Implications of event-based loss model structure on simulating large floods

Abstract

Event-based rainfall-runoff models used for design flood estimation and operational flood forecasting commonly represent losses (the amount of rainfall which does not appear as runoff) using lumped, parsimonious models. In gauged catchments, these loss values are typically estimated by calibrating models to observed flood events. Studies which compare loss models tend to focus on which model best reproduces observed events, however, flood modelling often requires estimating events larger than those used for calibration. It is thus of interest to examine how loss models perform over a range of flood magnitudes, particularly larger floods. Here, we evaluate the efficacy of different loss models by exploring the dependence between event magnitude and model performance. We assess four loss models commonly used in flood modelling: Initial Loss Continuing Loss (ILCL), Initial Loss Proportional Loss (ILPL), Soil Conservation Service Curve Number (SCSCN) and Probability Distributed Model (PDM).The four loss models were incorporated into a lumped rainfall runoff model that was calibrated to observed storm events for 35 small, rural Australian catchments using a total of 876 unique storm events. In calibration, ILPL was found to best reproduce observed hydrographs, however the difference in performance between the four different models in terms of NSE was small. Dependence between event magnitude and model performance was explored by re-simulating observed events with the loss model parameters fixed to represent typical catchment conditions. These typical parameters were estimated from the median of the observed parameter ranges in each catchment with the largest five events excluded so that the impact of extrapolating model parameters to events larger than which they were calibrated could be explored. When these typical parameters were applied to large flood events the ILCL model displayed minimal dependence between model performance and event magnitude. In contrast, ILPL, SCSCN and PDM all clearly displayed an underprediction bias which increased with increasing event size. We explain this behaviour through the loss model structure of ILCL where ongoing loss (the loss which occurs after runoff begins) is constant and independent of rainfall fvmagnitude. In contrast, the three remaining models represent ongoing loss as a proportion of rainfall. As the difference in magnitude between small and large floods increases with climate change, extrapolation of flood models is likely to become a more pertinent issue in the future. We conclude that loss models which are independent of event magnitude are better suited to flood modelling applications where extrapolation is required.