Streamflow prediction is generally made with a single hydrologic model. With many limitations and uncertainties associated with the current generation of hydrologic models, relying on a single model often underestimates the total predictive uncertainty.

Obtaining consensus prediction by using multi-model combination techniques such as Bayesian Model Averaging (BMA) has gained popularity in many scientific fields over recent years. It has been reported that the consensus prediction often has better predictive skills when compared with any single model prediction. Additionally multi-model framework has shown to provide better uncertainty quantification. In this study we present a Recursive Bayesian Model Combination (RBMC) scheme for streamflow prediction. The key idea is to produce consensus streamflow prediction that does well in all phases of the hydrograph (i.e., high peaks, mid-range flows, and low flows).

In this study three different hydrologic models are employed. These models are calibrated to historical streamflow observations using five different objective functions. The calibration approach places emphasis on the ability of the models to capture different phases of the hydrograph. The fifteen calibrated streamflow time series in our study form the streamflow prediction ensemble. Weights are then computed for each ensemble member using RBMC scheme for each hydrograph phase. The consensus streamflow prediction for each hydrograph phase is calculated as the weighted average of the entire ensemble. The overall streamflow prediction is the combined prediction of all hydrograph phases. This presentation describes the RBMC scheme and illustrates its use for streamflow prediction on a number of US hydrologic basins employed in the international Model Parameter Estimation Experiment (MOPEX).