Process consistency in models: The importance of system signatures, expert knowledge, and process complexity

Abstract

Hydrological models frequently suffer from limited predictive power despite adequate calibration performances. This can indicate insufficient representations of the underlying processes. Thus, ways are sought to increase model consistency while satisfying the contrasting priorities of increased model complexity and limited equifinality. In this study, the value of a systematic use of hydrological signatures and expert knowledge for increasing model consistency was tested. It was found that a simple conceptual model, constrained by four calibration objective functions, was able to adequately reproduce the hydrograph in the calibration period. The model, however, could not reproduce a suite of hydrological signatures, indicating a lack of model consistency. Subsequently, testing 11 models, model complexity was increased in a stepwise way and counter-balanced by ''prior constraints,'' inferred from expert knowledge to ensure a model which behaves well with respect to the modeler's perception of the system. We showed that, in spite of unchanged calibration performance, the most complex model setup exhibited increased performance in the independent test period and skill to better reproduce all tested signatures, indicating a better system representation. The results suggest that a model may be inadequate despite good performance with respect to multiple calibration objectives and that increasing model complexity, if counter-balanced by prior constraints, can significantly increase predictive performance of a model and its skill to reproduce hydrological signatures. The results strongly illustrate the need to balance automated model calibration with a more expert-knowledge-driven strategy of constraining models.

A constraint-based search algorithm for parameter identification of environmental models

“…Result of the recent study by Gao et al (2014) had shown that a more complex, but simultaneously constrained and calibrated model, can provide better predictions of extreme events outside of the calibration period as well as capturing the flow duration curve in nested catchments for the large scale Heihe basin in China. Meanwhile, Hrachowitz et al (2014) reported similar results for a small-scale catchment in France, using 11 different model structures and 20 different runoff signatures. In these studies, the introduction of constraints was found to be crucial for ensuring better model performance, particularly outside of a calibration period.…”

“…The multi-objective approach seeks to identify parameter sets that simultaneously provide optimal performance for different aspects of system response (Gupta et al, 1998;Boyle et al, 2000Boyle et al, , 2001. This can include constraining the model to reproduce multiple system fluxes and state variables such as runoff, evaporation, groundwater levels or tracer concentrations (e.g., Gupta et al, 1999;Bastidas et al, 1999;Freer et al, 2002;McDonnell, 2002, 2013;Khu and Madsen, 2005;Fenicia et al, 2008;Winsemius et al, 2008;Birkel et al, 2011;Hrachowitz et al, 2013).…”

“…Expert knowledge is brought to bear implicitly, by the prior selection of the model and the specification of parameter ranges that define the prior parameter space. Recently, several studies have tested strategies that relate the parameters of CRR models to catchment physical characteristics (Koren et al, 2000(Koren et al, , 2003Anderson et al, 2006;Yadav et al, 2007;Pokhrel et al, 2008Pokhrel et al, , 2012Kling and Gupta, 2009;Hrachowitz et al, 2013). The general picture that emerges from these studies is that exploiting expert knowledge (by somehow imposing more rigorous constraints…”

Glacio-hydrological melt and runoff modelling: a limits of acceptability framework for model selection

“…7a) were used to define 21 different river discharge signatures that cover a range of temporal scales and flow magnitudes. The majority of these signatures were based on previous studies (Coxon et al, 2014;Yilmaz et al, 2008;Westerberg et al, 2016;Shafii and Tolson, 2015;Hrachowitz et al, 2014;Schaefli, 2016;Viglione et al, 2013;Euser et al, 2013;Garavaglia et al, 2017;Yadav et al, 2007;Casper et al, 2012; (99-66% flow exceedance), high flow section (15-5% flow exceedance) and highest flow section (5-0.5% flow exceedance).…”

“…autocorrelation) of flows. They have shown to have more discrimination power than traditional error metrics (Hrachowitz et al, 2014;Shafii and Tolson, 2015;Euser et al, 2013;Schaefli, 2016) and, importantly, it is also possible to take account of their information content (i.e. their uncertainty) so that decisions about model appropriateness can 10 be made within the uncertainties of observation data used to evaluate the model.…”