Elsevier

Journal of Hydrology

Volume 411, Issues 1–2, 6 December 2011, Pages 66-76
Journal of Hydrology

A downward structural sensitivity analysis of hydrological models to improve low-flow simulation

https://doi.org/10.1016/j.jhydrol.2011.09.034Get rights and content

Summary

Better simulation and earlier prediction of river low flows are needed for improved water management. Here, a top–down structural analysis to improve a hydrological model in a low-flow simulation perspective is presented. Starting from a simple but efficient rainfall–runoff model (GR5J), we analyse the sensitivity of low-flow simulations to progressive modifications of the model’s structure. These modifications correspond to the introduction of more complex routing schemes and/or the addition of simple representations of groundwater–surface water exchanges. In these tests, we wished to improve low-flow simulation while avoiding performance losses in high-flow conditions, i.e. keeping a general model.

In a typical downward modelling perspective, over 60 versions of the model were tested on a large set of French catchments corresponding to various low-flow conditions, and performance was evaluated using criteria emphasising errors in low-flow conditions. The results indicate that several best performing structures yielded quite similar levels of efficiency. The addition of a new flow component to the routing part of the model yielded the most significant improvement. In spite of the close performance of several model structures, we conclude by proposing a modified model version of GR5J with a single additional parameter.

Highlights

► We detail a downward approach for improving a model for low-flow simulation. ► We compare several model structures on a large set of catchments and propose an improved model version. ► We compare the model with existing ones, with satisfactory results.

Introduction

The occurrence of low flows is perhaps less spectacular than high flows, but low-flow consequences can be as costly, because they correspond to crucial periods in the functioning of both ecological and water management systems. For example, the cost of damage caused by the drought events in the years 1988–1989 in the United States was approximately US$40 billion, whereas the cost of the 1993 flood event was US$18–20 billion (Demuth, 2005). Thus, we consider that the simulation and advanced prediction of river low flows is an important challenge to improve low-flow management, both in the present climate and under the projected climate changes, which may well result in an increase in the occurrence of low-flow events (see e.g. Boé et al., 2009, Feyen and Dankers, 2009).

While a variety of lumped rainfall–runoff models are available to simulate streamflow irrespective of the flow conditions (see e.g. Singh and Frevert, 2002a, Singh and Frevert, 2002b), only a limited number of modelling studies focus on low-flow simulation. This study aims at identifying a generic model structure for improved low-flow simulation. Note that given the complexity of hydrological processes and the specificities of each catchment, some modellers have argued that model structures should be catchment-specific (e.g. Fenicia et al., 2008). However, we believe that before identifying catchment-specific models, the best possible general model that would include the representation of most of the dominant processes at work on catchments should be identified. This is the approach followed in this paper.

To identify the general model structures that represent catchment behaviour, we followed a downward approach: a lumped representation of the catchment was used, in which only the main features of catchment hydrological behaviour are represented. This means that we did not attempt to build an explicit physical representation of the system but instead attempted to find the building blocks of the model that maximised modelling efficiency. The tests reported herein can be considered a structural sensitivity analysis. Some studies highlight the usefulness of sensitivity analysis for the improvement of hydrological models (see e.g. Andréassian et al., 2001, Oudin et al., 2006b, Tang et al., 2007, Bahremand and De Smedt, 2008, Ruelland et al., 2008). Other studies used sensitivity analysis to better understand model behaviour with respect to inputs such as precipitation and potential evapotranspiration (Oudin et al., 2005a, Oudin et al., 2005b, Xu et al., 2006, Meselhe et al., 2009). Here we will focus on the sensitivity of low-flow simulation to the change in the components of the model structure responsible for low-flow simulation.

The main objective of this article is to analyse the extent to which a downward sensitivity analysis can help identify ways to improve low-flow simulation, while keeping the hydrological coherence in simulating the other parts of the flow regime. The downward search starts from a robust and parsimonious model structure. Then we will analyse how sensitive low-flow simulations are to the formulation of the model structure. This is done in trial-and-error mode, by testing many alternative model structures on a large set of catchments representing various physical and hydrometeorological conditions. The best candidate towards which our search converged is finally assessed in comparison with other model structures available in the literature.

The number of catchment modelling studies focusing on low-flow simulation using hydrological models is quite limited. One of the major problems with low-flow simulation is to account for surface water–groundwater interactions. During low-flow periods, water exchanges occur through the stream bed: the river may be fed by groundwater or, conversely, it may leak to feed the aquifer. Therefore, groundwater significantly influences low flows. A few studies that investigated these issues can be mentioned here. Fleckenstein et al. (2006) clearly mentioned the river–aquifer interactions and the significance of groundwater contribution during low-flow periods. Herron and Croke (2009) noted the improvement of lumped model predictions with the incorporation of groundwater exchange functions. The conclusions by Anderson et al., 2004, Hughes, 2004 also suggest that the model simulation efficiency can be improved by the addition of functions which represent the interaction between channel and aquifer flows. This is clearly shown in the study by Le Moine et al. (2007), who tested several options to account for inter-catchment groundwater flows using two rainfall–runoff models. Their results indicate that explicitly accounting for these groundwater fluxes significantly improves modelling efficiency.

