The Research Of Multi-model Predictive Control Based On Clustering Method For MIMO Nonlinear System

Identification and control of general nonlinear systems is a difficult but important problem. Multi-model modeling method based on clustering algorithm with input-output data of systems reveals the relationship between input data and output data sufficiently, and partitions data into several subspaces corresponding to local linear models reasonably. This method is simple with fast convergence speed and good performance of dynamic characteristic for tracking the system. Recently, model predictive control (MPC) is considered as a novel method to satisfy requires from the practice because of its good control performance for nonlinear systems. So, it is very meaningful to study the multi-model predictive control based on a clustering modeling method and MPC strategy.For modeling of complex nonlinear systems, a multi-model modeling method is presented in this research. It is based on improved K-means clustering to conquer the weakness of classical K-means algorithm, which can not guarantee unique clustering with different initial clusters and assumes the priori knowledge of the number of clusters. With divide-and-conquer strategy, it divides the sample datasets into several subspaces owning distinct character of systems, then for each subspace, Partial least squares (PLS) constructs regression equations with data belong to this subspace. At last, switching scheme is used to select the proper model and corresponding controller for systems. Its applications in general modeling task or nonlinear systems' modeling are both studied and the simulation results express that the proposed algorithm enhances the convergence speed and modeling precision.Through the analysis for multi-model based modeling method, we find that the number of models is important to modeling precision and system stability. In this part, two optimization algorithms to the number of clusters are proposed. The former gives a dynamic threshold value on the basis of DBSCAN and multi-scale theory to merge the similar clustering into one and obtain the optimum number of models and scale value;the latter defines a performance index function to determine the optimum number of models and scale value. For a general clustering task and multi-model identification task, the simulation results testify that two optimization algorithms are validity. Finally, we try to study multivariate PLS regression algorithm for two-output pH processes modeling. This is the outset for further research of modeling strategy.On the basis of multi-model modeling method proposed in above, a multiple model predictive control (MMPC) approach is designed for MIMO systems. A constrained predictive control algorithm is also devised. With regard to switching scheme for multiple models, three method based on different rules are also discussed in this part. The simulation in pH neutralization processes shows that the present method is efficient and switch schemes can improve the transient respond and system stability.