Generalized Predictive Control Of Multi-Variable Non-Linear System Based On Neural Networks

Along with the development of the subjects such as voyaging control, process control, long-range robot control as well as biological medical project, have arisen some control problems of strong nonlinear and strong couple system correspondingly. The generalized predictive controlled (GPC) can consider various conditions in optimization to restrain, and have very good robustness and interference rejection as other predictive controlling method. So compared with conventional control, the GPC can realize development control better. On the other hand, GPC adopts controlled auto-regressive integrated moving average (CARIMA) model and its control law including integral action. So the GPC can automatically eliminate the deviation of stable state of controlsystem. Therefore in single variable weak nonlinear system it has gotten widely application. When applying this kind of successful control thought in strong nonlinear and strong couple system we need to solve firstly the difficulty of decoupling and building model. And with its mapping and approaching ability for general nonlinear function, the neural network offers a way in which we can decouple multi-variable nonlinear system and build mathematics model suiting to GPC of single-variable nonlinear systematic.In order to the advanced thought of GPC apply successfully in multi-nonlinear system, this paper use error back-propagation (BP) neural network to build system model. Use BP neural network that has one hidden layer to approach model of the nonlinear part of nonlinear system and the Davidon least-square that studies speed comparatively quickly to online adjust network weight values. Linear part adopts CARJMA model and RLS as its mathematics model and parameter updating way. Every circle expand network model in linear way and get linear regression model, and use nonlinear feed-forward gain way to compensation building-model error. Have established a kind of generalized predictive controllerthat suits single-nonlinear system. And simulation result with MATLAB language has proved this design validity.To be able to popularize this design to multi-variables nonlinear system, firstly disconnect controllers, use the one hidden layer BP neural network to decouple multi-variable nonlinear system with open loop, through studying the might controllers of generalized plant, train neural network and eliminate couple influence. Then connect the GPC and use neural network decoupling to eliminate influence be think as disturb coming from other loop. Last use the GPC of the single-variable nonlinear system to control every single-variable nonlinear system.Apply this method in strong nonlinear, strong coupling, big time delay, big inertia and change-time model such as coal-pulveriz -ing systems with ball mill. To simplify double-in double-out discrete mathematics models, firstly neural network coupling device decouples this model and use MATLAB language to study and analyze it. Then apply GPC of single variable nonlinear systems in two approximate without coupling system and use MATLAB language to simulation at the same time to study and analyze. Thesimulation results prove the GPC based on neural network multi-variable non-linear system that this paper presents Validity in the using of in coal-pulverizing systems with ball mill.