(40c) Subsystem Decomposition of Process Networks for Simultaneous Distributed State Estimation and Control

A process network comprises interconnected operating units that are coupled through material, energy and information flows. Advanced control strategies are typically used for these process networks to ensure operational safety and profitability and to meet the regulations on environmental sustainability [1, 2]. Distributed control has been recognized as a very promising advanced control framework for such process networks [1, 3]. Within the distributed control framework, a process network is decomposed into subsystems, and then a local controller is developed for each subsystem [4, 5]. The local controllers communicate with each other to coordinate their actions [3]. The distributed control framework inherits the structural flexibility of the decentralized control and provides much improved control performance. Most research efforts on distributed control (especially distributed model predictive control (DMPC)) have been devoted to the algorithm development and the analysis of feasibility, optimality and stability. However, the fundamental problem of subsystem decomposition has received relatively much less attention.

Subsystem decomposition may affect the performance of a distributed predictive control system significantly. The community detection concept originating from network theory provides a very promising way to address the subsystem decomposition problem [4, 6]. By means of the measure of modularity [7], community-based approaches have been proposed to find distributed control structures where the subsystems are made well-decoupled [8, 4]. It is worth mentioning that the existing methods require a full state feedback. However, full state measurements can be difficult to obtain online for many applications. One solution is to incorporate distributed state estimation and distributed control in one integration, such that distributed control can be implemented based on output measurements. From the perspective of implementation, maintenance and communication, it is more favorable if the local state estimators and local controllers are designed based on the same subsystem decomposition. However, a systematic approach to achieve this objective is not yet available.

In this work, we focus on subsystem decomposition of nonlinear process networks for simultaneous distributed state estimation and distributed control. To achieve this goal, we propose a systematic approach based on the concept of community structure detection. We resort to the measure of modularity to quantitatively assess the quality of different community structures. Specifically, the state, manipulated input and measured output variables of a process are taken into account and are viewed as vertices in a network. The way to construct a directed graph containing all the vertices and the corresponding adjacency matrix is defined. An implementation procedure based on approximate optimization of modularity is developed, such that subsystem models for simultaneous distributed state estimation and distributed control can be established by allocating vertices into communities based on modularity. Three chemical process examples of different complexities are used to illustrate the effectiveness and applicability of the proposed approach. While there have been several heuristic approaches on subsystem decomposition for distributed control (e.g., [8, 4]), to the best of our knowledge, the approach proposed in this work is the first one that enables a distributed scheme to require much less information from the process; that is, only output-feedback information is needed for distributed control design and implementation. Also, it is not a trivial extension of the existing approaches. Technical contributions of this work include the following aspects:

• an adjacency matrix which incorporates state, manipulated input, and measured output variables is defined;
• bidirectional edges are considered to characterize the connectivity between state and measured output variables and are used for constructing the direct graph and the adjacency matrix;
• a simple yet efficient method is proposed to calculate the adjacency matrix;
• a method for initializing the community structure is proposed to better handle the constraints on the subsystem structure;
• a systematic procedure is proposed based on the fast folding algorithm [9] to decompose the entire process network into subsystems of which the number is user-specified.

References

[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 and Chemical Engineering, 51:21-41, 2013.

[2] R. Scattolini. Architectures for distributed and hierarchical model predictive control - A review. Journal of Process Control, 19:723-731, 2009.

[3] P. Daoutidis, M. Zachar, and S. S. Jogwar. Sustainability and process control: A survey and perspective. Journal of Process Control, 44:184-206, 2016.

[4] W. Tang and P. Daoutidis. Network decomposition for distributed control through community detection in input-output bipartite graphs. Journal of Process Control, 64:7-14, 2018.

[5] P. Daoutidis, W. Tang, and S. S. Jogwar. Decomposing complex plants for distributed control: Perspectives from network theory. Computers and Chemical Engineering, 2017. in press, doi:10.1016/j.compchemeng.2017.10.015.

[6] D. B. Pourkargar, A. Almansoori, and P. Daoutidis. Impact of decomposition on distributed model predictive control: A process network case study. Industrial and Engineering Chemistry Research, 56(34):9606-9616, 2017.

[7] M. E. J. Newman and M. Girvan. Finding and evaluating community structure in networks. Physical Review E, 69(2):026113, 2004.

[8] S. S. Jogwar and P. Daoutidis. Community-based synthesis of distributed control architectures for integrated process networks. Chemical Engineering Science, 172:434-443, 2017.

[9] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, 2008(10):P10008, 2008.