Comment quantifier l’incertitude prédictive en modélisation hydrologique ? : Travail exploratoire sur un grand échantillon de bassins versants

Abstract : Hydrological modelling aims to quantify the rainfall-runoff relationship at the catchment scale. Although hydrological models are generally able to provide an acceptable representation of the catchment behaviour, this representation is necessarily simplified and imperfect, as a consequence an evaluation of uncertainties is desirable. This thesis focuses on the quantification of predictive uncertainty in hydrological modelling. Our main objective was to explore several methods that can be used to complete the deterministic predictions of a rainfall-runoff model with probabilistic distributions. Two prediction cases were distinguished, namely simulation and forecast. We set up a comparative framework to evaluate different uncertainty quantification methods on a large set of French catchments. In simulation mode, we focused on two methods related to the parameter estimation issue, the GLUE uncertainty method and the Bayesian framework, as well as two more pragmatic approaches, a multi-model approach and the post-processing method. Our results suggest that the GLUE-like methods based on ensembles of various parameter sets do not provide an adequate representation of the total predictive uncertainty. An external model error is necessary. Post-processing methods that are flexible enough to adequately describe the residual errors of the hydrological model obtained during calibration give more satisfactory estimates of the modelling uncertainty. We also present a method able to transfer model uncertainty estimates from gauged to ungauged catchments. Our results demonstrate that the method holds interesting perspectives, providing in most cases reliable and sharp uncertainty bounds at ungauged locations. In a forecasting context, we compared several post-processing methods and evaluated the interactions between data assimilation and post-processing in an ensemble forecasting modelling chain. Results show the crucial role of the lead time on the estimates of predictive uncertainty. They also suggest that some improvement can be achieved when the evolution of flows is better taken into account by a post-processing method. Last, we investigated the interactions between data assimilation and post-processing in hydrological ensemble forecasting and showed the benefits of using both in an ensemble forecast system.
Document type :
Theses
Complete list of metadatas

Cited literature [98 references]  Display  Hide  Download

https://pastel.archives-ouvertes.fr/tel-01130084
Contributor : Abes Star <>
Submitted on : Monday, March 9, 2015 - 5:54:56 AM
Last modification on : Wednesday, September 4, 2019 - 3:14:02 PM
Long-term archiving on: Monday, April 17, 2017 - 4:20:19 AM

File

2014AGPT0016.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01130084, version 2

Collections

Citation

François Bourgin. Comment quantifier l’incertitude prédictive en modélisation hydrologique ? : Travail exploratoire sur un grand échantillon de bassins versants. Sciences de la Terre. AgroParisTech, 2014. Français. ⟨NNT : 2014AGPT0016⟩. ⟨tel-01130084v2⟩

Share

Metrics

Record views

713

Files downloads

1088