The HBV-96 model is the operational model running within the Delft-FEWS plat-form at Deltares (see Chapter 4) (Berglov, 2009). HBV-96 is a conceptual semi-distributed hydrological model, originally created at the Swedish Meteorological and Hydrological Institute (SMHI) in 1996 from a re-evaluation of the lumped HBV hydrological model (Lindstrom et al., 1997; SMHI, 1999) (Figure B.1, Ap-pendix B).
The model runs with hourly time steps over sub-basins (about 40 km2) using the following input meteorological data: hourly precipitation and air temperature and potential evapotranspiration monthly mean averages. Computation of precipitation and air temperature is done through an optimal interpolation method, and potential evapotranspiration is corrected using altitude. Sub-basins are subdivided into zones defined by elevation and vegetation spatial distribution. Four different land use classes are used within the model: open areas, forests, lakes and glaciers (Berglov, 2009).
HBV-96 is structured in multiple routines: i) the precipitation routine, ii) the soil moisture routine and iii) the runoff response routine. The latter is composed of two zones, an upper non-linear reservoir and a lower linear reservoir. The upper zone reservoir generates a quick runoff flux, while the lower reservoir produces as an output a baseflow component. The outputs of the reservoir are then summed prior to being subjected to the MAXBAS tranformation function (Figure B.2, Appendix B). This function is a triangular weighting function and is used as a routing func-tion to compute simulated runoff. Between sub-basins, the routing is achieved by the use of modified Muskingum’s equations and delay parameters (lag and damps) (Lindstrom et al., 1997; Berglov, 2009).
The HBV model has applications in flood modelling and operational flood fore-casting in more than fifty countries all over the world. From multiple compari-son framework, it has been established that the latter produced moderate to good flood simulations and forecasts. This quality, adjoined to its simplicity, conferred by its low number of parameters, has made the HBV-96 model adequately ad-justable to an extensive number of cases, especially to large sub-basins such as the Rhine river basin (Berglov, 2009; Gorgen et al., 2010; De Wit, 2005; Drogue et al., 2010).
Many developments have been made for the operational use of the HBV-96 for flood forecasting within Delft-FEWS. These include the improvement of: the data interpolation method (Weerts et al., 2008), the evaporation calculations and the precipitation and temperature updating for forecasting (Berglov, 2009).
The GR4H model is the hourly version of the GR4J daily model (Mathevet, 2005). It is a four-parameter empirical lumped rainfall-runoff model created at IRSTEA in the early 1980s. This model aims at reproducing a catchment’s stream response to water inputs events by an ensemble of conceptual (non-physical) processes (Figure C.1, Appendix C; Perrin et al., 2003, 2007).
The GR4H model runs with hourly precipitation data (PK ) and average potential evapotranspiration data (E), lumped over the catchment area, as inputs and has four parameters (X1: the production store maximum capacity (mm), X2: the ground-water exchange coefficient (mm), X3: the routing store maximum capacity (mm) and X4: the time base of the unit hydrographs (hours)) optimized for each basin (Perrin et al., 2003, 2007).
Two conceptual reservoirs are present within this model, a production store and a routing store. The production store behaves conceptually similarly to an unsatu-rated reservoir, controlling the water infiltration into the ground. The routing store, as its name indicates, acts like a groundwater reservoir and participates to the water routing out of the system.
Water is routed linearly by two unit hydrographs (UH1 and UH2), which contri-bution is fixed. Indeed, 90% of the water is routed by UH1 and 10% of the water is routed by UH2. While conceptual, this routing method enables to simulate the time lag between a rainfall event and its corresponding peak discharge. This is done through the ordinates of the unit hydrographs, defined by S curves (SH1 and SH2), that distribute the water entering the unit hydrographs over time (Figure C.2, Appendix C) (Perrin et al., 2003, 2007). Nevertheless, these unit hydrographs can be related to physically based hydrological fluxes, being respectively baseflow and quick runoff. For this project, the GR4H model does not present any snow routine as it was not yet ready on the hourly time step. This should however not be problematic, considering that the snowmelt water contribution to the total discharge does not exceed 15% for the sub-catchments selected. Nonetheless, simulations were anal-ysed carefully for periods during which the contribution of snowmelt water was relatively high.
