Page 37 - Distributed model predictive control for plant-wide systems

Introduction 11

tolerance, and no global information requirements [27, 28]. The advantages of the distributed
predictive control are as follows:

• Its underlying ideas are easy to understand: the distributed predictive control is the dis-
tributed implementation of a set of predictive controllers, and these predictive controllers
consider the feedforward information from the predictive controllers which corresponds to
the subsystems they interacted with.
• The local predictive control can deal routinely with equipment and safety constraints.
• The local predictive control handles multivariable control problems naturally. It is more
powerful than PID control, even for single loops without constraints. It is not much more
difficult to tune, even on “difficult” loops such as those containing long time delay.
• It allows operation closer to constraints compared with conventional control, which fre-
quently leads to more profitable operation.
• Since the centralized predictive control is decomposed into many small-scaled predictive
controllers, the computational efforts in each small-scaled predictive control are much less
than that used for solving the centralized predictive control.
• If one or several errors occur in a subsystem, the other subsystem-based predictive con-
trollers are still able to work. There is a good error-tolerance characteristic.
• If some new subsystems are appended into the current system, it is not necessary to modify
all the local predictive controls. We should only modify the predictive control whose cor-
responding subsystem interacts with the new added subsystems. The distributed predictive
control owns high flexibility to the system structure.
• The “plug-in and plug-out” is also able to be realized if a suitable algorithm and an appro-
priate program are designed.

Due to these advantages, the distributed predictive control gradually takes the place of a
centralized predictive control for plant-wide systems. However, as pointed out in [27–33], the
optimization performance of distributed predictive control, in most cases, is not as good as that
of centralized predictive control. Thus, many different coordinating strategies are proposed to
solve this problem [27, 29, 31–44]. In most cases, the coordinating strategies are very important
to the performance of the closed-loop systems.

1.4.4 Classification of DMPC
To improve the global performance of the DMPC, several coordination strategies have
appeared in the literature, and can be classified according to the information exchange
protocol needed, and to the type of cost function which is optimized [6]. There are two classes
of distributed predictive control if we catalog it by the information exchange protocol.

• Noniterative-based algorithm: in this kind of distributed predictive control, each local
predictive control communicates only once with other local predictive control within
every single control period, and solves the local control law once in a control period,
e.g., [34, 44–47].
• Iterative-based algorithm: this kind of distributed predictive control assumes that the net-
work communication resources are abundant enough for supporting the fact that each local