Équifinalité, incertitude et procédures multi-modèle en prévision hydrologique aux sites non-jaugés

Abstract

The objective of the present project is to analyse different methods to predict streamflow in ungauged basins in order to define their limitations and, ultimately, improve upon them. This project contains three distinct parts, each one having contributed to at least one of the seven scientific papers presented in this thesis. The first part consists in reducing parameter uncertainty by improving the automatic calibration algorithms. A comparison and analysis of 10 optimization methods allowed identifying the most appropriate optimization algorithms in hydrological model calibration. The end result was the selection of efficient and robust calibration algorithms which helped in reducing uncertainty linked to parameter equifinality, which is an important hypothesis behind some regionalisation approaches. On a similar note, parameter reduction and fixing techniques (global sensitivity analysis methods in particular) were investigated to remove unnecessary parameters from hydrological models. The aim was to improve parameter identifiability by reducing the parameter space and model dimensionality. It was shown that for the HSAMI hydrological model, 8 to 11 parameters out of 23 could be fixed with little to no loss in validation and regionalisation performance. However, the improved identifiability and reduced parametric uncertainty did not result in improved regionalisation skill. This major finding is directly opposed to the generally accepted parsimony concept. The second part of the project involves multi-model averaging of hydrological simulations, in which it was shown that hydrographs generated by different models can, when weighted and averaged appropriately, offer better performance than that of any single model. A comparison of popular model averaging weighting schemes as well as the development of a new technique made it possible to select the best tools to attempt regionalisation approaches under a multi-model averaging framework. In this part of the study, three models were used to attempt predicting streamflows at ungauged sites with the main regionalisation methods. Even though the results showed that model averaging did not perform as expected in regionalisation, it was found that model robustness is important in this type of application. A second project made use of the multi-model averaging concept in a novel way. In that paper, models were run multiple times on the same catchment with different sources of inputs on each run. The resulting hydrographs were averaged according to the multi-model averaging framework. The results showed that the performance was better than by using a single source of climate data and multiple models. The best scenario was a combination of multiple models and multiple sources of input data. It is recommended that hydrological simulations be ran with multiple models and multiple inputs for more robust estimates of hydrological response. The last part of the project directly targeted the regionalisation methods themselves. It was shown that parameter equifinality plays a very minor role in the regionalisation methods’ skill. It was also demonstrated that the hydrological model’s performance is more important than parameter identifiability as it can simulate the complex interactions between the model parameters and catchment physics to a certain extent; something that the linear regression methods were constantly unable to do. Another important contribution made in this thesis was the creation of a new hybrid regionalisation method which slightly outperformed the classic methods. Finally, the regionalisation approaches were analysed in a virtual world setting, which is a reconstruction of the real world inside a Regional Climate Model. This allowed investigating their strengths and weaknesses from within a uncertainty-free environment. Climate, hydrometric and physical characteristic data are physically coherent and present the advantage of working in a "world" without missing data or measurement biases. Within this virtual-world, it was shown that the physical characteristics (catchment descriptors) are not sufficiently precise or adequate to explain the regionalisation mehtods’ performances. It was also presented that the uncertainty related to climate and hydrometric data are less important than previously thought.

Date	15 May 2015
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	François Brissette (Supervisor) & Daniel Caya (Co-supervisor)