Neural networks for prediction of ultrafiltration transmembrane pressure - application to drinking water production

Modelling of ultrafiltration plants for drinking water production appears as a necessary step before plants control and supervisory. It first requires a better knowledge about membrane fouling by natural waters. The phenomena involved are very complex, because of the nature of the fluid concerned: water. Thus up to now phenomenological model cannot be applied for resource waters. Because of their properties, new modelling tools called neural networks seem to be a promising way to model complex phenomena and therefore to be applied to water treatment. In the present study a neural network is used to model the time evolution of transmembrane pressures for ultrafiltration membranes applied to drinking water production. Different network structures and architectures have been elaborated and evaluated with the aim of computing the pressure at the end of a filtration cycle and after the next backwash. For some of these networks a very good accuracy is obtained for both pressures predictions. The inlets are permeate flow rate, turbidity during the cycle and pressure measurements at the cycle start and at the end of the previous cycle. These networks are able to model the effect of both reversible and irreversible fouling on pressures even if no inlet parameter concerning organic matters is considered. (C) 1998 Elsevier Science B.V. All rights reserved.