Coupling mesoscale numerical weather prediction models with conceptual rainfall-runoff models yields a potentially valuable hydrometeorological forecasting tool. However, there are a number of sources of uncertainty in this forecasting framework, including uncertainty in model parameters and initial states as well as in measured rainfall and runoff. Another substantial source of uncertainty is the quantitative precipitation forecast (QPF), which serves as the input for the rainfall-runoff model.

This paper addresses some of these sources of uncertainty in a probabilistic forecasting framework. A multi-model mesoscale ensemble is used to address the QPF component, with model uncertainty handled through a Bayesian Model Averaging (BMA) approach. The statistical model is a mixture of generalized linear models that treat model forecasts as predictors for a training set of observed precipitation. Spatial dependence of observed precipitation is another important characterisitic that is investigated. The meteorological component yields predictive distributions for precipitation at desired locations in addition to estimates for model weights in the BMA procedure.

The QPF predictive distribution is subsequently sampled to provide a mean areal precipitation input distribution for rainfall-runoff simulations over a moderatley large watershed. This is combined with prior distributions for rainfall-runoff model initial states and parameters in a Markov Chain Monte Carlo (MCMC) routine. The result is a predictive distribution for stream discharge that can be used for short-term probabilistic flood forecasting.