The Research Of Adaptive Multilayer Nonlinear Model Predictive Control Algorithm

With the growing scale and complexity of production, it becomes more and moredifficult to establish mechanism model of the system, then it has become a trend toestablish identification model using production data. In recent years, support vectormachine (SVM) has been successfully applied in the fields of pattern recognition, systemidentification. Support vector machine is injecting a flow of fresh vitality into establishingthe mathematical models of complex systems. With the development of production level,higher requirements of the process controller are demanded, however, the traditionalmultilayer control structure is for linear systems, the adaptability to external disturbanceand model mismatching is not strong and the weights adjustment are difficult, so theresearch on a new kind of multilayer model predictive control algorithm for nonlinearsystem is of great significance. For a class of complicate objects with nonlinearcharacteristics, this paper researches an intelligent modeling method based on supportvector machine and the corresponding multilayer nonlinear model predictive controlstrategy. Specific tasks are as follows:First of all, for a class of complex systems with nonlinear characteristics, this paperresearches a model identification algorithm-least square support vector machine based onparticle swarm optimization (PSO-LSSVM). PSO-LSSVM adopts PSO to optimizeregularization coefficient C and kernel function width, which affect thegeneralization ability of LSSVM, avoiding the blindness of parameters selection.PSO-LSSVM enhances the generalization ability of the model, and provides modelfoundation for the research of adaptive multilayer nonlinear model predictive controlalgorithm.Secondly, for multilayer optimization problems of nonlinear system, adaptivemultilayer nonlinear model predictive control (AMNMPC) algorithm is put forward,which divides the multilayer optimization of nonlinear system into three parts: localsteady-state optimization, adaptive steady-state optimization and dynamic optimization.Local steady-state optimization layer calculates expectant inputs and outputs of the system;adaptive steady-state optimization layer recalculates the optimal working point in eachcontrol period to enhance the system’s adaptability to external disturbance and modelmismatch, and coordinates multiple control objectives using goal programming method toreduce the weight adjustment difficulty; dynamic optimization layer adopts zone control method to provide enough degrees of freedom for system to track the optimal workingpoint.Finally, selecting the fourth generation of grate cooler as the study object, byanalyzing the process flows and control mechanism, this paper chooses the number of barmove back and forth and the speed of second wind room cooling fan as control variables,and chooses second wind temperature and the grate under pressure of second wind roomas controlled variables. The historical data of cement grate cooler system is obtainedthrough simulation platform of cement production process control system. For cementgrate cooler system, this paper uses PSO-LSSVM model identification algorithm toidentify the model and uses adaptive multilayer nonlinear model predictive controlalgorithm to make the predictive control simulation. The simulation results verify thecorrectness and effectiveness of the proposed algorithm.