Ensemble predictions of runoff in ungauged catchments

TLDR

The study argues that an ensemble‑based regionalization of conceptual rainfall‑runoff models can improve upon the traditional regression of parameter values against catchment descriptors. The authors introduce this ensemble approach and discuss options for addressing remaining issues. They use daily data from 127 UK catchments to build a probability‑distributed model structure, test alternative prior and posterior likelihood schemes, and select parameter sets from gauged catchments to evaluate ungauged predictions. The method delivers the best results when employing the ten most similar gauged catchments, outperforming regression especially for low flows, but the ensemble uncertainty estimates are unreliable and often miss peak flows.

Abstract

A new approach to regionalization of conceptual rainfall‐runoff models is presented on the basis of ensemble modeling and model averaging. It is argued that in principle, this approach represents an improvement on the established procedure of regressing parameter values against numeric catchment descriptors. Using daily data from 127 catchments in the United Kingdom, alternative schemes for defining prior and posterior likelihoods of candidate models are tested in terms of accuracy of ungauged catchment predictions. A probability distributed model structure is used, and alternative parameter sets are identified using data from each of a number of gauged catchments. Using the models of the 10 gauged catchments most similar to the ungauged catchment provides generally the best results and performs significantly better than the regression method, especially for predicting low flows. The ensemble of candidate models provides an indication of uncertainty in ungauged catchment predictions, although this is not a robust estimate of possible flow ranges, and frequently fails to encompass flow peaks. Options for developing the new method to resolve these problems are discussed.