Coordination Techniques for Distributed Model Predictive Control.

摘要

Industrial chemical plants are complex, highly integrated systems composed of geographically distributed processing units, linked together by material and energy streams. To ensure efficient operation in such integrated plants, multivariable optimal control methods like MPC are required.;This work addresses systematic development of CDMPC networks for plant-wide MPC of interconnected dynamical processes, by modifying the existing decentralized MPC network and designing coordinator. Goal Coordination, Interaction Prediction Coordination and Modified-Pseudo Model Coordination are the three coordination methods studied in this thesis to alter the network of decentralized linear constrained MPCs into CDMPC network. Convergence accuracy studies are provided for the proposed coordination algorithms. CDMPC networks are also developed to study the impacts of uncertainty on the CDMPC and coordinator design using an individual chance-constrained approach. By modifying the CDMPC and coordinator in the Goal Coordination method, it is shown that choosing efficient numerical strategies can improve convergence performance of the coordination algorithm. A novel linear CDMPC network, which has performance of centralized nonlinear MPC, is presented to address the plant-wide nonlinear MPC problem. Numerical simulations are provided to test performance of the proposed CDMPC networks.;Although centralized MPC may provide the best achievable control performance, issues such as lack of flexibility and maintainability make this approach impractical. The general industrial practice to plant-wide MPC is to recognize the distributed structure of the processing units to design a network of decentralized MPCs. Decentralized controllers avoid the disadvantages associated with centralized control at the expense of poorer plant-wide control performance. To improve the performance of decentralized controllers, Distributed MPC (DMPC) methods have become centre of attention in the plant-wide optimal control research community. DMPC methods are divided into two general classes of non-coordinated and coordinated approaches. Coordinated Distributed MPC (CDMPC) networks, which consist of distributed controllers and a coordinator, are able to yield optimal centralized solution under a wide range of conditions.