Étude d'un ensemble multi-modèle multi-calibration pour la détermination de l'ensemble optimal pour la prévision hydrologique

Abstract

With the drought situation becoming more and more serious in the province of Quebec, the government has decided to equip itself with a hydrological forecasting system with a 1 to 3 month horizon. The point of this project is to identify the optimal combination methods for hydrological model ensembles. There are two types of models used in this study: global (HMETS, HSAMI and MOHYSE) and spatially distributed models (HYDROTEL). These models were used on 10 North American Catchments. The global models are calibrated using 9 different metrics creating ensembles of 27 members for the first 9 basins. Hydrotel on the other hand uses 5 metrics on one basin to create an ensemble of 32 members. The hydrological ensembles were assigned weights by using 9 different weighting methods on an annual, seasonal, flow type and drought class basis. The three best performing combination weighting methods were consistently the Bayesian model averaging (BMA), Granger Ramanathan type C (GRC) and the shuffle complex optimizer (SCA) for the complete annual series. Separating the flows by season, flow type and drought class proved to be an advantage over the training period due to the increased amount of parameters (weights), but brought no consistent improvement when it came to the validation period over the complete annual series combinations. This was possibly due to an over fitting of the parameters. The amount of members used in the ensembles was then questioned. Limiting them to the 12 best according to the Nash-Sutcliffe coefficient proved efficient and increased the performances over the validation period. Following a multi objective calibration, it was determined that an ensemble could be identified for each basin that maximized the performances for the Nash coefficient. There existed no correlation between these selected models for all the basins, but when limiting the amount of members to roughly 5 (a big operational advantage), certain members tend to disappear, all the while maintaining the majority of the forecasting power of the ensemble optimising strictly Nash. Due to the nature of the ends, the simplest method to implant in a warning system has been suggested through the use of the GRC weighing scheme due to its simplicity and quickness of execution. This approach should be used with temporally invariable weights annually. For an operational application, ensemble reduction is greatly recommended.

Date	11 Nov 2014
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	François Brissette (Supervisor)