EGU2020-20280
https://doi.org/10.5194/egusphere-egu2020-20280
EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

How to Tailor my Process-based Hydrological Model? Dynamic Identifiability Analysis of Flexible Model Structures

Axel Bronstert1, Tobias Pilz1,2, Till Francke1, and Gabriele Baroni3
Axel Bronstert et al.
  • 1University of Potsdam, Institute of Environmental Sciences, Chair for Hydrology and Climatology, Potsdam-Golm, Germany
  • 2Potsdam Institute for Climate Impact Research, Water Research Group, Potsdam, Germany (tobias.pilz@pik-potsdam.de)
  • 3University of Bologna, Dipartimento di Scienze e Tecnologie Agro-Alimentari, Bologna, Italy

In the field of hydrological modeling, many alternative mathematical representations of natural processes exist. To choose specific process formulations when building a hydrological model is therefore associated with a high degree of ambiguity and subjectivity. Identifiability analysis may provide guidance by constraining the a priori range of alternatives based on observations. In this work, a flexible simulation environment is used to build a process-based hydrological model with alternative process representations, numerical integration schemes, and model parametrizations in an integrated manner. The flexible simulation environment is coupled with an approach for dynamic identifiability analysis. The objective is to investigate the applicability of the coupled framework to identify the most adequate model structure. It turned out that identifiability of model structure varies in space and time, driven by the meteorological and hydrological characteristics of the study area. Moreover, the most accurate numerical solver is often not the best performing solution. This is possibly influenced by correlation and compensation effects among process representation, numerical solver, and parametrization. Overall, the proposed coupled framework proved to be applicable for the identification of adequate process-based model structures and is therefore a useful diagnostic tool for model building and hypotheses testing.

How to cite: Bronstert, A., Pilz, T., Francke, T., and Baroni, G.: How to Tailor my Process-based Hydrological Model? Dynamic Identifiability Analysis of Flexible Model Structures, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-20280, https://doi.org/10.5194/egusphere-egu2020-20280, 2020

Display materials

Display file

Comments on the display material

AC: Author Comment | CC: Community Comment | Report abuse

Display material version 1 – uploaded on 29 Apr 2020
  • CC1: Comment on EGU2020-20280, Björn Guse, 03 May 2020

    Thanks for this very interesting study. I have some questions.

    Page 16:

    Could you explain why the conceptual approach of runoff concentration performs better during peak flows?

    How is hereby the impact of precipitation uncertainty (during peak flows)?

    Does this result mean that our hydrological understanding of how to model the runoff concentration during peak flows is not sufficient (to be implemented physically-based)?

    Page 17:

    Many models allow a selection of an ETP method. One of your results is that ETP methods are highly identifiable. Thus, would you agree that a justification for the selection of an ETP method should be generally given in model studies?

    • CC3: Reply to CC1, Tobias Pilz, 05 May 2020

      Dear Björn,

      thanks for your interest in our work.

      Could you explain why the conceptual approach of runoff concentration performs better during peak flows?

      This is a good question. A possible explanation: The conceptual approach is a parallel storage model with three storages (reflecting surface flow, interflow, baseflow) and associated retention constants, which are calibrated. It possibly performs better during high flows because it incorporates calibration parameters and more degrees of freedom. On the other hand, the physics-based approach seems to work better for low (peak) flow conditions (see the two small peaks in 2015). The physics-based approach is incorporated in the original WASA-SED model, which has been designed specifically for dryland environments and low flow conditions. It is based on the lateral re-distribution of surface and interflow along representative hillslopes without further retention parameters. As it seems, under such conditions, the conceptual parallel storage model is no plausible representation, even with the calibration parameters. Under high flows the physics-based WASA-SED approach is not working well because, as I assume, high rainfall (and consequently flow) events cause problems in the soil water module (numerical problems, leading to unrealistic model states and, consequently, flows). It would be interesting to see what would happen if one (significantly) increases temporal and spatial (soil horizons) resolution. However, this would cause such a computational burden that massive Monte Carlos analyses, as presented, are not possible. 

