HR: 15:10h
AN: H53H-07    [Abstracts]
TI: Ensemble modeling of flows in ungaged catchments
AU: McIntyre, N
EM: n.mcintyre@imperial.ac.uk
AF: Imperial College London, Department of Civil and Environmental Engineering, South Kensington Campus, London, SW72AZ United Kingdom
AU: * Wheater, H
EM: h.wheater@imperial.ac.uk
AF: Imperial College London, Department of Civil and Environmental Engineering, South Kensington Campus, London, SW72AZ United Kingdom
AU: Lee, H
EM: h.lee@imperial.ac.uk
AF: Imperial College London, Department of Civil and Environmental Engineering, South Kensington Campus, London, SW72AZ United Kingdom
AU: Young, A
EM: ary@ceh.ac.uk
AF: Centre for Ecology and Hydrology, Wallingford, Oxfordshire, OX10 8BB United Kingdom
AU: Wagener, T
EM: thorsten@engr.psu.edu
AF: The Pennsylvania State University, Department of Civil and Environmental Engineering, University Park, PA PA16802 United States
AB: The established approach to rainfall-runoff model regionalisation is regression of model parameters (MPs) against numeric catchment descriptors (CDs). We argue that, due to its fundamental limitations, further refinement of the regression method is not the optimum way forward, and we introduce an alternative method based on weighed averaging and ensemble modelling. The new method consists of the following basic steps: 1) A sample of successful models is identified for each of a number of `donor' gaged catchments. 2) Each model is assigned a weight based on how well it has performed. 3) This weight is updated based on the similarity of the associated catchment to the `target' ungaged catchment. 4) All models with non-zero weight are applied to the target catchment, to produce an ensemble time-series and a weighted average prediction. The theoretical advantage is that MP interactions are not neglected or linearized to facilitate regression. The practical attraction is the ease with which all sources of uncertainty (e.g. data, CD, equifinality, model structure) can be integrated into the pool of models and the weighting scheme. A case study of daily data from 127 non-urban UK catchments is presented. A single conceptual model structure is used (a five-parameter probability distributed model) so that, in this case, differences in models are defined only by the MP sets. Each of the 127 catchments is, in turn, considered to be ungaged, so that candidate models can be drawn from up to 126 donor catchments. Relative weights are proportional to a quantitative measure of donor-target catchment similarity. Various schemes for defining catchment similarity are applied, based on CDs relating mainly to soil type, catchment size and climate. Using the models of the ten most similar catchments provided the best weighted average simulations, both in terms of NSE and a low-flow objective function. Using this scheme, in 90% of low-permeability catchments the prediction NSE was within 0.05 of that achieved using the locally calibrated model. Performance was poorer when fewer or more donor catchments were used, and was a little poorer in high-permeability catchments. The results were not sensitive to how many models were drawn from each donor catchment. The method out-performed regression-based schemes; the improvement was small for NSE but was considerable for the low-flow objective function. Comparisons of the prior and posterior ensembles demonstrates the reduction in uncertainty permitted by the method. The ability of the ensemble to envelope the observed flood peaks was inconsistent - options for improving robustness include integration of alternative model structures and rainfall input realisations. In summary, the ensemble modelling and weighted averaging method is a theoretical improvement on established regionalisation schemes, and initial results are promising. The method provides an elegant basis for transfer of information between catchments and comprehensive analysis of uncertainty for PUBs.
DE: 1805 Computational hydrology
DE: 1873 Uncertainty assessment (3275)
DE: 1874 Ungaged basins
DE: 1879 Watershed
SC: Hydrology [H]
MN: Fall Meeting 2005