Votre recherche
Résultat 1 ressource
-
Abstract : Flood forecasting plays a pivotal role in effective water resource management and flood risk mitigation. Despite recent advancements, existing ensemble forecasting systems often grapple with issues of unreliability and under-dispersion. The uncertainty in ensemble forecasts can emanate from various sources, including inputs like precipitation, temperature, and streamflow, as well as initial conditions, model structure and parameters, and boundary conditions. Over the past two decades, numerous endeavors have been made to address bias and under-dispersion through the introduction of various statistical post-processing methods in meteorological and hydrological forecasts. However, a significant challenge lies in selecting the appropriate method and strategy in the forecast chain. Limited research has addressed the integration of pre-processing and post-processing in streamflow ensemble forecasting. Moreover, existing studies have generally focused on lumped hydrological models, while the performance of this integration in distributed hydrological models remains scarcely examined. On the other hand, streamflow data, often indirectly measured through rating curves, are particularly susceptible to errors. While various methods have been proposed to estimate rating curve uncertainty (RCU), the impact of RCU on streamflow forecasting system performance remains underexplored. In addition, current post-processing methods utilized in streamflow forecasting systems have disregarded the inherent uncertainty in observational streamflow data. Thus, this thesis delves into the potential and hurdles associated with considering different sources of uncertainty in flood forecasting systems and their impact on system performance. The primary objectives entail 1) assessing and identifying optimal scenarios from pre-processing of both temperature and precipitation and post-processing of streamflow forecasts approaches, or a combination thereof, and 2) evaluating the consideration of rating curve uncertainty on flood forecasting system performance through incorporating into targeted post-processing methodologies. The thesis focuses on a short-range ensemble streamflow forecasting framework spanning lead times of 1 to 5 days for the au Saumon watershed in southern Quebec, Canada. This watershed, being flood-prone, encountered significant challenges during the 2019 and 2017 flood events, including widespread inundation, road closures, and damage to property and infrastructure. Enhancing streamflow ensemble forecasts and upgrading flood forecasting systems holds the potential to greatly benefit decision-makers and the local populace. Chapter 4 (Results part 1), presented as a scientific article, thus delves into assessing and employing various pre- and post-processing strategy scenarios within the flood forecasting framework employing a spatially distributed hydrological model. By applying different statistical processing and bias correction techniques, the performance and quality of ensemble streamflow forecasts were evaluated across different scenarios. The findings highlight biases and under-dispersion as significant factors affecting raw ensemble forecasts. Pre-processing partially improves raw forecasts but doesn't fully address bias in under-dispersed forecasts. Combining pre- and post-processing enhances forecast skill and reliability, albeit with some variations compared to post-processing alone. Integrating flood events into the training dataset and optimizing its length improves the effectiveness of processing methods, underscoring the critical role of data management strategies in enhancing streamflow forecasting systems. Subsequently, Chapter 5 (Results part 2) , evaluates the consideration of rating curve uncertainty, derived from the Voting Point Method (VPM) and Bayesian Rating curve (BaRatin) estimation methods, on the performance of streamflow ensemble forecasts by integrating into a targeted post-processing approach using Weighted Ensemble Dressing (WED) and Cumulative Distribution Function Matching (CDFM) post-processing methods. Post-processing without RCU enhances forecasting system skill and reliability but overlooks the inherent uncertainty in observational data, posing concerns about its effectiveness. Conversely, integrating RCU improves forecast skill, reliability, and effectively addresses observational data uncertainty, offering a notable advantage over traditional post-processing methods. Comparing WED and CDFM post-processing methods highlighted nuanced differences in forecast outcomes, influenced by uncertainty estimation techniques. Both methods showed favorable accuracy metrics, with RCU integration, especially from VPM, notably enhancing forecast quality. However, variations in RCU from different estimation methods may lead to forecast underestimation or overestimation, warranting careful consideration. These findings collectively shed light on the potential and challenges associated with incorporating different sources of uncertainty into flood forecasting systems.