Predicting Liquid-vapor (LV) composition at distillation column

This paper will present the development of nonlinear model of distillation column using neural networks approach. The model is accomplished in Nonlinear Auto Regressive with exogenous input (NARX) structure. This distillation column has two input and two output variables. The input variables are heat duty on the reboiler (Qr), and reflux flowrate (L), while the output variables are mole fraction of distillate (Xd) and mole fraction bottm product (Xb). The training as well as validation data were generated using Amplitude Pseudo Random Binary Signal (APRBS) as excitation signal. The structure of neaural networks is feedforward networks, which consists of three layers: input, hidden and output layer. Levenberg-Marquardt algorithm is used as learning algorithm to adjust the weight matrices of the networks. The results show that NN soft sensor base on flow rate correlation is easy to build, fast response, no need special instrumentations, better of reliability compare to analyzer reliability, cheaper, low operational cost, low maintenance cost, and has good Root Mean Square Error (RMSE).