MPC (Model Predictive Control) is an important control strategy based on control theory and computer technology. The classic MPC use linear model as their predictive model, and are therefore not generally suitable for the industry processes which are subjected to strong nonlinearities. It is very important to study new ways so as to combine nonlinear system model theory, nonlinear optimization theory and model predictive control theory and to design nonlinear model predictive control algorithms for industrial processes. This PhD thesis presents novel studies on nonlinear model predictive control method based on artificial neural network, where the research objective is to apply artificial neural network theory and nonlinear optimization theory to the design of model predictive control so as to establish novel techniques to deal with nonlinear control problems. The main study and contributions are as follows: (1) The basic principles and typical algorithms of model predictive control theory, system identification theory, artificial neural network theory and nonlinear optimization theory are systematically reviewed. These are the foundation of the follow-up chapters of the thesis. (2) The nonlinear one-step predictive model based on a BP neural network was derived according to the mathematics expression of NARMAX model. Then the one-step predictive control rule without constraints was firstly designed by using Newton iterative algorithm. This is followed by the development of a one-step predictive control rule with constraints using quadratic programming algorithm. The capability of one–step predictive model based on BP neural network and the performance of the one-step predictive control rules were validated through some computer simulations. (3) The nonlinear multi-step predictive model based on a two-layer BP neural network was derived through iterative calling the two-layer BP neural network. Then the multi-step predictive control rule without constraints was designed by using Newton–Rhapson algorithm. The multi-step predictive control rules with constraints were then designed by using sequential quadratic programming algorithm and genetic algorithm, respectively. The capability of multi–step predictive model based on two-layer BP neural network and the performance of the multi-step predictive control rules were validated through the computer simulations.
修改评论