Analysis And Design Of Simplified Dual Neural Network Based Model Predictive Controller

In the past30decades, model predictive control (MPC) technology hasbeen widely applied to a variety of industrial processes. A crucial advantageof MPC technology over other control strategies lies in its capability ofhandling constraints straightforwardly by solving an optimization problemthat takes all system constraints into account. Generally, the optimizationproblem in MPC technology is a constrained quadratic programming (QP)problem whose solution gives the control law. Because the MPC controllerhas to solve the mentioned constrained QP problem on-line at every samplinginstant, the resulting computational burden limits its application in localcontrol systems. Hence, in recent years, implementing the MPC technologyon embedded hardware devices and then applying to local control systemshave attracted broad attentions. Under such a background, this dissertationinvestigates the theoretical issues of discrete-time simplified dual neuralnetwork (SDNN) for solving convex QP problems and on this basis,implements the neural network MPC controller on digital signal processor(DSP) platform. The main work of this dissertation can be summarized as thefollowing four aspects:1) Investigation of analog circuit implementation of the continuous-timesimplified dual neural network (SDNN). The analog circuit for eachfunctional module of the continuous-time SDNN is designed. The analogimplementation design is simulated and studied on Multisim. The drawbacks and difficulties of analog implementation are analyzed and discussed.2) Convergence analysis of the discrete-time simplified dual neuralnetwork (SDNN). From the perspective of implementation, this dissertationmakes time discretization of the continuous-time SDNN and enables it to beimplemented on digital circuits. The convergence property of thediscrete-time SDNN is first studied. By choosing a proper Lyapunov function,a sufficient condition for global exponential convergence is obtained. Basedon this condition, we further obtain an enhanced convergence condition byintroducing a new scaling parameter to the network. The enhanced conditionenlarges the convergence region while remains the global and exponentialproperty at the same time.3) DSP implementation of the discrete-time SDNN model predictivecontroller. An integrated design of software and hardware is developed toimplement the discrete-time SDNN on DSP and a prototype system of theneural network MPC controller is built on a TMDSEVM6678L DSPdevelopment board.4) Air separation unit (ASU) application. The proposed DSP-basedneural network MPC controller is applied to an ASU system and achievessatisfactory control performance. This illustrates the feasibility andeffectiveness of the proposed design.