(137c) Quasi-Decentralized Output Feedback Model Predictive Control of Networked Process Systems Using Forecast-Triggered Communication

El-Farra, N. H., University of California, Davis
Hu, Y., University of California, Davis

Chemical plants are typically large-scale processes involving many units that have complex dynamical behavior; tight interconnections usually exist among the various units and thus the dynamics of each unit are strongly coupled with the dynamics of the rest of the plant through the exchange of material and energy. The traditional solution for control of plants with interconnected and distributed units usually falls within either the centralized or the decentralized control frameworks. The centralized framework offers satisfactory performance for small- to medium-scale plants as the single control agent that controls the entire plant is able to account for the interconnections of all the units; however, for large-scale plants, the single agent has to maintain control for a large number of subsystems which may have different objectives, and this poses significant problems to the computation and coordination of the plant-wide control. While for decentralized control where each subsystem is controlled independently by an agent, computational burden and coordination of all the subsystems are no longer a problem, but it may lead to poor and unreliable performance when the various subsystems have strong interactions with each other.

The issues associated with the traditional solutions for control of large-scale plants have motivated the development of the distributed control framework (e.g., see [1]–[4] for some results and references in this area) which, on the one hand, avoids the complexity and inflexibility that are typical of centralized approaches by adopting a decentralized topology, and on the other hand, addresses the stability and performance issues associated with decentralized approaches by handling the inter-subsystem interactions in a more systematic way.

The increase in size and complexity of industrial processes is also accompanied by the increasing use of communication networks (e.g., see [5], [6]). Control over networks, however, introduces new issues that challenge the assumptions in traditional process control theory due to the inherent limitations on the capabilities of the networked devices and communication medium. These issues, for example, resource constraints, data losses and communication delays, may limit the overall achievable control quality if they are not explicitly handled in the control system design. In addition, for many practical applications, not all process states can be measured, and this limits the applicability of distributed control methods developed under state feedback [7] because the control systems designed using those methods require full-state measurements to monitor and regulate the process.

Motivated by these considerations, we propose in this work a quasi-decentralized Lyapunov-based MPC framework for networked process systems under output feedback with a forecast-triggered communication strategy. Specifically, each subsystem of the networked process system is controlled by a local MPC controller, and these quasi-decentralized MPC controllers are designed based on distributed Lyapunov-based output feedback control. To handle the lack of full-state measurements, a supervisory observer that is able to access the process input and output information generates estimates of the process state. We show that practical stability can be guaranteed if the state of the model that is embedded within each controller is updated by the observer estimate at every sampling instant. In order to minimize the information transfer between the supervisory observer and the local control systems while still maintaining practical stability, an adaptive forecast-triggered communication strategy is proposed. The key idea of this strategy is to forecast the worst-case scenario of the future evolution of the process by combining our knowledge about the stability properties of each subsystem with the information from the process itself that reflects its current operating status. If the forecast indicates possible instability in the future, then the communication is established to curb the growth of the local model estimation error; otherwise, the communication is suspended for as long as closed-loop stability is not at risk. The implementation of the developed methodology is demonstrated using a simulated model of a chemical process network.


[1] E. Camponogara, D. Jia, B. Krogh, and S. Talukdar, “Distributed model predictive control,” IEEE Control Syst. Mag., vol. 22, no. 1, pp. 44–52, 2002.

[2] M. R. Katebi and M. A. Johnson, “Predictive control design for large-scale systems,” Automatica, vol. 33, no. 3, pp. 421–425, 1997.

[3] Y. Sun and N. H. El-Farra, “A quasi-decentralized approach for networked state estimation and control of process systems,” Ind. Eng. Chem. Res., vol. 49, no. 17, pp. 7957–7971, 2010.

[4] P. D. Christofides, J. Liu, and D. Mu˜noz de la Pe˜na, Networked and Distributed Predictive Control: Methods and Nonlinear Process Network Applications. London: Springer-Verlag, 2011.

[5] B. E. Ydstie, “New vistas for process control: Integrating physics and communication networks,” AIChE J., vol. 48, no. 3, pp. 422–426, 2002.

[6] P. D. Christofides, J. F. Davis, N. H. El-Farra, D. Clark, K. Harris, and J. N. Gibson, “Smart plant operations: Vision, progress and challenges,” AIChE J., vol. 53, no. 11, pp. 2734–2741, 2007.

[7] Y. Hu and N. H. El-Farra, "Adaptive Quasi-Decentralized Model Predictive Control of Networked Process Systems," Distributed MPC Made Easy, R. Negenborn and P. Maestre (Eds.), in press, Springer.