Flow duration curve prediction in ungauged basins: a model intercomparison study

The flow duration curve (FDC) has long been used in water resources research and practice.  We compared three approaches to FDC estimation in ungauged basins, ranging in model complexity and richness of input data.  FDCs were estimated by 1) assuming daily runoff is log-normally distributed and predicting distribution parameters from catchment descriptors, 2) ensemble averaging of nearest and most physically similar gauged neighbours, and 3) neural network rainfall runoff modelling. When evaluated on a hydrologically diverse sample of 712 catchments around British Columbia, Canada, we found the more complex neural network model provided little performance advantage over a simpler nearest-neighbour ensemble approach, and inter-model ensembles yielded equal or better performance than individual components.  Models were evaluated by four performance measures to highlight different notions of dissimilarity expressed by conventional residual error based metrics versus an information measure, the Kullback-Leibler divergence.