Identification of rainfall-runoff models for operational applications

The problem of selecting appropriate objective functions for the identification of a lumped conceptual rainfall runoff model is investigated, focusing on the value of the model in an operational setting. A probability-distributed soil moisture model is coupled with a linear parallel routing scheme, and conditioned on rainfall-runoff observations from three catchments in the southeast of England. Using an abstraction control problem, which requires accurate simulation of the intermediate flow range, it is shown that using the traditional RMSE fit criterion, produces operationally sub-optimal predictions. This is true in the identification period, when applied to a testing period, and to proxy catchment data. Using a second case study of the Leaf River in Mississippi (USA), where the focus changes to predicting flood peaks over a specified threshold, also suggests that the relevant flood threshold should govern the objective function choice. It is concluded that, due to limitations in the structure of the employed model, it would be counter-productive to try to achieve a good all-round representation of the rainfall-runoff processes, and that a more empirical approach to identification may be preferred for specific forecasting problems. This leaves us with the question of how far hydrological realism should be sacrificed in favour of purpose-driven objective functions.