Robust Model Predictive Control For Constrained Systems And Its Iterative Optimization

In practice,the control system is generally multi-input and multi-output(MIMO),besides its control input and state are constrained,which is a kind of common MIMO constrained control problem in engineering.Recently,model predictive control(MPC)is increasingly used to solve this kind of constrained problem in the applications.However,in the practical applications,the MPC needs to face two key issues,one is the robustness,the other is the computational efficiency.Therefore,this dissertation focuses on how to improve both of the robustness and the computational efficiency of real-time optimization for the MPC's applications.The content of this dissertation is divided into two parts:one is to make a research on the robust and fast MPC technology of the large-scale MIMO nonlinear autonomous underwater vehicle-manipulator system(UVMS),and the other is on the MPC iterative learning optimization technology of the MIMO nonlinear system.The main work of this dissertation is as follows:1.A fast MPC(FMPC)based on the EKF disturbance observer is proposed for the MIMO UVMS in consideration of the anti-disturbance,the control constraints and the online computational efficiency.In order to reduce the complexity of online quadratic programming,the FMPC adopts the future approximated predictive control law instead of computing the future accurate optimal control sequence of the cost function.This scheme can make up for the shortcoming of the H? robust control in terms of online real-time computational efficiency(Riccati equation of H? can only be calculated offline),and strictly guarantee the motion constraint of each degree of freedom(DOF),which can provide the robust optimal control for the UVMS in accordance to the online constraints of the DOFs.Then the Lyapunov stability analysis of the FMPC is presented.Finally,the simulation results show that the proposed control scheme can satisfy the mechanical/physical constraints and the computational efficiency of the control law.2.Furthermore,considering the control problem of redundancy and grasping moving target of the MIMO UVMS,a fast tube MPC(FTMPC)scheme based on the EKF observation of moving target trajectory is proposed.The EKF is an observer of moving target,which can overcome sensory measurement noise.The robust FTMPC consists of the nominal fast MPC and online nonlinear feedback controller,which integrates the kinematic redundancy problem into the fast nominal MPC optimization(no longer decouples the planning and the control as before)and uses the feedback loop to make the controlled system error between the actual state variable and the nominal state variable robustly stable to a fixed domain(Tube).In order to reduce the complexity of online quadratic programming,the fast nominal MPC adopts the kinematic-based approximated predictive control law instead of optimal control sequence for the cost function.The proposed scheme can solve the problem of the kinematic planning and the anti-disturbance grasping moving target of the UVMS.Then the Lyapunov stability analysis of the FTMPC is presented.Finally,the simulation results show that the proposed UVMS control scheme not only can compute fast but also satisfy the constraints and redundancy.3.For a class of MIMO discrete Lipschitz nonlinear system control problem,the adaptive dynamic programming(ADP)theory is used to realize the fast approximated optimal solution of the tube MPC in order to avoid the direct solution of convex optimization of the quadratic cost function.The tube robust feedback loop can provide the anti-disturbance capability for the Lipschitz nonlinear system.At the same time,the implementation flow and stability analysis of the proposed neural networks are given.Finally,the simulations are provided to validate the proposed method.4.For a class of MIMO discrete nonlinear system control problem with unknown model,a critic-only network Q-learning that does not depend on model information is proposed to solve the min-max MPC equivalently,which avoids the quadratic optimization of the min-max cost function.Then the realization process,consistency and stability proof of the proposed neural networks are given.Finally,the comparative simulations show that the proposed scheme is more efficient and robust than the convex optimization of min-max MPC.