Improving hydrological multi-model prediction by elimination of double counting in the ensemble

• Published in: Journal of Hydrology Vol. 58; no. 2; pp. 81 - 104
• Main Authors: Singh, Shailesh Kumar, Pahlow, Markus, Duan, Qingyun, Griffiths, George
• Format: Journal Article
• Language: English
• Published: Wellington New Zealand Hydrological Society 01.12.2019
• Subjects: Accuracy; Case studies; Catchment area; Computing costs; Costs; Depth perception; Histograms; Hydrologic models; Hydrologic processes; Hydrology; Mathematical models; Model accuracy; Multiscale modeling; Nautical charts; Precipitation; Predictions; Probability theory; Representations; River networks; Simulation; Simulation methods; Skills; Statistical methods; Stream discharge; Stream flow; Streamflow; Uncertainty; New Zealand
• ISSN: 0022-1708; 2463-3933; 2463-3933

Hydrological models are a mathematical representation of heterogeneous and non-linear hydrological processes. In the past, a great many simple to complex hydrological models have been developed, but none of these models is superior to the others for all types of practical applications. These hydrological model alternatives have different strengths in representing and capturing complex natural hydrological processes. Yet generally a single hydrological model is used in practice, which may represent certain processes of the catchment well and may be less adequate for others. Moreover, the use of a single hydrological model is restrictive, as the conceptual uncertainty associated with the model structure cannot be identified and quantified. To overcome these issues, the multi-model ensemble approach has recently been applied more commonly to take advantage of the diverse skills of different models. In this study, the multi-model ensemble approach to account for model structure uncertainty is employed to improve hydrological model prediction. While a certain hydrological model may represent particular processes or (extreme) events better than another, two distinct models may represent these processes or events with comparable accuracy. If members of a hydrological ensemble model capture the same process and if they are similar in process representation, then these members will not supply any additional information for prediction and therefore will not improve the accuracy. Hence, by identifying similar models, there is potential to increase the reliability of hydrological ensemble predictions and to reduce computing costs without reducing accuracy. In this study a methodology is presented to identify similar models. The methodology is applied and tested for the Tuapiro catchment in New Zealand. A range of verification statistics are computed to ascertain the validity of the approach. Overall, the multi-model ensemble-based hydrological prediction where non-informative members have been removed is shown to not compromise prediction accuracy. For the case study streamflow prediction an increased flatness of the rank histogram, insignificant changes in the continuous rank probability score, and improved accuracy in terms of Nash-Sutcliffe coefficient, Kling-Gupta efficiency and Root Mean Square Error were found, at lower computing costs.