Skip to Main content Skip to Navigation
Theses

Application de l'apprentissage artificiel à la prévision des crues éclair

Abstract : The need for accurate predictions of flash floods has been highlighted by the recent occurrences of catastrophic floods. The scope of this thesis is therefore to apply machine learning to forecast cévenol flash floods, which have caused casualties and huge damage in France over the last 20 years. The work was performed in the framework of the Bassin Versant Numérique Expérimental Gardon project, initiated by the French Ministry in charge of Sustainable Development. In this context, this work addresses the capability of machine learning to provide flood forecasts in the absence of rainfall forecasts The first part of this manuscript describes the family of functions chosen in the present study - neural networks-, their ability to forecast the behavior of non-linear dynamic processes, their fundamental properties (universal approximation and parsimony), as well as the conventional methods used to prevent overfitting. The second part of this work presents the river under investigation, the Gardon d'Anduze, as well as related hydrological studies. The third part presents the application of two traditional regularization methods: early stopping and weight decay. In order to allow the prediction of very intense floods, an original variable selection method is proposed: “partial cross validation”. After careful variable and model selection, the ability of models, obtained by either regularization method, to predict the most dramatic event of the database (September 2002) is assessed, thereby allowing an early warning of the populations. Thus, this work demonstrates that, in contrast to statements found in many publications on neural networks applied to flash-flood forecasting, the prediction of an event that is more intense than the events present in the database is feasible, provided a rigorous methodology is used. For this reason, this work opens the way to making current models more adaptive, and to applying the method to ungauged basins.
Complete list of metadatas

Cited literature [24 references]  Display  Hide  Download

https://pastel.archives-ouvertes.fr/pastel-00005626
Contributor : Ecole Espci Paristech <>
Submitted on : Thursday, October 7, 2010 - 9:45:27 AM
Last modification on : Tuesday, October 20, 2020 - 10:47:08 AM
Long-term archiving on: : Monday, January 10, 2011 - 10:55:06 AM

Identifiers

  • HAL Id : pastel-00005626, version 1

Citation

Mohamed Samir Toukourou. Application de l'apprentissage artificiel à la prévision des crues éclair. Hydrologie. École Nationale Supérieure des Mines de Paris, 2009. Français. ⟨NNT : 2009ENMP1669⟩. ⟨pastel-00005626⟩

Share

Metrics

Record views

1185

Files downloads

5254