Model-order reduction based on artificial neural networks for accurate prediction of the product quality in a distillation column

The main aim of this paper is to establish a reliable model of a process behaviour both for the steady-state and unsteady-state regimes. The use of this accurate model allows distinguishing a normal mode from an abnormal one. Therefore, the neural black-box identification by means of a non-linear auto-regressive with exogenous model has been chosen. This study shows another technique for neural model reduction into account the physical knowledge of the process. An analysis of the inputs choice, time delay, hidden neurons and their influence on the behaviour of the neural estimator is carried out. After describing the system architecture, a realistic and complex application as a distillation column is presented in order to illustrate the reliability of the prediction and model reduction. Satisfactory agreement between identified and experimental data is found and results show that the neural model successfully predicts the evolution of the product composition.