Bayesian calibration of HEC-HMS model using two different algorithms

Abstract

The assessment of hydrological model parameters is impacted by uncertainties that may be attributed to several sources of errors. Many methods have been developed and applied to various hydrological models in order to address these uncertainties. The present study aims to apply two uncertainty analysis methods, i.e., generalized uncertainty estimation (GLUE) and differential evolution adaptive metropolis (DREAMzs), for the purpose of evaluating the parametric and predictive uncertainty of the HEC-HMS hydrological model. The HEC-HMS model was coupled to these two approaches in order to simulate flash floods in the Mekerra watershed (Algeria) and to estimate the uncertainty of 14 model parameters which are due to be calibrated. For this, six flood events were selected for the purpose of calibrating and validating the model, two informal likelihood measures L1 and L2 were applied in GLUE and one formal likelihood measure was applied in DREAMzs. Considering the difference between the results, and not the principles of the two methods, the analysis indicated that these parameters can be better identified by DREAMzs algorithm. Indeed, the coefficients of variation of their posterior distributions are generally lower. Compared to GLUE, DREAMzs algorithm is more efficient in exploring the parametric space. With regard to the predictive uncertainty, 95% of the uncertainty intervals were estimated by both methods and then analyzed. The DREAMzs algorithm results suggested a good dynamics of the simulated floods, associated with a significant uncertainty that is due not to the parametric uncertainty but to other sources of errors. The GLUE algorithm generates similar results but does not distinguish between the total uncertainty and parametric uncertainty. Further studies would be desirable to evaluate the impact of other likelihood measures on the simulation results with both algorithms