The contribution discusses the problems with modelling design floods for water structures. The statistical extrapolations of observed flood series of, for example, 80 years "only" to the annual exceedance probability AEP = 0.01 is difficult due to the large variability in extreme values. For large dams, however, the AEP = 0.001 or 0.0001 is required. Most of the uncertainties in hydrological modelling are epistemic (uncertainties in model structure, model parameters, inputs, calibration data, and in measurements) and moreover some measurements can be disinformative.With powerful computers, it is now possible to produce very long series (100 to100,000 years in hourly time step) using precipitation and temperatures computed with a weather model. Within the framework of the Generalised Likelihood Uncertainty Estimation (GLUE) many (thousands) of such continuous simulations are produced and compared to the observed historical data. According to Keith Beven's Manifesto for the equifinality thesis the differences between modelled and observed values should not be larger than some limits of acceptability based on what is known about errors in the input and output observations used for model evaluation (e.g., for flow the current metering data are used). The unacceptable realisations are rejected. We have been working with the frequency version of TOPMODEL in various versions according to the unique characteristics of each catchment. Design hydrographs for water structures are then extracted from the acceptable realisations. The continuous simulation with uncertainty estimation seems nowadays the most promising method of computing design hydrographs for important water structures, even if issues associated with epistemic uncertainty of model assumptions remain.