An Approach to Robust and Flexible Modelling and Control of pH in Reactors

Abstract

Preliminary investigations into the potential application of static feed forward neural networks in the dynamic modelling of pH in complex, time-varying systems have been carried out. To assist in network training and testing, a simplified, ‘global first principles’ (FP) model of the pH of such systems was developed, and used successfully to simulate input-output data. Neural networks with input information vectors enhanced by the introduction of auxiliary variables derived from acid-base principles were trained and tested on this data, using both Levenberg-Marquardt (L-M) and heuristic training algorithms. Both algorithms produced good predictions, but the heuristic algorithm required data pre-treatment to minimize its error. However, it trained much faster than the standard, L-M algorithm.