Distributed model predictive control of large-scale networked systems
The control of large complex networked systems may be accomplished by applying local modeling and control techniques to its smaller, more manageable constituent subsystems. Because there is little cooperation between the local controllers, they can interact in unexpected ways, not considered in the design phase. As a result, the full system may display fragility and even instability in the face of unmodeled disturbances. An excellent illustration of this phenomenon was the failure of the North American power system in August 2003.
The PIs plan to develop new methods for the control and optimization of large, complex, networked systems. They will design these methods to be more robust than existing methods in the presence of large disturbances and component or subsystem failure, with performance approaching that of a fully centralized methodology. At the same time, these methods will be implementable in practical settings, by taking advantage of the currently deployed subsystem models and controllers, and avoiding the onerous modeling requirements and organizational/institutional obstacles associated with a centralized control methodology. The key to the approaches will be communication of information between subsystems, and cooperation between their controllers. The forecasts produced by model predictive control techniques provide rich information about future behavior of each subsystem. In some situations, sharing of this information between subsystems is itself sufficient to achieve almost all of the potential benefits of centralized control. In more tightly coupled systems, more extensive cooperation is required. One form of cooperation is to give the subsystem controllers a common objective-a straightforward modification if the subsystems already use model predictive control. If the subsystems currently use PID or some other low-level control scheme, it is not difficult to first replace these controllers with model predictive controllers before applying the techniques developed in this research. Cooperation can also involve sharing of information between subsystems, possibly more than once within a single sample period, with local reoptimizations performed after each exchange of information. A crucial component of the project will be the design of robust and rapidly converging optimization schemes that perform most of their computations locally within each subsystem and exchange limited information between subsystems.
To ensure industrial relevance and impact, the PIs have established collaborations with six industrial partners to demonstrate these new approaches on two economically significant application classes: electric power networks and large-scale, integrated chemical plants exchanging raw materials and products. They plan to collaborate with the industrial partners to test and refine the proposed methods, to demonstrate the benefits with actual industrial operating data, and to transfer the technology into practice. Testing and implementation of the proposed methods will provide a vital educational experience for the graduate students supported by the project.
The broader impact lies in the opportunity to demonstrate methods that increase the reliability of critical infrastructures that are composed of many highly interacting subsystems. Infrastructure of this type is already becoming pervasive in highly technological societies, and the need for improving the reliability of this infrastructure is increasingly urgent.
