A Study On Generalized Predictive Control Algorithms And Their Simulations

Generalized Predictive Control (GPC) has emerged as a powerful control strategy that has been widely applied to complex industrial processes. Since 1987 when it was presented, GPC has been fully developed in theory and proved to have a good control performance. As is logical, however, it also has its disadvantages in the practical applications:1.Huge computation, including the on-line estimation of the process parameters, the computation of the Diophantine equation and, especially, the computation of the inverse matrix as well as other computation, like one used to handle the constrains.2.The tuning parameters are not easy to determine, include the prediction horizon, the control horizon, the control weighting parameter and the output softness parameter.3.GPC still needs a relatively precise model of the process.Therefore, GPC is studied here in the above aspects, and the thesis is organized as follows:1.The first chapter deals with the background of presentment and the basic principles of Model Predictive Control algorithms, including Model Algorithmic Control (MAC), Dynamic Matrix Control (DMC) and Generalized Predictive Control (GPC). Following this, GPC algorithm is introduced in detail.2.In the second chapter, to avoid huge computation of the traditional GPC, a fast algorithm of GPC is presented by exactly computing the current controlaction and approximately computing the future control sequence off line. The algorithm is simple and it needn't solving Diophantine equation and the inverse matrix so that the on-line amounts of computation are reduced strongly. The simulation results show that the algorithm has perfect control quality and has strong constraints on the control actions. And then it is extended to Multiple Input Multiple Output (MIMO) linear systems.3.In the third chapter, a GPC algorithm with input constraints is presented in which the concept of output softness was used to soften the inputs. As a result, the constraints are simplified to be the only one constraint on the current control increment witch can be computed directly, no matter what the prediction horizon is. At the same time, it needn't computing the inverse matrix and thus reduces large computation. Moreover, it guarantees the feasibility of the algorithm and has good control performance. At last, it is extended to MIMO linear systems.4.ln the fourth chapter, to overcome the difficulty in the choice of tuning parameters in GPC, a GPC algorithm with variable parameter design is presented based on BP Neural Network trained by modified BP algorithm, in which the tuning parameters are tuned on line, including the output softness parameter and the control increment weight (or input softness parameter). Simulation results show that the algorithm is superior to the GPC with invariable parameters design in track capacity and control precision, and it also puts a check on the disturbance in some sense.5.ln the fifth chapter, to overcome the influence of the modeling errors on predictive control performance, a GPC algorithm based on predictive error correction by Neural Network is presented, in which the linear model of the controlled system is identified by RLS, and the model of predictive errors is by a Feedforward Neural Network trained by Powell Method. The network needn't training off line in advance, and can be directly put to use in the close-loop control. At the same time, it can guarantee that the algorithm has good stability, track ability and robustness. The simulation results show its effectiveness.Finally, the thesis concludes with a summary and perspectives of future...