Plug and Play Distributed Model Predictive Control with Dynamic Coupling: A Randomized Primal-Dual Proximal Algorithm

This paper proposes an algorithm for distributed model predictive control that is based on a primal-dual proximal algorithm developed recently by two of the authors. The proposed scheme does not require strong convexity, involves one round of communication at every iteration and is fully distributed. In fact, both the iterations and the stepsizes are computed using only local information. This allows a plug and play implementation where addition or removal of a subsystem only affects the neighboring nodes without the need for global coordination. The proposed scheme enjoys a linear convergence rate. In addition, we provide a randomized variant of the algorithm in which at every iteration subsystems wake up randomly independent of one another. Numerical simulations are performed for the frequency control problem in a power network, demonstrating the attractive performance of the new scheme.