LS-SVR Based Nonlinear Predictive Control Method And Application

Model predictive control,which can deal with the problem of process control with constraints,multivariate and uncertainty,has been widely used in industrial process.With the complexity and large-scale of modern industrial engineering,the traditional linear predictive control is difficult to achieve satisfactory control objectives.As a result,the nonlinear model predictive control algorithm has become the focus of research.However,the nonlinear predictive control model cannot reflect the dynamic characteristics of the whole process and online nonlinear optimization is very complex,which cannot make the nonlinear model predictive control achieve the expected achievement in practical application.Therefore,the need of improving the accuracy of the nonlinear prediction model,solving effectively the model mismatch and improving the optimization efficiency of predictive control are particularly important in the whole process of nonlinear model predictive control.This thesis relies on the National Natural Science Foundation of China,"a large-scale blast furnace high-performance operation control method and implementation technology"(Grant number:61290323)and "Experimental verification platform construction and application verification of large blast furnace high-performance operation control"(Grant number:61290321).In order to solve the above problems,a multi-output least squares support vector regression(M-LS-SVR)algorithm is proposed for multi-input multi-output(MIMO)nonlinear systems.A nonlinear predictive controller based on sequential quadratic programming(SQP)is designed for this model to achieve the control of the MIMO nonlinear systems.The main contributions are given as follows:(1)M-LS-SVR based nonlinear predictive control of MIMO system.Considering the existence of multiple output in most of the actual industrial processes and the conventional LS-SVR can only be modeled for a single output system,a nonlinear predictive control method based on the M-LS-SVR model is proposed.Firstly,the objective function of LS-SVR is changed to M-LS-SVR.Since the regularization parameter and kernel parameter of M-LS-SVR have a great influence on the learning ability and generalization ability of the model,particle swarm optimization(PSO)is used to select the optimal parameter set of M-LS-SVR.In order to validate the modeling results,a comparison experiment with the modeling method based on random weight neural network(RVFLN)is conducted.The results show that M-LS-SVR has better effect in model prediction.Secondly,according to the final established M-LS-SVR prediction model,a nonlinear predictive controller based on SQP algorithm is designed.Finally,a standard test is implemented.The results show that the nonlinear predictive control method based on M-LS-SVR model is feasible and has better control effect than the nonlinear predictive control method based on RVFLN model.(2)M-LS-SVR based nonlinear adaptive predictive control of MIMO system.Aiming at the problem of model mismatch between the plant and the M-LS-SVR prediction model when the internal parameters of the system are changed,an adaptive nonlinear predictive control strategy is proposed.There are two nonlinear optimization layers in this method.The first layer is a model parameter estimator,which adjusts the M-LS-SVR prediction model parameters through real-time optimization of the prediction model and the output error of the actual system.The second layer is a nonlinear predictive controller which can calculate the nonlinear predictive control law in real time by using the SQP algorithm.Finally,a standard test is conducted.The results show that the nonlinear adaptive predictive control method based on M-LS-SVR model is better than the M-LS-SVR nonlinear predictive control method when the parameters of the controlled plant are changed.(3)Aiming at the two nonlinear predictive control methods proposed in this thesis,they are applied to the control of multiple molten iron quality indicators(silicon content,molten iron temperature)in the actual blast furnace iron-making process.The M-LS-SVR prediction model is established by preprocessing the production data of the iron-making process,and then the two methods mentioned above are used to design the nonlinear predictive controller.They can make the silicon content and molten iron temperature reach the expected values.Furthermore,when the prediction model is mismatched with the blast,the nonlinear adaptive predictive control method still keeps a satisfactory control effect,which verifies the feasibility and validity of the proposed method.