Optimisation of ultrafiltration pilot plants requires a better knowledge of membrane fouling. Zn the field of drinking water production, phenomena involved in fouling are very complex and interdependent because of the numerous compounds contained in raw waters. As no knowledge model is available for this application, a statistical modelling tool called neural network is used in this paper to predict the total hydraulic resistance at the end of a filtration cycle and after next backwash, using some parameters concerning water quality (turbidity and temperature) and operating conditions, for a given experimental site. Different network structures have been evaluated, using information concerning the current filtration cycle and the previous cycle. Some of them allow a prediction of resistance with a very good accuracy. They take into account as network inlets the permeate flow rate, pressure and water turbidity, and are able to model the effects of reversible fouling on resistance.