The Method And Application Of Nonlinear Model Predictive Control Based On Support Vector Regression

Model predictive control(MPC)is one of the most mature control algorithms in current advanced control application,which shows excellent control performance in environments such as strong nonlinearity,time-varying and time delay compared with the traditional methods.In nonlinear model predictive control(NMPC)system,the actual nonlinear process cannot be described accurately by the traditional linear modeling method.At the same time,there exists some problems such as difficulty in solving the nonlinear controller and the complicated changes in actual industrial state.Therefore,obtaining a more accurate predictive model,improving the solution efficiency of the control algorithm,and strengthening the adaptive ability of the model are the key issues to be solved urgently in the NMPC system.Support vector regression(SVR)is intended to establish the predictive model for the above problems in the NMPC.Firstly,in order to improve the modeling accuracy of the SVR,a method selecting model order that only relies on the historical input and output data of the system is proposed to determine the input vector dimension of the SVR in advance.Secondly,one-step and multi-step prediction models of the SVR are derived and the Levenberg-Marquardt(LM)algorithm is used to solve the corresponding control variables.Finally,the offline model is combined with online learning,so that the prediction model has an effective update strategy,which can adapt better to the changing working conditions and interference in the real production environment.The highlights of the research work are as follows:(1)Considering the correlation between the order of the system and the model dimension of the input variables during modeling,the false nearest neighbor algorithm(FNN)and Gaussian mixture model(GMM)clustering are used to select the model order.According to the clustering results,the gradient of the most principal component of each data cluster is solved to approximate the Jacobian matrix in the FNN.Therefore,the limitation of the FNN algorithm is overcome,and the model order is obtained eventually.(2)It is difficult to build accurate models in the NMPC,so the SVR is used to build the nonlinear model where one-step prediction and multi-step prediction models are designed.Considering complicated solving process of control variables,the LM algorithm is adopted to derive the optimal solution of the quadratic objective function.Meanwhile,three hyperparameters of the SVR need to be set in advance,so that the bat algorithm is developed to further optimize hyperparameters in the modeling.(3)Considering the fact that the offline model can not have high adaptability in the online production process,an update strategy with online SVR that using incremental deduplication-reduction pruning based on sample distance is proposed to make the model obtain better anti-interference and adaptive adjustment.Finally,two experiments are carried out in the strong nonlinear simulation function and the isothermal polymerization reaction process to verify the superior performance of the proposed methods.