Does a large number of parameters enhance model performance? Comparative assessment of common catchment model structures on 429 catchments

Abstract

Hydrological models must be reliable and robust as these qualities influence all applications based on model output. Previous studies on conceptual rainfall–runoff models have shown that one of the root causes of their output uncertainty is model over-parameterisation. The problem of poorly defined parameters has attracted much attention but has not yet been satisfactorily solved. We believe that the most fruitful way forward is to improve the structures where these parameters act. The main objective of this paper is to examine the role of complexity in hydrological models by studying the relation between the number of optimised parameters and model performance. An extensive comparative performance assessment of the structures of 19 daily lumped models was carried out on 429 catchments, mostly in France but also in the United States, Australia, the Ivory Coast and Brazil. Bulk treatment of the data showed that the complex models outperform the simple ones in calibration mode but not in verification mode. We argue that the main reason why complex models lack stability is that the structure, i.e. the way components are organised, is not suited to extracting information available in hydrological time-series. An inadequate complexity typically results in model over-parameterisation and parameter uncertainty. Although complexity has been used as a response to the challenge of predicting the hydrological effects of environmental changes, this study suggests that such models may have been developed with excessive confidence and that they could face difficulties of parameter estimation and structure validation when confronted with hydro-meteorological time-series. This comparative study indicates that some parsimonious models can yield promising results and should be further developed, although they are not able to tackle all types of problems, which would be the case if their complexity were ideally adapted.

Introduction

After more than 30 years of ‘classical’ hydrological model development, there is an increasing trend in hydrology today to use new tools, that provide distributed information of catchment characteristics. It is tempting to introduce newly available information into increasingly complex catchment models, when one is faced with the task of accurately representing the inherent complexity of real systems. However, although this approach might be useful in terms of knowledge of the processes, it has limitations when applied in an operational context. Conversely, simple catchment models that lump catchment heterogeneities and represent the transformation of precipitation into streamflow, conceptually or empirically, are generally easy-to-use tools with low data requirements. In spite of the crude approximation resulting from their lumped nature and simple structure, such models have proved efficient in many case studies reported in the literature and they are undoubtedly useful for engineers and water managers. Daily conceptual models (of interest here) have successfully fulfilled most operational requirements, such as flood frequency assessment (Cameron et al., 1999, Uhlenbrook et al., 1999), reservoir management (Yang et al., 1995), flood and drought forecasting (see, e.g. Yang and Michel, 2000). Although these models have, up to now, been unable to predict the change in streamflow caused by land-use changes, they are apt at objectively detecting such changes (see, e.g. Lavabre et al., 1993, Lørup et al., 1998).

Since the early 1960s, many hydrologists have concentrated their efforts on designing rainfall–runoff models. Because they have attracted such widespread interest and so many different ones are being developed, the need for comparative studies was expressed quite early on (WMO, 1975) to evaluate the ability of models to simulate streamflow and to provide guidelines for end-users. Model assessment however is a tricky exercise and the conclusions of such experiments generally depend on the methodology of the comparisons and the characteristics of the test catchments. This is all the more true as models are mostly tested on limited numbers of catchments. Due to increasing computational capacity, it is possible today to extensively test simple models against great many catchments, under a wide range of climate conditions. However it would be naı̈ve to believe that, one might, from a broad-based comparison, identify one single outstanding model according to any assessment criterion that would satisfy all water stakeholders (Leviandier, 1988). Nevertheless, we believe that the robustness and reliability of a model resides primarily in its ability to perform under as varied a set of hydrological conditions as possible. From this standpoint, comparative assessments serve to highlight strengths and weaknesses of modelling approaches of various complexity.