4.2

**Towards Improved Ensemble Hydrologic Predictions: A Multi-Model Multi-Objective Bayesian Approach**

**Qingyun Duan**, LLNL, Livermore, CA; and J. A. Vrugt

In a series of papers since 1992, Keith Beven and coworkers have been advocating the use of a Generalized Likelihood Uncertainty Estimation (GLUE) framework to generate probabilistic hydrologic predictions and to quantify the associated uncertainties. They presented a series of pseudo-likelihood measures to discriminate behavioral models (or model parameters) from non-behavioral models (or model parameters). Beven deems all models or model parameter sets which yield model simulations with pseudo-likelihood measures above certain threshold values as equally good or equifinal. In presenting the GLUE framework, Beven has intentionally left it open to the users on what sampling strategy to be used to search for behavioral models or model parameters. Poor or inefficient sampling strategies would run the danger of producing predictions that are too uncertain or too confident to be of practical use. Here, we present a strategy to extend and improve the GLUE methodology. As in GLUE, multiple hydrologic models and a large basket of pseudo likelihood measures will be used to generate many sets of ensemble hydrologic predictions. A Bayesian Model Averaging (BMA) scheme is then used to combine these ensemble sets and produce consensus probabilistic hydrologic predictions which have better overall predictive skill as measured by pseudo likelihood measures and better reliability as measured by statistical consistency (better calibration subject to sharpness). A key to improve GLUE sampling effectiveness and efficiency will be the use of an efficient Monte Carlo Markov Chain (MCMC) method known as the Shuffled Complex Evolution Metropolis (SCEM-UA) method. In this presentation, we will present the mathematical methodology and demonstrate its usefulness using different hydrologic models and a selected set of hydrologic basins in the US. .

Session 4, Hydrologic Data Assimilation, Parameter Estimation, And Uncertainty

**Thursday, 2 February 2006, 1:30 PM-5:15 PM**, A403** Previous paper Next paper
**