Modeling of distillation column and reactor dynamics using artificial neural networks

Artificial neural networks (ANN's) have recently attracted much attention from the chemical engineers because of their potential advantages as follows: ANN's are inherently parallel and can learn nonlinear relations from sets of data. The application of ANN's to the modeling of dynamic chemical engineering processes has recently been discussed in several papers, however the applications were limited to some specifically selected examples. In this thesis, several typical chemical engineering processes are classified according to their degree of nonlinearity and dynamics, and representative examples in each of four classes are selected, i.e. distillation column, isothermal CSTR, nonisothermal CSTR, and batch reactor. The dynamic modeling of the selected processes using ANN's, mainly recurrent networks, is studied. Some basic research about the various parameters of ANN's is also discussed in the beginning of the thesis.; As a result of basic research, some of the rules for the structural simplification of ANN's are verified. Noise filtering ability of ANN is also discussed. Several variations of the fully recurrent network for better applications are proposed. Through the study of application of ANN to a distillation column, the importance of considering the process dynamics is emphasized. Through the dynamic modeling of an isothermal CSTR using recurrent nets, the relation between the physical parameters and the ANN parameters are discussed. Various dynamic behaviors of a nonisothermal CSTR are modeled using recurrent nets, and a training strategy to overcome the convergence problem of ANN for difficult examples is proposed. Also a control method based on the developed ANN models is proposed. Finally, ANN model for the early on-line runaway prediction of a batch reactor is developed and its robustness to noise and process changes is demonstrated.