^{1,2,a}, Oswaldo Morales-Nápoles

^{1}, and Matthijs Kok

^{1,2}

^{1}Civil Engineering and Geosciences, Delft University of Technology, The Netherlands.

^{a}g.w.f.rongen@tudelft.nl

^{2}HKV Consultants, The Netherlands

Statistics of extreme discharges along the Meuse are needed for reliability analysis and design of flood defenses. These are often obtained through a series of models that generate long synthetic time series, in which the extremes should be represented. In this work, extreme discharge are generated from measurements and statistical models based on observed discharges and geographical characteristics. The statistical model is based on a Generalized Extreme Value distribution for each catchment, and a Non Parametric Bayesian Network (NPBN) that correlates the discharges. We used hierarchical graph configurations of the BN, in which latent variables group catchments based on location and catchment characteristics. The hierarchical configuration of the Bayesian Network did not represent the dependence structure better than the 'conventional' direct graph structure, but a combination of direct and hierarchical did. The model forms a flexible and refreshing alternative to conventional hydrological models.