An Approach to Auto-Tuning Distributed Model Predictive Control Suboptimality by Using Neural Networks

In this paper, an approach to automatic tuning of suboptimal distributed MPC (DMPC) of linear interconnected systems with coupled dynamics subject to both state and input constraints is proposed. The purpose is to obtain a desired closed-loop performance without exceeding a limit on the online computational complexity. The approach includes three stages which are performed offline. First, the optimal tuning of the MPC cost function parameters is obtained for different values of the suboptimal DMPC design parameters by adjusting the DMPC closed-loop performance. Then, a neural network is used to approximate the influence of the design parameters on the performance and the computational complexity. As a third stage, the best choice of the design parameters is determined by solving an optimization problem based on the obtained neural network model. The suggested approach would be appropriate for embedded distributed MPC since it will reduce the complexity of the online MPC computations and simplify the software implementation.