Empirical and feed forward neural networks models of tapioca starch hydrolysis

Abstract

The aim of dynamic modeling of the tapioca starch hydrolysis process is to generate models for forecasting the future product concentration (glucose) from the initial conditions of available process measurements. This paper compares two methods of modeling the tapioca starch hydrolysis process: (1) The empirical approach and (2) the feed forward neural network (FFNN) approach. Experiments were conducted to obtain a set of data for the modeling purpose. The Gauss-Newton method was used for parameter estimation in the empirical analysis and a multilayer neural network with one hidden layer was utilized in the neural networks approach. This study indicates that the FFNN model of tapioca starch hydrolysis produces better predictive accuracy, that is simpler to develop and has a generalization capability compared with the empirical model.