Use Genetic Programming for Selecting Predictor Variables and Modeling in Process Identification

Availability of an accurate and robust dynamic model is essential for implementing the model dependent process control. When first principles based modeling becomes difficult, tedious and/or costly, a dynamic model in the black-box form is obtained (process identification) by using the measured input-output process data. Such a dynamic model frequently contains a number of time delayed inputs and outputs as predictor variables. The determination of the specific predictor variables is usually done via a trial and error approach that requires an extensive computational effort. The computational intelligence (CI) based data-driven modeling technique, namely, genetic programming (GP) can search and optimize both the structure and parameters of a linear/nonlinear dynamic process model. It is also capable of choosing those predictor variables that significantly influence the model output. Thus usage of GP for process identification helps in avoiding the extensive time and efforts involved in the selection of the time delayed input-output variables. This advantageous GP feature has been illustrated in this study by conducting process identification of two chemical engineering systems. The results of the GP-based identification when compared with those obtained using the transfer function based identification clearly indicates the outperformance by the former method.