(504c) Architectures for Distributed Model Predictive Control of Nonlinear Process Systems

Augmenting dedicated, local control systems (LCS) with control systems that may utilize real-time sensor and actuator networks gives rise to the need to design/redesign and coordinate separate control systems that operate on a process. Model predictive control (MPC) is a natural control framework to deal with the design of coordinated, distributed control systems because of its ability to handle input and state constraints, and also because it can account for the actions of other actuators in computing the control action of a given set of control actuators in real-time. With respect to available results on distributed MPC design, several distributed MPC methods have been proposed in the literature that deal with the coordination of separate MPC controllers that communicate in order to obtain optimal input trajectories in a distributed manner.

From augmentation of existing LCS with NCS point of view, in our previous work [1], we introduced a distributed model predictive control method for the design of NCS where one Lyapunov-based MPC (LMPC) was designed for the LCS and one LMPC was designed for the NCS. The distributed MPC design requires one-directional communication between the two distributed controllers and was proved that it guarantees practical stability of the closed-loop system and has the potential to maintain the closed-loop stability and performance in the face of new or failing controllers/actuators and to reduce computational burden in the evaluation of the optimal manipulated inputs compared with a fully centralized LMPC of the same input/output-space dimension.

In the present work, our objective is to extend our recent results [1] to develop distributed model predictive control methods including multiple controllers in both LCS and NCS for nonlinear systems. We propose two different approaches to design the distributed model predictive control. In the first approach, different MPCs are evaluated in sequence; each MPC is evaluated once at a sampling time; and only one directional communication is required. In the second approach, different MPCs are evaluated in parallel; each MPC is evaluated once (decentralized, no communication among controllers is needed) or more than once (iterative communication) at a sampling time depending on iteration times; and two directional communication is required. In both approaches, the distributed MPCs are designed via Lyapunov-based model predictive control. Sufficient conditions are given for each approach under which the state of the closed-loop system is guaranteed to be maintained ultimately in an invariant region including the origin. Extensive simulation studies of the two proposed distributed MPC designs are carried out through an alkylation of benzene with ethylene process.

[1] J. Liu, D. Munoz de la Pena, and P. D. Christofides. Distributed model predictive control of nonlinear process systems, AIChE Journal, vol. 55, pp.1171-1184, 2009.