Page 13 - Distributed model predictive control for plant-wide systems

Preface

There is a class of complex plant-wide systems which are composed of many physically or
geographically divided subsystems. Each subsystem interacts with some so-called neighbor-
ing subsystems by their states and inputs. The technical target is to achieve a specific global
performance of the entire system.
The classical centralized control solution, which could obtain a good global performance,
is often impractical for application to a plant-wide system for computational reasons and lack
of error tolerance. When the centralized controller fails or a control component fails, the entire
system is out of control and the control integrity cannot be guaranteed.
The distributed (or decentralized) framework, where each subsystem is controlled by an
independent controller, has the advantages of error-tolerance, less computational effort, and
flexibility to system structure. Thus the distributed control framework is usually adopted in this
class of system, in spite of the fact that the dynamic performance of centralized framework is
better. Thus, how to improve the global performance under distributed control framework is a
valuable problem.
Model predictive control (MPC), as a highly practical control technology with high perfor-
mance, has been successfully applied to various linear and nonlinear systems in the process
industries, and is becoming more widespread. The distributed framework of MPC, distributed
MPC (DMPC), is also gradually developed with the development of communication network
technologies in process industries that allow the control technologies and methodologies to
utilize their potentials for improving control.
For the MPC algorithm applied to the plant-wide systems, the system’s architectures can be
divided as follows:

1. Centralized MPC, which is a MIMO system architecture;
2. Decentralized MPC, one controller-one subsystem, but no information exchange between
controllers, and
3. Distributed MPC, which assumes that each subsystem can exchange information with its
neighbor’s subset of other subsystems.

Since the centralized MPC is forbidden for the large-scale plant-wide system with hun-
dreds (or thousands) of inputs and outputs variables due to its lesser flexibility, weak error
tolerance and the large cost of computation, the distributed framework is usually adopted
despite its lower global performance. The schematic of distributed MPC is shown in Figure 1,
the whole system is composed by many spatial distributed interconnected sub-systems. Each