Amélioration des prévisions hydrologiques d’ensemble dans un contexte de modélisation multi-modèle

Abstract

Whether it is in the context of hydroelectric reservoir management, flood management or any other operation aimed at controlling water inflows, hydrologists use hydrological forecasts to make their operational decisions. To do so, they must consider financial, environmental and, above all, public safety risks. Ensemble forecasting is a type of hydrological forecasting that allows for risk assessment by taking into account the uncertainties in these forecasts. However, hydrological ensemble forecasts are often biased, and their distribution is often undispersed when compared to observations. This problem stems mainly from the uncertainty in the meteorological ensemble forecasts used to feed hydrological models to obtain forecasts of river streamflow, and also from the model(s) used. These erroneous predictions can thus contribute to operational decisions that can endanger the public and infrastructure. This research project presents a new methodology that aims to improve the accuracy of ensemble hydrological forecasts used in water resource management. This methodology has been evaluated on five catchments of the industrial partner Rio Tinto in the Lac-Saint-Jean region of Quebec. ECMWF weather ensemble forecasts (50 members) from 2015 to 2019 were used to inform eight global hydrological models over a nine-day horizon. Following the calibration of the hydrological models, a data assimilation method based on Kalman ensemble filters (EnKF) is used to modify the initial conditions of each hydrological model to represent the different sources of uncertainty in the observations. Then, a post-processing of the hydrological ensembles of each model is performed using the quantile bias correction (QM) method. These same hydrological ensembles are finally grouped into a large multi-model ensemble with the objective of having a better sampling of the total uncertainty, and thus to standardize its distribution. In order to evaluate the results obtained by the methodology and its different steps, several performance criteria are used. Talagrand diagrams, two metrics for quantifying reliability and precision, as well as the Kolmogorov-Smirnov statistical test, are used to evaluate the performance of the methodology. This evaluation is performed over four distinct periods, representing different hydrological regimes over the catchments. Despite difficulties during the spring freshet period, the results obtained indicate that each step of the methodology improves the accuracy of hydrological ensemble forecasts over the five catchments under study. Finally, the use of the multi-model significantly improves the accuracy and extent of hydrological ensembles over a nine-day horizon. This project demonstrates that it is therefore possible to correct biases and improve the reliability of hydrological ensemble forecasts.

Date	11 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor) & Jean-Luc Martel (Co-supervisor)