Research Of Nonlinear Model Predictive Control Algorithm Based On Neural Network

Model Predictive Control (MPC) is a promising new predictive control algorithm with the development of industrial practice in1970s. MPC algorithm has the characteristics of low requirements of model, strong robustness and good control performance. MPC also can deal the problems effectively with varying delays, multi-variable and constraint.Now MPC has been widely applied in the fields of petroleum, chemical, metallurgy, machinery, robotics, biomedicine in which significant economic benefits has been achieved.The actual industrial production systems have characteristics of strongly coupled, nonlinear, time delay and varying due to the complexity of the industrial process. Facing with complex industrial subjects and higher control performance requirements of plants, Generalized Predictive Control only adapted to the linear systems is difficult to achieve rapid and effective control in nonlinear systems. Therefore, the researches on Model Predictive Control of nonlinear systems give rise to more concern about the issue.Applying Artificial Intelligence Technology on control system becomes the trend of development of the subject of modern automatic control. As a cutting-edge technology of multidisciplinary field, Artificial Neural Network (ANN) can be used to approach any complex nonlinear systems and be able to learn and adapt to the uncertainty systems’dynamic characteristics. Taking into account these characteristics, ANN is can be applied in solving the problem of modeling and controlling of nonlinear and uncertain systems. So, Nonlinear Model Predictive Control based on neural network gradually which combines the advantages of MPC and ANN gradually becomes an important method of solving complex nonlinear system control problems in industrial process.Based on the above theories and ideas, this paper firstly introduces the development and main methods of Nonlinear Model Predictive Control (NMPC)as well as the research situation of NMPC. Then the emergence reason and development of Model Predictive Control is described, especially emphasis on Generalized Predictive Control algorithm and parameter selection principle. The paper also points out that MPC proposed for linear system and it is difficult to establish accurate models for nonlinear system for multi-step prediction and obtain satisfactory control effect Therefore, the neural network is applied in the nonlinear system identification in the paper. Neural network chooses Elman neural network which is a local dynamic feedback network with good approximation performance. On the basis of theories, the new method of applying modified Elman as a predictive model of multi-step prediction in the GPC is proposed. First of all the neural network is used as a predictive model to output multi-step prediction value, then optimum control is obtained by optimizing the objective function using optimization algorithms. Then Particle Swarm Optimization is adopted to avoid complexity of the recursive algorithm in the GPC controller. Finally, the results by the simulation of nonlinear system show that the new algorithm has a better control effect.