NARX Neural Network Modeling of Batch Distillation Process

The batch distillation model is highly nonlinear due to the influence of mass and composition of the initial material to be separated. It is also caused by the thermodynamics of the system which is not ideal. Therefore, analytical batch distillation modeling does not fulfill superposition theory, or is nonlinear. Thus, any well-established linear control schemes can not be applied here. Meanwhile, modeling a system using the Nonlinear Autoregressive with eXogenous inputs (NARX) is quite widely used today because of its ability to represent non-linear system dynamics. This paper propose a system identification using NARX Neural Network (NARX-NN) to modeling batch distillation process. In order to prove the superiority of NARX-NN's accuracy in the system identification process, a comparison with linear ARMA models is done. In this study, ARMA approximation was carried out in two ways. The first way is to use a neural network, while the second method is through approximation to the discrete transfer function. The validation results show that the NARX-NN achieves significantly better fit compared to the linear models. NARX-NN and ARMA-NN were compared and their MSE ratio for delay input and output 1 and 3, respectively have the smallest value, i.e 1.71e-04