Bibliographie complète

Reliable long-term streamflow forecast is essential in water resources management and plays a key role in reservoir management and hydropower generation. Properly framing the uncertainty is the key issue in providing a reliable long-term streamflow forecast, and probabilistic forecasts have been used to this effect. In a probabilistic approach, each observed historical data is taken as a possible realization of the future. Non stationarity of hydrometeorological variables, either due to the climate internal variability or anthropogenic change, is another important problem for long-term streamflow forecasts as it is becoming increasingly clearer that past historical data may not adequately represent the current climate. Therefore, there is a need to develop flexible approaches taking into account non-stationarity for long-term streamflow forecasts. Resampling past historical time series is the main approach used for probabilistic long term streamflow forecasts. However, non-stationarity is a key issue of resampling approaches. One possible approach is to make use of a stochastic weather generator coupled to a hydrological model to generate long-term probabilistic streamflow forecasts. Weather generators can easily be modified to account for climatic trends and therefore have the potential to take non-stationarity into account. However, before weather generators can be modified to account for climate non-stationarity, it is first necessary to evaluate whether the modeling chain consisting of a stochastic weather generator and a hydrological model can generate probabilistic streamflow forecasts with a performance similar to that of more traditional resampling approaches. The first objective of this study is therefore, to compare the performance of a stochastic weather generator against that of resampling historical meteorological time series in order to produce ensemble streamflow forecasts. Results indicate that while there are differences between both methods, they nevertheless largely both perform similarly, thus showing that weather generators can be used as substitutes to resampling the historical past. Based on these results, two approaches for taking non-stationarity into account have been proposed. Both approaches are based on a climate-based perturbation of the stochastic weather generator parameters. The first approach explored a simple perturbation method in which the entire length of the historical record is used to quantify internal variability, while a subset of recent years is used to characterize mean climatic values for precipitation, minimum and maximum temperatures. Results show that the approach systematically improves long-term streamflow forecasts accuracy, and that results are dependent on the time window used to estimate current mean climatic estimates. The second approach conditioned the parameters of a stochastic weather generator on largescale climate indices. In this approach, the most important climate indices are identified by looking at yearly correlations between a set of 40 indices and precipitation and temperature. A linear model is then constructed to identify precipitation and temperature anomalies which are then used to induce perturbations in the stochastic weather generator. Five different time windows are defined to determine the optimal linear model. Results show that temperatures are significantly correlated with large-scale climate indices, whereas precipitation is only weakly related to the same indices. The length of the time window has a considerable impact on the prediction ability of the linear models. The precipitation models based on short-duration time windows performed better than those based on longer windows, while the reverse was found for the temperature models. Results show that the proposed method improves long-term streamflow forecasting, particularly around the spring flood.