Bibliographie complète

Streamflow forecasting is important for managing water resources in sectors like agriculture, hydropower, drought management, and urban flood prevention planning. Our study examines short and long lead-times to create a framework for streamflow forecasting that can benefit water resource management and related sectors. To improve streamflow forecasts for up to ten days of lead-time, the study first focuses on improving initial conditions using an ensemble Kalman filter as a data assimilation method. The goal is to regulate the hyperparameters of the ensemble Kalman filter for each season to produce more accurate forecasts. A sensitivity analysis is conducted to identify the best hyperparameter sets for each season, including uncertainty in temperature, precipitation, observed streamflow, and the water content of three state variables - vadose zone, saturated zone, and snowpack - from the CEQUEAU model. Results indicate that improving initial conditions with the ensemble Kalman filter produces more skillful forecasts until a 6-day leadtime. Temperature uncertainty is particularly sensitive and varies across seasons. The vadose zone state variable was identified as the most important and sensitive state variable, and updating all state variables systematically may not be necessary for improving forecast skill. Recent machine learning advances are improving short-term streamflow forecasting. One such method is the Long Short-Term Memory (LSTM) model. In general, neural networks learn from regression as relationships exist between input-output. However, LSTM models have a feature named ‘forget gate’, which enables them to learn the relationship between inputs (e.g., temperature and precipitation) and output (streamflow), and also to capture temporal dependencies in the data. The study aimed to compare the performance of the Long ShortTerm Memory (LSTM) model with data assimilation-based and process-based hydrological models in short-term streamflow forecasting. All three models were tested using the same ensemble weather forecasts. The LSTM model demonstrated good performance in forecasting streamflow, with a Kling-Gupta efficiency (KGE) greater than 0.88 for 9 lead-times. The LSTM model did not incorporate data assimilation, but it benefited from observed streamflow until the last day before the forecast. This is because the LSTM model learned and incorporated knowledge from the previous days while issuing forecasts, similar to how data assimilation updates initial conditions. The study results also showed that the LSTM model had better performance up to day 6 of lead-time compared to the data assimilation-based models. However, training the LSTM model separately for each lead-time is a time-consuming process and is a disadvantage compared to the data assimilation-based methods. Nonetheless, the study demonstrated the potential of machine learning techniques in improving streamflow forecasting. The forecasting of streamflow for long lead-times such as a month usually involves the use of historical meteorological data to create probable future scenarios, as meteorological forecasts become unreliable beyond this lead-time. In this study, we proposed a novel method for streamflow forecasting based on ensemble streamflow forecasting (ESP) filtering, using a Genetic Algorithm (GA) to filter forecast scenarios. This method quantifies the potential of historical data for each basin. This potential could be utilized to enhance the accuracy of streamflow forecasts. We sorted the selected and unselected scenarios to find out the common features between them, but the results did not help distinguish between the two groups. Nonetheless, the GA method can be used as a benchmark for future studies to improve longterm streamflow forecasting. This method can also be used to compare different forecast methods based on the potential shown by the GA method for a specific size of ESP members. For instance, if a method uses large-scale climate signals to filter ESP members, the forecast skill result could be compared with the potential of historical data for that particular size of ESP members.