Page 68 - Distributed model predictive control for plant-wide systems

42 Distributed Model Predictive Control for Plant-Wide Systems

predicting the evolution of a subsystem over a horizon involves even more uncertainty
than when a centralized MPC is employed. This is one of the reasons why typically the
decentralized/DMPC control structure has a lower performance than that of the centralized
MPC. The challenge in implementing such a decentralized/DMPC strategy comes from
ensuring that the actions that the individual MPCs choose result in the performance that is as
good as when a hypothetical centralized MPC control structure in which all information is
available would be used.
Compared with decentralized MPC and DMPC, the advantage of the decentralized over
the single-layer DMPC control structures is that there is no communication between the
MPCs, resulting in lower computational requirements and faster control. However, this
advantage will typically be at the price of decreased overall performance. The advantage of
a DMPC control structure is therefore that improved performance can be obtained, although
at the price of increased computation time due to cooperation, communication, and perhaps
negotiation among subsystem-based MPCs. Consequently, with the development of network
communication, fieldbus, and computer technologies, the study on how to design DMPC
coordination strategies for improving the performance of the entire system has become a very
popular topic.

3.4 Hierarchical Distributed MPC
When there are multiple MPC controllers, and some of the MPC controllers can force or guide
other MPC controllers, we call the control structure a multilayer DMPC structure, as illustrated
in Figure 3.3. A typical multilayer MPC control structure is the one in which MPC deter-
mines a set-point to a group of other MPCs that work in a decentralized or distributed way.
Due to the authority relationship between controllers or groups of controllers, the multilayer
MPC control structure is also referred to as a supervisory control structure, or a hierarchical
decentralized/DMPC structure.
The hierarchical DMPC structure provides the possibility of obtaining a tradeoff between
system performance and computational complexity. A higher layer controller considers a large
part of the system and can therefore direct the lower control layer to obtain coordination. Such
a hierarchical control structure can thus combine the advantages of the centralized MPC struc-
ture with the DMPC control structure. The approaches that the higher layer MPC coordinates
with the low-layer controllers could be enforcing penalty terms, providing additional con-
straints, or providing set-points. The advantage of the higher layer controller is in particular
clear when the MPCs of the lower layer are decentralized, i.e., not communicating with one
another. Consequently, the hierarchical decentralized MPC has been widely used in chemical
processes in recent years, regardless of the fact that the communication in a hierarchical decen-
tralized/DMPC control structure is more complex than that in a centralized MPC structure and
a decentralized/DMPC control structure.
An important issue that should be addressed when designing MPC for hierarchical DMPC
control structures is that, in the hierarchical decentralized/DMPC, the supervisory MPC can
also be regulated by an MPC in a higher layer of the control structure. The higher layer MPC
can coordinate with the lower layer MPCs, which may consist of controllers using hierarchical
decentralized/DMPC and single-layer decentralized/DMPC. Another issue is the choice of the
prediction model that the higher layer control agent uses. A controller in the higher layer has
to be able to make relevant predictions of the physical system. Since the physical system is