Numerous intercomparison projects including the GR4H model have lead to the conclusion that the latter performed satisfactorily on average and that an increase in the number of parameters did not show significant improvements of the streamflow simulations quality (Perrin et al., 2003; Mathevet, 2005; Gorgen et al., 2010; Lerat et al., 2006; Perrin, 2000). Nonetheless, several comparisons have also shown that more flexible models perform better than the GR4H model (Demirel et al., 2013; Esse, 2012; Esse et al., 2013; Pushpalatha et al., 2011).
The GR4H model is not used as thus at IRSTEA for operational forecasting pur-poses. Instead, the forecasting model used nowadays, called GRP, was born from several research activities which lead to the modification of the original GR4H structure, as well as the addition of several assimilation procedures to the latter (Tangara, 2005; Berthet, 2010).
OpenStreams is a project initiated by Deltares. Still in progress, this project’s aim is to enhance hydrological models (or model components) re-use in different frameworks. The models or parts of models can be used independently through a command line interface or they can be associated through the use of a layer inherent to OpenStreams, which aim is to execute a standard interface that can process different programming languages (Schellekens, 2012).
Within OpenStreams there is wflow, a distributed hydrological modelling platform. The wflow models are written as a set of PCRaster python scripts imbricated into each other (Wesselung et al., 1996).
The maps, input data, parameters and models states needed for the computation of the models are organized within folders, they constitute the models ’database’.
The computation of the models is done over grid cells, using static PCRaster maps. For the purpose of this project, these maps have a 1 km by 1 km resolution. They are created using a digital elevation model (dem) and a land use map of the consid-ered basin. These maps include: a catchment delineation map, dem maps, a map of the gauges, the original land use map (the land use classes are the following: infrastructures, arable land/pastures/permanent crops, forested areas, other nature, wetlands and water bodies), a local drainage direction (ldd) map, a soil map, a rivers map, a sub-catchment delineation map and maps of the spatial distribution of physical parameters used for the models computations.
Input data are hourly precipitation, potential evapotranspiration and temperature point data interpolated over the maps grid cells.
Model states encompass all the information needed (reservoir contents, fluxes, etc) for starting the model from a selected moment in time. They are kept in the model’s memory after each time step, for the following step.
The routing procedure within the wflow models consists of a kinematic wave func-tion (Shultz et al., 2008), which is an approximation of Saint-Venant’s equations. This function routes water over the ldd network and depends on several parameters, such as channel width and the manning roughness coefficient.
The wflow platform presently includes the wflow hbv and the wflow sbm mod-els; a short description of these distributed hydrological models is given thereafter (Schellekens, 2013).
The wflow hbv model is a fully distributed version of the HBV-96 model, where the original routing function (MAXBAS) was traded for a kinematic wave function (Schellekens, 2013). A complete diagram of the hydrological processes involved in the wflow hbv model is presented in the Appendix D (Figure D.1), together with a description of the model’s variables (Figure D.2, Appendix D). This diagram was produced based on the wflow hbv python script.
The interception values of the wflow hbv model were involuntarily set to zero prior to the model simulations and hindcasts. This was discovered at a late stage of the project, disabling any additional model runs. Therefore, the wflow hbv model results were analysed but did not permit any reliable comparison with the other model results.
Due to their recentness, there is no published literature referring to the wflow sbm and the wflow hbv models performances, apart from the wflow documentation (Schellekens, 2013).
The wflow sbm model originates from the distributed topog sbm simple bucket model developed by Vertessy and Elsenbeer (Vertessy et al., 2000) in the sense that it follows the same soil modelling concept. The following description of the wflow sbm model is based on the wflow documentation (Schellekens, 2013) and on the wflow sbm python script from which a diagram of the processes inherent to the model was made (Figures D.3 and D.4, Appendix D).
The originality of the wflow sbm model lies in its direct runoff component. Indeed, the precipitation water that does not infiltrate into the soil (FreeWaterDepth) runs off directly onto the soil’s surface (similar to an overland flow) and contributes to the Kinematic wave.
