Separable Nonlinear Model Predictive Control via Sequential Quadratic Programming for Large-scale Systems

Abstract

In this paper, a sequential quadratic programming method is presented for largescale nonlinear and possibly non-convex model predictive control (MPC) optimization problem which is often set up with a separable objective function. By introducing the so-call consensus constraints to separate the couplings among the subsystems. The resulting QP subproblem is formulated in a separable form, which makes it possible to use the existing alternating direction methods, like ADMM, to efficiently compute Newton steps for the overall system in a distributed way. In order to enforce the convergence rate of the distributed computation, a distributed line search with local merit functions is also proposed.