Neural networks: a tool to improve UF plant productivity
Résumé
The aim of this work is to develop a predictive control algorithm to improve the productivity of an ultrafiltration pilot plant producing drinking water from surface raw water. The objective is to avoid irreversible fouling, that means to get a constant quality of the membrane after backwash and at the same time to optimise the productivity of the process. This control strategy is based on a model able to predict long-term performances of the ultrafiltration pilot plant. This model consists in two interconnected recurrent neural networks coupled with the Darcy's law. The parameters taken into account are water quality parameters, operating conditions during filtration time and during the backwash procedure. The model allows predicting satisfactorily the filtration performances of the experimental pilot plant obtained for different water quality and changing operating conditions, during several hundred filtration cycles.