      How is hereby the impact of precipitation uncertainty (during peak flows)?

      The impact of precipitation uncertainty is certainly, at least in some cases / events, significant. However, I don't think this can explain the whole pattern of runoff concentration identifiability. I would think the explanation given above is more likely.

      Does this result mean that our hydrological understanding of how to model the runoff concentration during peak flows is not sufficient (to be implemented physically-based)?

      As I already suggested in my answer to the first question, I believe the problem is rather related to numerical problems. During model development I conducted some test runs of the soil water routines with varying temporal and spatial resolution and different numerical solvers and it turned out that with decreasing resolution and solvers with low order of accuracy (e.g. explicit Euler) the model produces unreasonable model states and flows, especially in case of high precipitation input. Calibration parameters are, to some extent, able to compensate for such effects. However, this is just an assumption, would be interesting to find out.


      Many models allow a selection of an ETP method. One of your results is that ETP methods are highly identifiable. Thus, would you agree that a justification for the selection of an ETP method should be generally given in model studies?

      Evapotranspiration is certainly an important process, especially in cases where it accounts for a large part of the water balance, such as in dryland regions. I think that, ideally, one should choose a model according to how well it incorporates the most important processes of a study area (not just ET). Especially if the model flexible in a way that it contains multiple choices for a specific process representation. This should also be discussed in publications and, to my experience, reviewers often ask for it if this is not adequately done. However, my personal perception also is the following: Model selection is in practice often based on convenience, i.e. modellers use their favourite model which they already know best, even when it is not well suited for a particular catchment. The use of a model is then solely justified by calibrating the model to achieve high Nash-Sutcliffe values, even though internally the model produces unrealistic behaviour which is simply ignored as long as the NSE is 0.9 (which is not even that difficult when you apply the model in catchments with strong seasonality). This, in my opinion, should be much more critically reviewed in model studies.

      Kind regards,

      Tobias

  • CC2: Comment on EGU2020-20280, Xiangqian Zhou, 04 May 2020

    thanks for your interesting study. My question are :

    Page 6:

    What are the two realisations of runoff concentration processes ? 

    How does the model partition the discharge ?

    Page 11:

    Could be falling limbs of discharge events not well matched attribute to the realisations of runoff concentration processes?

    • CC4: Reply to CC2, Tobias Pilz, 05 May 2020

      Dear Xiangqian,

      thanks for your questions.

      What are the two realisations of runoff concentration processes ?
      How does the model partition the discharge ?

      The conceptual approach is a parallel storage model with three storages (reflecting surface flow, interflow, baseflow) and associated retention constants, which are calibration parameters. The physics-based approach is the lateral re-distribution of surface and interflow along representative hillslopes (from upslope to downslope locations, randomly distributed to the soil-vegetation components, and finally into the river) without further retention parameters. That means, runoff concentration is implicitly incorporated in the complex landscape discretisation scheme of the model (slide 5). Water flows include surface runoff and interflow (also baseflow, but groundwater is represented only in a simplified way). Soil water movement (interflow) is explicitly modelled. Besides, water can exfiltrate from and re-infiltration into the soil and a small share is bypassed from upslope locations directly into the river to represent preferential flow paths. The approach is described by Güntner and Bronstert, 2004, https://doi.org/10.1016/j.jhydrol.2004.04.008 (though not explicitly under the term runoff concentration; look at the lateral redistribution scheme).

      Could be falling limbs of discharge events not well matched attribute to the realisations of runoff concentration processes?

      It might be that it can be attributed to the parametrization of the storage constants, as the red line in the figure of slide 11 was produced with the conceptual runoff concentration approach (not shown in the presentation). Still the red line reflects the best discharge simulation in terms of RMSE and NSE, possibly as absolute differences are comparably small and therefore are not much penalised by RMSE and NSE, which are based on squared residuals and put more emphasize on large values. However, we haven't analysed the simulated discharge dynamics in detail.

      Kind regards,

      Tobias