In many countries, rivers are the primary supply of water. In France in 2013, 73 % of total withdrawals (38 km3) came from rivers (Banque National des Prélèvements en Eau, Chataigner and Michon, 2017). A number of uses are concerned (thermal power plant cooling, hydropower, drinking water, irrigation, industry, navigation) and can be strongly affected by water shortages in rivers (Bousquet et al., 2003). Furthermore, uses should be compatible with maintaining the quality of aquatic life, through environmental constraints like minimum environmental flows (Acreman and Dunbar, 2004).

There is a need for the early anticipation of low-flow periods to improve water management and to take more timely measures to mitigate the socio-economic and ecological impact of water shortages (Chiew and McMahon, 2002; Karamouz and Araghinejad, 2008). Extreme droughts which occurred in France in 1976, 2003, and more recently in 2011, 2015 and 2017 underline the need for forecasting systems, which is strengthened by the perspective of having more frequent and severe low flows in summer in the context of climate change.

In 2011, the French Agency for Biodiversity (formerly ONEMA) and the Ministry for the environment launched a research project to compare and evaluate the ability of various hydrological models to produce low-flow forecasts useful for real-time decision making. This project led to the development of a low-flow forecasting tool that includes the tested hydrological models.

There are a few existing approaches and tools for low-flow forecasting. A detailed review can be found in Nicolle et al. (2014). More recently, a few works have been carried out on this topic. Some examples are commented here:

conditioning methods for input scenarios for seasonal streamflow forecasting, as tested by Crochemore et al. (2017). These authors investigated the impact of conditioning methods on the performance of seasonal streamflow forecasts, to identify forecast attributes leading to improvement or deterioration using these methods.

In France, a few operational tools have been recently developed, indirectly or directly linked to low-flow forecasting. One can mention:

the Aqui-FR project (Habets et al., 2015) that aims integrating hydrogeological models to monitor and forecast groundwater resource at medium range to seasonal scale, on the main aquifers of France.

The objectives of the article are to present the main characteristics of the low-flow forecasting tool and the results obtained in operational conditions over the 2017 summer period on a case study.

Models are expected to forecast streamflow from time steps t+1 to t+L (with L the lead time), knowing both observed meteorological inputs and streamflow until time step t and making assumptions (i.e. choosing scenarios) for the future meteorological inputs from t+1 to t+L. Streamflow observations can be used within an assimilation scheme and/or a statistical correction procedure.

Figure 2 presents the successive steps of the forecasting method for low-flow forecasting with hydrological models:

internal states of hydrological models are initialized using climatic observations of past conditions until the day of forecast;

This approach is quite classical. The originality in the case of PREMHYCE is that it is applied in a multi-model framework. This has two potential advantages: to improve the resulting forecasts and to better account for structural uncertainty.

The PREMHYCE operational tool adopts the classical ESP approach (Day, 1985) in terms of future meteorological inputs. For a given catchment, let us consider that N years of past historical climatic observations are available. In real time, one wishes to make a forecast on a calendar day t of a year Y within the test period, i.e. to forecast flows between calendar days t+1 and t+L. The observed meteorological data available between days t+1 and t+L in the years 1 to N (i.e. N scenarios) are used as input scenarios to the model, considering that they are likely meteorological conditions for this period of the year. Here, 57 years (1959–2016) of daily climate data from the SAFRAN reanalysis were used for the tests during the 2017 year and 58 scenarios (1959–2017) for the 2018 year.

A zero-precipitation scenario (i.e. precipitation equal to 0 for the L next days) is also used to provide the worst-case streamflow forecast. It is associated to a daily interannual average of potential evapotranspiration.

PREMHYCE is a low-flow forecasting platform that aims at providing low-flow forecasts at a daily time step for the next 90 d to end-users. It is based on the airGR package (Coron et al., 2017a, b) and includes five hydrological models: Gardenia (BRGM), GR6J (Inrae), Mordor (EDF, currently implemented under a simplified form for technical reasons, here called IrMo), Presages (Lorraine University) and the Safran-Isba-Modcou (SIM) modelling suite (Météo-France). More detailed information on these models and their practical implementation for forecasting are given by Nicolle et al. (2014). Note that the SIM model is run by Météo-France independently from the platform and only the SIM outputs are uploaded into the PREMHYCE tool. This platform is currently hosted by a server at Inrae and allows data exchange via FTP protocol.

The tool includes two main modules:

An off-line module for the calibration of hydrological models, to estimate models parameters and evaluate models reliability. The module functions allow to:

import the catchment database (observed time series of streamflows, precipitation, PE and temperature);

The model parameters can be either calibrated within the platform or specified by the user. As a physically-based model, SIM has been set up all over France and its parameters are not further tuned for the PREMHYCE objectives. Note that SIM is the only model for which no calibration against observed flow data at the catchment outlet is performed. The spatially distributed parameters used in this model are estimated regionally.

Users provide a list of catchments that have to be in the French HYDRO database. For each catchment, up to four operational streamflow thresholds can be also provided, which will be used to interpret the severity of future forecasted low flows.

Models are calibrated using a gradient-type method for the GR6J, IrMo and Presages models. Two objective functions (KGE, Gupta et al., 2009; NSE, Nash and Sutcliffe, 1970) can be used with three prior transformations of streamflow (Q, (Q), ln(Q)). A compromise can be sought between several parameter sets obtained by these various functions. Gardenia is calibrated using the Rosenbrock method with the NSE objective function calculated on ln(Q).

The tool can be run every day to provide low-flow forecasts at the daily time step. Given the computing constraints associated with the SIM model, the SIM streamflow forecasts are provided only once a month.

Several operations are made every day to provide low-flow forecasting.

The last observed data until the day of forecast are first used to update model internal states. A data import process has been implemented to retrieve meteorological and hydrological data.

Every day, Météo-France provides real-time gridded SAFRAN data (precipitation, temperature) for the day before forecasting, as daily data for the day of forecast is not yet available. Every month, Météo-France provides the gridded SAFRAN reanalysis of these meteorological data for the month before, to consolidate real-time SAFRAN data. These data are averaged at the catchment scale for each catchment.

Users can provide the last observed streamflows up to the day before forecasting, as daily streamflow for a catchment at the day of forecast is not yet available. These observed streamflow data are used within assimilation schemes or post-correction procedures for low-flow forecasting. If observed data is not available during the seven last days, models do not use assimilation schemes or post-correction procedure.

The database containing the streamflow and meteorological data is updated on a daily basis and initial models states are computed with these data at the day of forecast.

The forecasts are then computed by applying the ESP method described above.

For each model, the PREMHYCE tool provides streamflow forecasts for each scenario and for the no precipitation scenario, for the next 90 d, under numerical or graphical formats. Figure 3 shows an example of synthesis plot provided by the tool for a forecast issued on 11 April 2017 for the Ill River at Didenheim with the GR6J model. It represents:

Quantiles (0.1, 0.25, 0.5, 0.75 and 0.9) of the distribution of streamflow forecasts for the next 90 d (blue envelop curve and blue dashed line).

Here are also represented observed streamflow (black line), and simulated streamflow by GR6J using observed meteorological P and PE (brown line), that are normally not available in real-time.