On the evaluation of grid and grid-to-grid rainfall-runoff models and their differences with physical benchmarks

Abstract

In the first part of our study, we demonstrated how a simple physical benchmark model can be used to assess assumptions of the conceptual models, based on a lumped Probability Distributed Model (PDM) formulated by Lamb (1999). In this second part, we extend the scope of our study to distributed models, which aim to represent the spatial variability of model's elements (e.g. input precipitation, soil moisture levels, flow components etc.). For demonstration purposes, we assess the assumptions of the Grid and Grid-to-Grid models, commonly used for flood real-time forecasting in the UK. While the distributed character of these models is conceptually closer to the physical model, we demonstrate that its exact implementation leads to many qualitative and quantitative differences in the model behaviour. For example, we show that the main assumption, namely that the speed of surface and subsurface flow is constant, causes the Grid-to-Grid model to significantly misrepresent scenarios with no rainfall, leading to too fast river flow decay, and scenarios with upstream rainfall, failing to capture characteristic flash flood formation. We argue that this analytical approach of finding fundamental differences between models may help us to develop more theoretically-justified rainfall-runoff models, e.g. models that can better handle the two aforementioned scenarios and other scenarios in which the spatial dependence is crucial to properly represent the catchment dynamics.