Analyse du post-traitement par transformation quantile-quantile pour le pré- et post-traitement de prévisions hydrologiques

Abstract

This study delves into the enhancement of short-term probabilistic meteorological and hydrological forecasts through statistical treatments, aimed at optimal water resource management and risk prevention associated with extreme events. The study focuses on forecasts produced by 8 lumped conceptual rainfall-runoff models for 42 Quebec watersheds, employing the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm for calibration. The forecasts rely on ensemble meteorological data, and they are enriched through data assimilation, culminating in a set of 1250 hydrological scenarios. Emphasis is placed on the pre- and post-processing of data using Quantile Mapping (QM) with a sliding window approach for calibration data selection, assessed through the Continuous Ranked Probability Score (CRPS). The findings show that pre-processing considerably improves temperature forecasts but has a minimal impact on precipitation. Additionally, the benefits of pre-processing weakly propagate following simulation by the hydrological models. Conversely, post-processing leads to substantial alterations in simulated flows, with both positive and negative impacts. Postprocessing with QM enhances the performance of less accurate models but can slightly degrade the performance of the most efficient models. Among the 42 basins studied, raw forecasts remain superior for 16, indicating that no single treatment method distinctly outperforms across the analysed models and basins. The data selection method used for QM indicates that wider windows (±30 days) provide more effective corrections. However, the study also uncovers critical limitations, notably the limited quantity of available data. Consequently, the initial choice of a high number of quantiles for QM, relative to the available data, underscores a crucial methodological consideration. Acknowledging this, a reduced number of quantiles would have been more appropriate given the data limitations. This realization highlights the importance of matching statistical methods with the volume of available data and suggests that alternative strategies might be explored in future research to overcome such limitations. Although the study demonstrates the ability of statistical treatments to adjust and enhance hydrological forecasts, it also sheds light on the inherent complexity of predicting hydrological regimes and the challenges associated with statistical treatment methods. These observations encourage a more targeted approach and critical reflection on the choice of methods based on the specific characteristics of each watershed and the available data.

Date	18 Mar 2024
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor) & Annie Poulin (Co-supervisor)