Chemical Process Benchmark Intelligent Control Technology Research And Applications

With the development of science, controlled objects become more complicated, but high precision is given, so traditional theory of controlling is not well to solve this problem. On account of it, the theory of intelligent controlling based on logical dianoetic, expert system and neural network etc. is developed rapidly. Neural network control is one of important parts of intelligent control technique. In this dissertation, neural network characteristics, learning algorithm and learning parameters analyzed.There are not only inner feedback connections but only forwardback connections in the Recurrent Neural Network(RNN), which can token dynamic characteristics of objects very well. After analysed latest topology of recurrent neural network and representative neural network model, new recurrent neural network and its learning algorithm are proposed. The standard backpropagation learning algorithm, the size of samples, initial weights, learning rate and monentum are discussed, which influence the stability and convergence of neural network. The dynamic adaptive adjustment strategy based on the training error for learning rate and monentum is presented.In the dissertation, two industrial reaction proccesses are simulated successfully with the use of recurrent neural network with two context sub-layers and learning algorithm. The result shows that this recurrent neural network can provide good convergence and high precision, and that the model modeled ty this network is more reasonable and gives dynamic characteristic of object nicely.The implementing techniques of non-linear system control are analyzed. The neural control system based on recurrent neural network is given, at the same time, neural controller and learning algorithm are disccussed. In addition, the control deviation is used as one input on the neural controller, which can improve the response capacity of controller. At last the control result obtained in practice on continuousdynamic process CSTR is illustrated by using the neural control system mentioned above.Finally, the intelligent modelling and control system is applied into the AIChE benchmark plant, which is a nonlinear, time-lag, mlti-variables unknown system. In this benchmark plant, there are 12 process manipulated variables, 41 process measurements and 20 process disturbances. Besides this benchmark only provides control constraints and objectives. Control result shows that the cascade-loop control system comletely achieves the control objectives. In view of all results given, it can be concluded that the neural control system based on RNN is doable.