Analyse des approches multi-modèle en prévision hydrologique

Abstract

The present study aims to analyze the contribution of multi-model combination methods in order to improve the representation of the variability of ensemble hydrological forecasts generated by simple hydrological models. In the context of the management of hydroelectric reservoirs it often happens that the predictions of hydrological models do not represent the reality. This is partly due to the use of hydrological models greatly simplifies the complex process of the hydrological cycle, thus producing sets of forecasts that do not adequately represent forecast uncertainties. The multi-model combination is therefore an interesting avenue, especially since previous studies have indicated its ability to improve the quality of hydrological forecasts over short periods. Three hydrological models (GR4J, MOHYSE and HSAMI) were used to generate ensemble forecasts issued on the first day of each month and lasting from 2 weeks to 6 months on the Ashuapmushuan watershed in Saguenay-Lac-St-Jean. The combination of these three models by the GRC weighting method allowed to create a multi-model and thus to generate a fourth ensemble forecast. Forecasts were generated for each available year, from 1967 to 2001, using observed weather data to feed the models. Subsequently, combinations of the ensembles were made to add noise to the ensemble. These are the grand ensemble, which is the combination of the three simple models, and the super ensemble, which is the combination of the three ensembles of simple models and the multi-model ensemble. The quality of these sets is then assessed using Talagrand diagrams and two statistical tests, the ABDU scores and the Kolmogorov-Smirnov test. In all cases, the combinations of ensembles demonstrated an improvement in the correction of the sub-dispersion initially observed. The addition of the multi-model ensemble, the super ensemble, in comparison to the grand ensemble, is very beneficial to the short-term dispersion, but this benefit fades over longer horizons. In terms of achieving the objective of producing forecasts with adequate uncertainty, the grand-ensemble and super ensemble combinations show better results, especially during the warmer months when precipitation is in liquid form. The project demonstrates that there is an opportunity to use multi-model methods successfully, but further research is needed to determine how to maximize performance based on forecast periods and the desired time horizon.

Date	1 May 2018
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor) & François Brissette (Co-supervisor)