École Nationale Supérieure de
Techniques Avancées

Résumé

The hydrometeorological low-water forecasting chain at EDF-DTG involves using air temperature and precipitation ensemble data as inputs to a hydrological model called MOR- DOR. Once this model has been calibrated for the catchment studied, it transforms the meteorological forcings into flow data at the catchment outlet. These ensemble forecasts enable us to produce statistics on the obtained flow distributions. In order to evaluate the forecasting chain, low-water forecasts are replayed: 6 seasonal forecasts per year over the past 35 years. The CRPSS is the metric used to assess the quality of the forecasts in relation to a reference distribution, which is the set of climatological flow observations. To improve low-flow forecasts, several modifications are tested both upstream and downstream of the model. Upstream, we experiment with different input forcing data, data assimilation techniques to better initialize the model before the forecast begins, and calibration of one of the model’s reservoirs using piezometric data. Downstream, we employ statistical post-processing methods such as bias correction, recalibration, and dressing directly on the forecasts. An important finding from these tests is the strong dependence of low-water forecasts on the groundwater component of the catchment. Greater groundwater weight leads to a better predictability. Forecasts are clearly improved, nearly systematically by statistical post-processing, which corrects errors that develop within the modelling. Data assimilation improves low-water forecasts significantly when they are made well in advance, whereas recalibration proves more effective for short-term forecasts. Finally, the use of seasonal rainfall and air temperature forecasts substantially improves short and medium-term hydrological forecasts and offers promising prospects for enhancement.