An evaluation of the Canadian global meteorological ensemble prediction system for short-term hydrological forecasting

摘要

Hydrological forecasting consists in the assessment of future streamflow.Current deterministic forecasts do not give any information concerning theuncertainty, which might be limiting in a decision-making process. Ensembleforecasts are expected to fill this gap.In July 2007, the Meteorological Service of Canada has improved its ensembleprediction system, which has been operational since 1998. It uses the GEMmodel to generate a 20-member ensemble on a 100 km grid, at mid-latitudes.This improved system is used for the first time for hydrological ensemblepredictions. Five watersheds in Quebec (Canada) are studied: Chaudière,Chateauguay, Du Nord, Kénogami and Du Lièvre. An interesting17-day rainfall event has been selected in October 2007. Forecasts areproduced in a 3 h time step for a 3-day forecast horizon. Thedeterministic forecast is also available and it is compared with theensemble ones. In order to correct the bias of the ensemble, an updatingprocedure has been applied to the output data. Results showed that ensembleforecasts are more skilful than the deterministic ones, as measured by theContinuous Ranked Probability Score (CRPS), especially for 72 h forecasts.However, the hydrological ensemble forecasts are under dispersed: asituation that improves with the increasing length of the predictionhorizons. We conjecture that this is due in part to the fact thatuncertainty in the initial conditions of the hydrological model is not takeninto account.