Analyse de l’impact de l’assimilation de données par filtre d’ensemble de Kalman sur la performance des prévisions hydrologiques à court et moyen terme

Abstract

This study presents an analysis of the impact of data assimilation using the Ensemble Kalman Filter on ensemble hydrological forecast performance. The main objective is to analyze and quantify the impact of data assimilation on the characterization of uncertainty in hydrological forecasts. More specifically, it aims to assess the impact of data assimilation on the performance of short- to medium-term hydrological forecasts, using metrics such as CRPS and Talagrand diagrams. The study was conducted on the Lac-Saint-Jean watershed using three lumped conceptual hydrological models: MOHYSE, Blended and GR4J-CN. The models were calibrated using the Dynamically Dimensioned Search optimization algorithm. Data assimilation was implemented using the Ensemble Kalman Filter with 25 members. Meteorological data were drawn from the ERA5 reanalysis dataset for the period 1960-2023, while meteorological forecasts were obtained from the European Centre for Medium-Range Weather Forecasts ensemble, consisting of 50 ensemble members over 14 days, covering the period 2016-2023. Hydrological forecasts were generated over a 14 days lead time for all three models. To assess the performance of hydrological forecasts, Talagrand diagrams were used to evaluate forecast reliability, while the Continuous Ranked Probability Score was used to assess forecast accuracy. The results show that data assimilation improves forecast quality, with the most impact observed for the GR4J-CN model. This improvement is reflected in reduced forecast errors, increased temporal stability and enhanced reliability. However, the impact of data assimilation varies depending on the model. For MOHYSE, the gains are more limited and a slight degradation in performance is sometimes observed. The assimilated models better capture forecast uncertainty by mode effectively encompassing the observations within their prediction intervals. This trend is observed across all season, although it is more pronounced in winter and autumn.

Date	21 Jul 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor)