations in separate processors while achieving the closed-loop plant objectives cooperatively and efficiently [1].

While model-based controllers provide a natural control framework for distributed control systems due to their abilities of handling multi-variable ineractions and constraints, the performance of the Distributed Model Predictive Control (DMPC) systems will be heavily dependent on the accuracy of the model, which may not be always available or remain accurate throughout the plant operation. In this work, we develop a Lyapunov-based Distributed Model Predictive Control system using RNN prediction models. A hierarchical DMPC architecture is proposed which allows easy integration of new control loops with pre-existing stabilizing control loops to further improve closed-loop performance. Working with a general class of nonlinear models, and assuming that there exists a Lyapunov-based controller that stabilizes the nominal closed-loop system using only the pre-existing control loops, two separate Lyapunov-based MPC’s are designed, coordinating their actions in an efficient manner, to improve overall closed-loop performance while preserving the stability properties and reducing computational effort relative to that required in a centralized MPC design. With a dataset generated from extensive open-loop simulations within the desired operating region, RNN models are trained with a sufficiently small modeling error such that closed-loop state boundedness and convergence to the origin can be achieved. The proposed RNN-DMPC framework is applied to a nonlinear chemical process example consists of two continuously stirred tank reactors connected in series.

[1] Christofides, P. D., Liu, J., and Muñoz de la Peña, D. Networked and distributed predictive control: Methods and nonlinear process network applications. Springer Science & Business Media, 2011.

[2] Schmidhuber, J. Deep learning in neural networks: An overview. Neural networks, 2015, 61: 85-117.

[3] Venkatasubramanian, V. The promise of artificial intelligence in chemical engineering: Is it here, finally?. AIChE Journal, 2019, 65: 466-478.

[4] Wang, Y. "A new concept using LSTM Neural Networks for dynamic system identification." In Proceedings of American Control Conference (ACC), pp. 5324-5329. IEEE, 2017.

[5] Wu, Z., Tran, A., Rincon, D., Christofides, P.D., 2019. Machine learning-based predictive control of nonlinear processes. part I: Theory. AIChE Journal 65, e16729.