Présentations Colloques

    Oral Presentation
    Session 8.05: Karst aquifers
    Savary Michaël
    Prediction of turbidity using neural networks- application on chalk aquifer of Normandy (France)
    Approximately 25% of the world population drinking water supply depends on karst aquifers. Understanding and protecting karst aquifers is thus a crucial task whose development increases with fresh water needs.**In Normandy, northern France, drinking water is provided by the chalk aquifer which is a highly capacitive and intensively karstified aquifer. In this context, the Yport pumping well managed by the CODAH delivers half of city of “Le Havre Great Le Havre” area, drinking water. Due to the high infiltration properties of the alimentation basin, the rainfall water washes the compound of the surface (biological organism, chemical product,…) and brings them rapidly to the pumping well. This phenomenon causes turbid events forcing the manager (CODAH) to decrease the pumping rate in order to allow treatment. In this context, the aim of the present work is the prediction of turbid event to allow the CODAH to optimally adapt its production process (create water storage before the turbid event to compensate the pumping rate decreasing). Due to the lack of knowledge about (i) physical processes inducing turbidity and (ii) the complex rainfall turbidity relationship, a black box model seems suitable. To this end, thanks to their ability to identify nonlinear behaviours, neural networks are investigated. The database provided by the CODAH is composed of rainfalls recorded on the alimentation basin between 01 07 2009 and 28 04 2015 and the turbidity at the entry of the station between 23 10 1993 to 06 02 2015, both at a hourly time step. The first part of the work consists in analysing data in order to identify links between the variables and the system dynamics, using correlation and spectral analysis. In a second part, the black box model based on neural networks is designed thanks to previously cited variables. Finally, the ability of the neural network model to predict the turbid event will allow to anticipate on the 100 NTU threshold overtaking, 12 hours in advance.


    Nous vous invitons à mettre à jour vos coordonnées postales, téléphoniques et surtout MAIL sur votre fiche d'identification. Merci !

    Logo PFE