Multi-model ensemble simulation of flood events using Bayesian Model averaging

Abstract

Flood event simulation using hydrological model is subject to various uncertainties. Multi- model ensemble simulation approach has proved to improve forecasting uncertainty by reducing the systematic bias in comparison with its single model. In this paper, the Bayesian Model Averaging (BMA) approach, a statistical scheme based on multi-model ensemble, was applied for flood prediction and uncertainty estimation of flood event predictions in Qingjianhe Catchment. Five hydrological models including GR4J, HYMOD, Simhyd, XAJ and modified SCS were employed and calibrated with two objective functions NSE and WR 2 . Ten ensemble simulations were then used for further BMA analysis. The results showed that the five hydrological models performed reasonably well in QJH catchment, but with significant difference in the simulated hydrographs. The modified SCS model performed the best among the five models in term of NSE and WR 2 . The BMA weights for the model prediction were roughly consistent with the model performance. GR4J model weighted higher than other models. Predictions with BMA median performed less well than those from the best individual model SCS, especially for peak flows. However, BMA gave more reliable predictions. For most flood events with different recurrent periods in the study catchment, the 50% confidence interval seemed sufficient to bracket the observed flood discharge. It indicates that BMA approach is helpful in reducing uncertainties, thereby increasing the level of confidence in prediction results. The prediction uncertainty quantified via BMA can be very helpful for decision makers to develop flood management strategies.