A distributed MPC approach — Local agents decide network convergence

Abstract

In an earlier work, the Limited Communication-Distributed Model Predictive Control (LC-DMPC) scheme for controlling networks with dynamically coupled and locally constrained linear systems is presented. The scheme has an iterative and cooperative structure in which the systemwide optimum point is achieved by the distributed controllers requiring only coupled agents to cooperate. For assessing the network convergence, it is essential to possess complete information pertaining to all subsystems which has to be available to a central monitor. The current work endeavors to investigate this challenging point by distributing the network convergence within the local agents. With the new version of the algorithm, the convergence of the network is now guaranteed through the dissipativity of the local information exchange dynamics in the iteration domain. This is accomplished by introducing a set of free design variables into the distributed problems which are utilized by the agents to fulfill a simple local LMI problem employing local information only. Despite that the new approach is eliminating the necessity for a centralized observer, it may result in suboptimal local solutions. This is because the convergence of the information sharing loop between the coupled subsystems is insured by the small gain theorem. The new algorithm exhibits enhanced modularity due to the novel introduced convergence condition, implying that any updates to a subsystem physicals or design parameters do not require corresponding updates to neighboring subsystems or the network. The presented concepts are demonstrated by simulating a network of eight interconnected tanks.