Stratégies optimales de calibration de modèles hydrologiques en contexte de changements climatiques

Abstract

Since the beginning of the Anthropocene, never have such drastic changes been observed in the concentration of greenhouse gases (GHGs) in the atmosphere. This phenomenon is the instigator of current and future disturbances of the hydrological cycle, at global and regional scales. To date, the hydrological models used to simulate streamflows require calibration based on historical hydrometeorological data. In a future where the climate would be considerably altered, these models may therefore be ineffective in studying the impacts of climate change if the current calibration processes remain in place. This thesis aims to study various calibration strategies useful for climate change impact studies. The analysis of these strategies was carried out using the GR4J-Cemaneige rainfall-runoff model, on a sample of 921 North American watersheds. A total of five calibration strategies were tested based on 46 non-consecutive years of hydrometeorological data. In order to target periods with distinct anomalies, validation periods consisted of five years selected according to five different criteria: highest average temperatures, highest/lowest annual precipitation and highest/lowest average streamflows. The remaining 41 years were used for calibration, with samples ranging from 5 to 40 years (increments of 5). Using this methodology, the research also tested the performance of the model when calibrated over time periods of various lengths. This thesis also includes a study of parameter stability through different climates experienced in calibration, which allows us to deepen our knowledge on the temporal transposability of hydrological models. Although there is a lot of across-catchment variability in the results, some general conclusions can nonetheless be drawn. Results show that removing some of the coldest years from the calibration period is a reliable practice when the validation period consists of the warmest years. For the remaining four calibration strategies, where the validation conditions are heavily influenced by precipitation, removing a large number of most dissimilar years is the most effective solution. A calibration approach based on all available years (40 years in this case) provides good performance in most cases but is not optimal. On the other hand, using a small number of years (5 to 10 years) is not an optimal approach. In terms of parameter stability, results show that the changes undergone by the parameters X1, X2 and X3 as years of calibration are added, are generally in opposite directions for contrasting calibration climates. For example, the parameter X2 mainly shows a decrease for the dry years strategy, while it undergoes an increase in the case of wet years. Regardless of the calibration strategy used, a high calibration performance of the model on a watershed guarantees increased stability of the four parameters of GR4J. Globally, the two CemaNeige parameters seem more stable for catchments where the average annual snowfall exceeds 200 mm and the calibration years include more cold years. This stability is however more obvious for the coefficient of weighting of the thermal state of the mantle (X5) than for the degree-day factor (X6).

Date	3 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	François Brissette (Supervisor)