Networked Decentralized Model Predictive Control

In process control it is usual to deal with different interconnected plants characterized by significant interactions. For these processes three different control solutions can be developed: a centralized control solution, where all the interactions are considered, a decentralized control solution where each independent control agent is able to acquire estimations or measurements of the interactions, or a hierarchical solution in which a coordinator manages the information exchange between the agents and coordinates them. In general, the optimum policy for each agent does not guarantee the global optimum. Global objectives, such as closed-loop stability or some performance requirements for the global process require coordination among the control agents. This concepts have been widely developed in game theory and originates from the well known prisoner’s dilemma. In this thesis a contribution to decentralized implementation of Model Predictive Control strategy is proposed. The innovative solution is based on independent agents and on a local network used for exchanging a reduced set of information. This Networked Decentralized Model Predictive Control architecture, guarantees satisfactory performance also in the case of strong interactions among subsystems. The performance of the proposed solution is tested on a process characterized by strong interactions and here assumed as a case study. For understanding the reason of using MPC as control strategy, let consider a human being driving a vehicle. It generates steering-wheel commands by forecasting or predicting over a finite time-horizon the (possible) vehicle stateevolutions on the basis of vehicle current state and dynamics, and a virtual or potential steering-wheel command sequence. Then, among such sequences, a particular action which fulfills safety constraints and meets performance requirements is sorted out. Only a short initial portion of such a sequence is applied by the driver to the steering-wheel, while its remaining part is discarded. After such an initial portion is applied, the driver repeats the whole operation by restarting predictions over a moved-ahead or receded time-horizon from the updated vehicle state as determined by the applied command. In a similar manner, Model Predictive Control acts according to the receding horizon principle: at each sampling time, using a predictive model of the system dynamics, the response of the process to changes in manipulated variables over a fixed horizon is predicted. Based on a proper objective function, a finitehorizon optimal control problem is solved to obtain current and future moves of the manipulated variables. Only the first computed move is actually applied to the real system whereas all other control actions in the optimal control sequence are discarded. The same procedure is repeated at the next control step based on the new measurement. Although this computation hinges upon an open-loop control problem, MPC yields a feedback-control law. Therefore, Model Predictive Control is a human inspired method for generating feedback control actions and can be defined an "Intelligent Control Technique". Also the communication based coordination is an intelligent behavior because all the agents can take advantage of the information exchange, whereas the egoistic conduct penalizes the overall system. The general nature of the Model Predictive Control principle allows to deal with complex, multivariable, nonlinear and constrained systems but, at the same time, it arises stability and robustness problems, especially in the decentralized framework. A stability analysis of a model predictive control strategy is proposed in this dissertation for control agents operating in a decentralized control architecture. The current status of the art in Model Predictive Control field is briefly introduced in Chapter 1. The main concepts relying behind the receding horizon principle are presented in Chapter 2 then centralized and decentralized implementations are defined and compared in Chapter 3. Chapter 4 provides a stability analysis either for the centralized or the decentralized case. For validating the obtained results, the proposed analysis is tested on different plants.