(497a) Distributed Model Predictive Control of Complex Plants: A Systematic Study of Decomposition Effects

The decomposition of a large scale complex plant into constituent subsystems is an essential step for the synthesis of distributed model predictive control (DMPC) architectures, which directly affects the closed-loop performance and computational effort [1-3]. While there have been some initial attempts to address the decomposition problem for specific classes of systems [4, 5], developing a systematic framework for optimal decomposition into subsystems with desirable interaction characteristics is still an open problem.

From a network theory perspective, this problem can be posed as identifying weakly-connected subsystems whereby the variables of each subsystem are strongly connected. Such a point of view has been adopted to generate hierarchies of system decompositions with different degrees of decentralization based on input-output connectivity, following an agglomerative or divisive clustering procedure [6]. An alternative method based on the maximization of the modularity of the system equation graph was recently proposed to decompose the network into ”communities” that include inputs, states, and outputs [7]. The latter approach minimizes the interactions between the resulting communities and appears well suited for distributed control.

In this work, we provide a systematic evaluation of the impact of system decomposition on DMPC designs for a benchmark integrated plant. An iterative DMPC architecture is applied to a reactor-separator system containing two continuous stirred tank reactors and a vapor-liquid separator. The performance and computational effort of implementing the iterative DMPC for different decompositions identified through community detection methods [6, 7] or by intuition [3, 8], are analyzed and compared to those obtained by a centralized model predictive control (CMPC) design. State/output regulation and tracking case studies are considered. In each case, it is documented that the optimal decomposition obtained by the community detection method, which minimizes the interactions between the subsystems, enables closed-loop performance close to that of the CMPC, while reducing the computation effort significantly.

[1] P.D. Christofides, R. Scattolini, D. Munoz de la Pena, and J. Liu. Distributed model predictive control: A tutorial review and future research directions. Computers & Chemical Engineering, 51:21–41, 2013.

[2] B.T. Stewart, S.J. Wright, and J.B. Rawlings. Cooperative distributed model predictive control for nonlinear systems. Journal of Process Control, 21:698–704, 2011.

[3] P.D. Christofides, J. Liu, D. Munoz de la Pena. Networked and Distributed Predictive Control. Springer-Verlag, London, 2011.

[4] W. Al-Gherwi, H. Budman, and A. Elkamel. Selection of control structure for distributed model predictive control in the presence of model errors. Journal of Process Control, 20:270–284, 2010.

[5] N. Motee and B. Sayyar-Rodsari. Optimal partitioning in distributed model predictive control. In Proceedings of the American Control Conference, 5300–5305, Denver, CO, 2003.

[6] S. Heo and P. Daoutidis. Control-relevant decomposition of process networks via optimization-based hierarchical clustering. AIChE Journal, 62(9):3177–3188, 2016.

[7] S.S. Jogwar and P. Daoutidis. Community-based synthesis of distributed control architectures for integrated networks. Chemical Engineering Science, Submitted, 2016.

[8] M.J. Tippett, J. Bao. Distributed model predictive control based on dissipativity. AIChE J., 59(3):787– 804, 2013.