The water that infiltrates reaches the soil water reservoir which consists of a bucket composed of an unsaturated store (UStoreDepth) and a saturated store (FirstZoneDepth), working intrinsically (Vertessy et al., 2000). Part of the water present in the satu-rated store seeps to the surface (ExfiltWater and SubCellGWRunoff ) and is added to the Kinematic wave. The other part of the saturated store water content moves horizontally between cells (FirstZoneFlux) based on a D8 network 1, acting as a groundwater flow (Schellekens, 2013).
For the purpose of this project, the Inflow and the Reinfilt fluxes were deacti-vated.
From GR4H to wflow gr4
In order to be executable within Delft-FEWS, the GR4H model had to be incor-porated in the wflow platform according to a pre-defined framework (Schellekens, 2013).
Thus, the original FORTRAN code was first reprogrammed in the PCRaster Python dynamic language. Subsequently, the wflow gr4 model’s database (static maps, initial states, parameters) was created, prior to the model’s link to the platform.
A set of static maps with a resolution 1 km by 1 km was first generated for each of the five sub-basins using as initial maps the dem and land use maps of the wflow hbv and wflow sbm models for the Meuse River basin.
Initial states (the two unit hydrographs water content, the routing store and the production store water content) were then obtained for each sub-basin for the wflow gr4 model by running simulations on the command line using one week of forcing precipitation and evaporation data. These states were used at the start of the simulation period as cold initial states by the wflow gr4 model within Delft-FEWS.
A set of parameters (X1, X2, X3 and X4) for each of the five sub-basins was finally obtained through the wflow gr4 model calibration, performed using a proprietary software developed at IRSTEA, called AirGR. The calibration was based on ob-served hourly discharge measurements using the Nash-Sutcliffe model efficiency coefficient (NSE) as objective function (formula proposed in Nash and Sutcliffe (1970)).
Data import and transformation
Prior to running the hindcasts, precipitation and temperature COSMO-LEPS data were imported into the Delft-FEWS platform and interpolated according to the Kriging interpolation method described previously (see Section 5.2.1).
Subsequently, the interpolated precipitation was perturbed in order to obtain fif-teen different precipitation scenarios. These scenarios were obtained by adding and multiplying the original COSMO-LEPS interpolated precipitation data with a varying factor. These ensembles permitted the production of probabilistic hind-casts.
The hindcasts were executed for the four models (HBV-96, wflow sbm, wflow hbv and wflow gr4) on an hourly basis, from the 5TH of January 2011 at 1 am until the 20TH of January at 1 am. The hindcasts were produced with a lead time of five days.
Furthemore, an error correction procedure was applied to the models hindcasts. This correction procedure, called autoregressive (AR) error correction method, is an automatic algorithm that combines the simulated discharge, the observed dis-charge and the forecast discharge time series of a specific location in order to up-date the forecast discharge for each time step. By establishing the residual between the observed and the forecast discharge, the so called model error, this algorithm predicts the future error in order to correct for it (Minnet, 2014; Broersen and Weerts, 2005).
Table of contents :
2 Research area
2.1 Meuse River basin
2.2 Sub-basins selection
3 The models
3.3 OpenStreams models
3.3.1 wflow hbv
3.3.2 wflow sbm
4 The forecasting platform
5.1 From GR4H to wflow gr4
5.2.1 Data import and transformation
5.2.2 Simulations execution
5.2.3 Performance assessment
5.3.1 Data import and transformation
5.3.2 Hindcasts execution
5.3.3 Error correction
5.3.4 Performance assessment
6.1 From GR4H to wflow gr4
6.2.1 Overall statistical model performance assessment
6.2.2 Model dynamics performance assessment: peakflows and low flows
6.3.1 Ensemble skill analysis
6.3.2 Ensemble distribution analysis
7.1 HBV-96 model performance
7.2 wflow gr4 model performance
7.3 wflow hbv model performance
7.4 wflow sbm model performance
7.5 Simulated peaks time lag error
7.6 The case of the Chiers in Carignan
8 Conclusions and recommendations
A Rhine river system
B HBV-96 model
C GR4J model
D OpenStreams models
E The forecasting platform
F Data import and transformation
G Calibration of the wflow gr4 model
H Simulations results