Online Implementation of Local Linear Neuro-fuzzy Technique on Partially Simulated Reaction Process

Abstract

This study presents a methodology for on-line identification of the nonlinear reaction process in presence of measurement noise and uncertainty, for accurate simulation of this process, a link between HYSYS (chemical software) and MATLAB was generated by the author. In this link HYSYS simulates the continuous stirred tank reactor (CSTR) and MATLAB performs the data acquisition algorithm. The chemical process used to illustrate this study is the production process of propylene glycol. The local linear models tree (LOLIMOT) identification algorithm has been introduced. The nonlinear model parameters are determined by an incremental tree construction algorithm in a heuristic manner in order to avoid the application of nonlinear optimization techniques. The results show that the LOLIMOT gives best fitting on the train and test data. On the other hand, the limited flexibility of the local estimation reduces the variance error due to the bias/variance dilemma.