Along with groundwater exchange functions, additional stores in the routing module can also enhance model performance, especially in the case of delayed flows (Wagener et al., 2004, Mathevet, 2005). Lang, 2007, Lang et al., 2008 analysed the performance of lumped models with respect to the addition of routing stores (to account for different water pathways underground) in an existing structure. Their study showed that some improvement can be achieved in the low-flow simulation, although they conclude that further work would be needed to improve lumped models for low-flow simulation. In a recent study, Kim et al. (2011) used the IHACRES-3S (3 Storage) model to evaluate the low-flow simulation together with the integration of base flow. The results showed a slight improvement in the model’s performance, but they concluded that further studies are needed to obtain better low-flow simulation results. Last, Staudinger et al. (2011) analysed the sensitivity of recession simulation to various storage configurations on a snow dominated catchment in Norway within the FUSE framework. They conclude that the structural sensitivity is different in the winter and summer seasons, but that tests on a larger set of catchments are needed to get more general conclusions.

This article presents the end result of a long downward sensitivity analysis process that led to proposing an improved version of the GR4J catchment model (Perrin et al., 2003). Although our aim was to improve low-flow simulation specifically, we intended to find a generic solution, i.e. one that would improve low-flow representation without affecting the representation of high flows. This study builds on the previous studies by Mathevet, 2005, Le Moine, 2008 who have already conducted tests to modify the existing model structures to improve modelling efficiency on a wide variety of catchments.

The next section discusses the data set and testing methodology. Then the results are presented and discussed before the concluding remarks.

Section snippets

Data set, models and methodology

This section presents the data set, models and testing methodology used for the analysis.

Results and discussion

In the following, we present the main results and discuss how sensitive low-flow simulations are to model formulation, following the modifications presented above. The selected versions of GR5J and their formulations are briefly presented in Table 1. As the number of modifications is almost infinite, we chose to present only a few of them to answer a number of simple questions that may arise when discussing the model’s structure. Although these questions are sometimes interrelated, they are

Conclusions

Improving the low-flow simulation ability without impacting the high-flow simulation ability: this was one of our objectives in this study. We chose to proceed by trial and error, as recommended by Nash and Sutcliffe, 1970, Michel et al., 2006. Working on a large set of catchments proved to be a good way to prevent undue complexity in the proposed modified versions of the model. Here we started from the simple GR5J model and tested a number of modified versions, some having higher performance

Acknowledgements

We are very grateful to Météo–France for providing meteorological data and SCHAPI for providing flow data. We also wish to thank Domaine d’Intérêt Majeur Agrosciences, territoires, écologie, alimentation (DIM ASTREA) of Région Ile–de–France and Office National de l’Eau et des Milieux Aquatiques (ONEMA) for their financial support to conduct this study.

References (52)

  • C. Xu et al.

    Evaluation of seasonal and spatial variations of lumped water balance model sensitivity to precipitation data errors

    J. Hydrol.

    (2006)
  • M.L. Anderson et al.

    Modeling low flows on the Cosumnes River

    J. Hydrol. Eng.

    (2004)
  • V. Andréassian et al.

    Introduction and synthesis: why should hydrologists work on a large number of basin data sets? Large sample experiments for hydrological model parameterisation: results of the model parameter experiment – MOPEX

    IAHS Publ.

    (2006)
  • V. Andréassian et al.

    Editorial – The court of miracles of hydrology: can failure stories contribute to hydrological science?

    Hydrol. Sci. J.

    (2010)
  • A. Bahremand et al.

    Distributed hydrological modeling and sensitivity analysis in Torysa Watershed, Slovakia

    Water Resour. Manage

    (2008)
  • S. Bergström et al.

    Development of a conceptual deterministic rainfall–runoff model

    Nord. Hydrol.

    (1973)
  • K. Beven et al.

    A physically based, variable contributing area model of basin hydrology

    J. Hydrol. Sci.

    (1979)
  • J. Boé et al.

    Projected changes in components of the hydrological cycle in French river basins during the 21st century

    Water Resour. Res.

    (2009)
  • B. Davison et al.

    Low-flows in deterministic modelling: a brief review

    Can. Water Resour. Assoc.

    (2008)
  • S. Demuth

    Low flows and droughts – a European prospective

    IAHS Newsletter

    (2005)
  • Edijatno et al.

    Un modèle pluie-débit journalier à trois paramètres

    La Houille Blanche

    (1989)
  • Edijatno et al.

    GR3J: a daily watershed model with three free parameters

    Hydrol. Sci. J.

    (1999)
  • F. Fenicia et al.

    Understanding catchment behavior through 664 stepwise model concept improvement

    Water Resour. Res.

    (2008)
  • L. Feyen et al.

    Impact of global warming on streamflow drought in Europe

    J. Geophys. Res.

    (2009)
  • J.H. Fleckenstein et al.

    River–aquifer interactions, geologic heterogeneity, and low flow management

    Ground Water

    (2006)
  • V. Fortin et al.

    Le modéle hydrologique MOHYSE

    (2007)
  • Cited by (131)

    • Vulnerability of water resource management to climate change: Application to a Pyrenean valley

      2022, Journal of Hydrology: Regional Studies
      Citation Excerpt :

      However, the data length for SB2 is from 07/2014 to 12/2018. The conceptual lumped rainfall-runoff model GR6J, developed to improve low-flow simulation based on the extensively used GR4J model (Perrin et al., 2003) for French basins, was adopted to simulate the daily inflow into the reservoirs of the water system (Pushpalatha et al., 2011). The GR6J model was largely applied in studies including reconstruction of low-flow events (e.g., Caillouet et al., 2017), climate change projections (e.g., Givati et al., 2019), and streamflow forecasts (e.g., Crochemore et al., 2016).

    View all citing articles on Scopus
    View full text