Adaptive Neural Network Modeling And Predictive Control Of Thermal Processes

With the application of DCS in automatic control systems for thermal power plant, advanced control strategies have been adopted in thermal processes control. But as what can be seen in many thermal power plants, we yet cannot obtain good effects on thermal processes control, because the traditional PID control systems have been widely used in almost all the thermal power plants. And this is of disadvantage to the safety and economical efficiency of power plants. The answer to this question involves two main reasons as follows.First, when the thermal system's state changed, the dynamic performance of controlled plants often varied greatly, that is to say, there lies serious nonlinearity in controlled plants. So the traditional PID control system-which is designed on the basis of plants performance on certain load point-can not be used to realize process optimal control in whole load range. What is more, on the contrary to the linearity of PID control systems, there are nonlinear responses in controlled plants. As we all know, no matter how we improve the capability of PID control systems, there must be a limit in the improvement of control effects. We should adopt new optimal control systems which are based on whole nonlinear models for thermal processes to improve control effects. So it is obvious that whole nonlinear modeling for thermal processes is the base to the optimal control in whole load range. Due to their powerful ability of approximating nonlinear functions, and with the characteristics of adaptive learning, parallel and distributed processing, strong robustness and fault tolerance, Artificial Neural Networks (ANN) have been an effective approach to model and control the unknown and uncertain nonlinear systems. Just with this background, we study to use ANN to solve the problems in modeling and control of thermal processes.Among all the algorithms, Radial Basis Function (RBF) Neural Networks have attracted extensive attention and been widely studied these years mainly because of its ability to model arbitrary nonlinear mapping, simple network structure, linear relationship between the network weights and the output such that a linear optimal algorithm can be employed for weight updating. Especially, the great progress has been done in the study of dynamic RBF Neural Networks. Resource Allocating Network (RAN) not only gave the solution to the problem that static NN cannot be used in sequential learning, but also take great success in application in correlative fields. But in the research, we can see that there are yet many problems in using RAN to build thermal processes models, such as people can hardly choose adequate net parameters. And those problems should be overcame in applications.On the basis of careful studying on RBF NN, this paper gives a in-depth research on RAN and Minimal RAN (MRAN) further. To solve the problem that there are too many adjustable parameters in RAN and MRAN for people to adjust and the final network cannot be forecasted in the beginning, this paper proposes a new dynamic RBF NN algorithm―Approximate Correlation Network (ACN)―which was based on the criterion of the correlation in vectors of the output matrix of hidden units. Further, this paper proposes another algorithm-Local Projection Network (LPN). The new algorithm worked out a solution for the problem proposed above. Through many comparisons between the new algorithms and RAN, we can draw a conclusion that the proposed algorithms are superior to RAN in all aspects.With the new algorithms proposed, we obtained accurate nonlinear NN models of typical thermal processes. The new algorithms are adaptive in modeling and the application is easy to realize, which prove that using NN to build nonlinear thermal processes models is available.Finally, the paper carefully studied the theory of nonlinear NN thermal processes models based Dynamic Matrix predictive Control (DMC). Through having applied dynamic RBF NN into DMC, this paper proposed a nonlinear NN models based predictive control algorithm. After detai...