20190208_Neural networks for prediction of ultrafiltration transmembrane pressure – application

1. Title, Journal and Authors

Title: Neural networks for prediction of ultrafiltration transmembrane pressure – application to drinking water production

Journal: Journal of Membrane science 150 (1998) 111-123

Authors: N. Delgrange1.a, C. Cabassuda,*, M. Cabassudb, L. Durand-Bourlierc, J.M. Lainec

aLaboratore d’Ingenierie des Procedes l’Environment, Institue National des Sciences Appliquees, Complexe Scientifique de Rangueil, 31077 Toulouse, Cedex, France

bLaboratorie de Genie Chimique, LGC-CNRS UMR5503, ENSIGC 18, Chemin de la loge, 31078 Toulouse, Cedex 4, France

cCentre International de Recherche sur l’Eau et l’Environment, Lyonnaise des Eaux, Avenue du President Wilson, 78230 Le Pecq, France


2. Summary

Ultrafiltration process that has good efficiency to remove particles and to keep good water quality is upward trend in the field of drinking water. It is essential with modeling of ultrafiltration process to solve the membrane’s reversible fouling caused by particle deposition and irreversible fouling caused by adsorption. In this paper, for the modeling of it, Neural network was used. Neural network does not need complicated formula but it is able to obtain theoretical and practical solutions.

The objectives of this paper are developing the suitable neural network to predict transmembrane pressure which is indicator of membrane fouling and adapting it to the plant. Thus, inlet parameters, experimental data and comparisons between predicted and experimental value are needed to be discussed. And a quasi-Newton learning algorithm was used. A pilot plant was schematized and some parameters are measured.

For operating the neural network, turbidity, Qp, Ptm-b, Ptm-e were considered as inlet parameters and three networks were used to verify the model. R1 had three, R2 had five, R3 had four inlet parameters. Seven sets of data are used for training and three sets of them are used for testing. As a result, R2 which had five hidden layers and R3 which had four hidden layers had the lower errors and no differences. But R1 which had four hidden layers had the higher errors comparing others.

As a result of first test, turbidity and Ptm-e had a peak and became stabled. Backwash had good effects and it meant reversible fouling. In second test, turbidity and Ptm-e had a peak and remained high values. Because backwash did not have good effects so it meant irreversible fouling. The last graph meant if there were no other data such as TOC or UV, we should examine the previous data.


3. Contact

Dae Seong Jeong / Intern student


Phone : +82-10-2003-7860


E-mail : jeongds92@hanmail.net

첨부 (0)