Flash Flood Forecasting Combining Meteorological Ensemble Forecasts and Uncertainty of Initial Hydrological Conditions

Flood forecasting for fast responding catchments encounters problems especially in terms of short warning periods and a very limited reliability. Within a new stochastic framework based on rigorous rainfall-runoff modelling and Monte Carlo simulations we consider uncertainties of two sources: (i) uncertainty from the estimation of initial hydrological conditions and (ii) the uncertainty of the meteorological rainfall forecast. We avoided the high computational demand of extensive Monte Carlo simulations by using a symbiosis between physically based hydrological modelling and computationally highly efficient artificial intelligence techniques. The new PAI-OFF methodology (Process modelling and artificial intelligence for online flood forecasting, see Schmitz et al., 2005 and Cullmann et al., 2006) employs a physically based hydrological/hydraulic model of the considered catchment for generating in a first step the complete range of realistic possible flood scenarios on the basis of a catchment specific meteorological analysis. The resulting database of corresponding input/output vectors - supplemented by generally available hydrological and meteorological data for characterising the catchment situation prior to a storm event - serves in a second step for setting up a set of task specific artificial neural networks (ANN) which finally portray both the rainfall-runoff process and the hydrodynamic flood wave propagation in the river. We subsequently use this tool for investigating the global uncertainty of flash flood forecasting in a small to medium sized catchment on the basis of a comprehensive Monte Carlo analysis. Along these lines, the computationally highly efficient PAI-OFF technique allowed performing an extensive number of runs for obtaining ensembles of predicted stream flow which can be used to evaluate probabilities of exceedance of critical river stages/flows via an integration of both the hydrological uncertainty and the meteorological uncertainty. This approach was then implemented and applied to the Freiberger Mulde catchment in the Ore Mountains in Eastern Germany (with an Area of about 3000 square kilometres). The results of the overall ensemble predictions in the form of the ensemble mean values unveiled an astonishing underestimation of the recorded flood peak - most due to the bias of the considered initial hydrological conditions.