Hydraulics/Hydrology Seminar Series
Bayesian Model Average for flood prediction under uncertainty
Mr. Zhu Liu
Tuesday, November 15, 2016
Flood predictions are carried out by dynamic models. Depending on the model setup, the predicted flood inundation or stage has uncertainty associated with it. However, only one prediction is needed for decision making purpose. In this research, we apply a statistical method Bayesian Model Average (BMA) to develop a more skillful probabilistic prediction from multi-model simulations. Contradict to the statistical analysis which proceeds conditionally on one assumed/deterministic "Best Model", Bayesian Model Average is weighted average of the individual models in ensemble. The BMA method is tested on the 20000 km2 Black River Basin in Arkansas with the 2008 March-May flood event and 18 LISFLOOD-FP models considering uncertainty of channel shapes, widths and bed elevations are built to form the prediction ensembles. The BMA prediction indicates that it does not outperform the best model, however, it can provide a more consistent and reliable result than a single model across the basin.
Zhu Liu, PhD student of Dr. Venkatesh Merwade in Lyles School of Civil Engineering at Purdue University. Zhu obtained his bachelor degree in Hydro-power Engineering in China (2012) and his master degree in Civil Engineering in UC-Irvine (2014). His current research is flood modeling and Bayesian uncertainty analysis.