Research On Model Predictive Control For Data Driven Based Distributed Parameter Systems

Distributed parameter systems,the systems of spatiotemporal characteristics,are widely existed in the process industry.The spatiotemporal systems,also called spatial distributed systems,have infinite dimensional properties and strong nonlinear dynamics.Because of these attributes,modeling and control research are challenging for distributed parameter systems.In practice,uncertainty and lack of effective information are widespread in process industry,which make it difficult to obtain partial differential equations description precisely.Therefore,the mathematical description obtained by theoretical modeling method is not applicable to controller design directly.Without prior knowledge,the data sets based on input and output are fully learned.Meanwhile,the distributed parameter systems described by unknown mathematical description is approximated by finite dimensional model.Model predictive control is widely used to solve the control problem of process industry.Especially,due to the effectiveness of handling constraints,several predictive controllers have been successfully designed for large-scale and complex industrial systems.Based on a class of parabolic distributed parameter systems,the time coefficients obtained by time-space decomposition are modeled.After that,the predictive controller is designed for temporal estimated model.The main contents of the dissertation are as follows:First of all,the principle of time-space decomposition is mainly introduced.The training process of spatial basis function based on data driven is deduced in detail,and the solving steps of K-L decomposition-based basis function are given.In addition,the least squares support vector machine is selected to learn temporal series for weakly nonlinear parabolic spatiotemporal systems.For the multivariable autoregressive exogenous model approached by least squares support vector machine,a generalized predictive controller is designed.Thirdly,considering the standard nonlinear distributed parameter system,an extreme learning machine is chosen to identify the temporal coefficients of model reduction.The temporal dimensional nonlinear model regressed by extreme learning machine is used to design the predictive controller.The path integral optimal control algorithm is imported to solve the nonlinear predictive control problem,and the reference trajectory is approximated by random sampling repeatedly.Finally,the proposed control algorithm is utilized for two typical cases of chemical industry,respectively.A spatiotemporal autoregressive exogenous model based generalized predictive algorithm is designed for the catalytic rod described by a weakly nonlinear distributed parameter system.Concurrently,the selection of kernel functions and constrained control problem are discussed.For the continuous convection-diffusion-reaction process composed of tubular reactor,the path integral optimal control strategy based on nonlinear predictive model is presented to solve the concentration control problem of reactor.Simultaneously,the control effect for the number of sampling trajectories and disturbance hyper parameter is analyzed.