In recent years, various controller design techniques using neural networks have been developed for nonlinear process control. In the typical approach, neural networks are trained to model chemical processes and subsequently used in model based control schemes (Aoyama et al., 1995, Draeger et al., 1995, Psichogios and Ungar, 1991, Turner et al., 1996). However, previous attempts to use neural networks in the process modelling and control were mainly on simulation experiments or small scale laboratory processes with a few exceptions (Turner et al., 1996). This paper describes a step-by-step procedure of a nonlinear ARX model identification of the %C5 content of the LPG output of a Catalytic Reformer Debutaniser Column using a neural network. The data used were one month’s normal operation data. The result shows that the importance of data pre-processing. The neural network identified using only the selected key inputs out-performed the one identified using all available data. This work is significant not only for its application of a neural network to an industrial process, but also for the development of a methodology in which a nonlinear modelling is coordinated with various linear statistical analysis techniques.