Korrektur von Modell- und Vorhersagefehlern und Abschätzung der prädiktiven Unsicherheit in einem probabilistischen Hochwasservorhersagesystem

The main objective of a probabilistic flood forecasting system is the reliable estimation of the Predictive Uncertainty (PU), which contains all information available about the forecast variable given the history of observed and predicted values. This uncertainty can be separated into the hydrological uncertainty comprising model, parameter and measurement uncertainties and into the input uncertainty caused by weather forecasts. Within the European Flood Awareness System (EFAS) an error correction methodology based on wavelet transformations and Vector AutoRegressive models with eXogenous Input (Wavelet-VARX) has been developed in order to capture the hydrological uncertainty. The error-corrected stream-flow simulations, resp. forecasts, are combined with a Hydrological Uncertainty Processor (HUP) by applying the Bayesian theorem. The input uncertainty is estimated by taking different weather forecast systems and Ensemble Prediction Systems (EPS) as driving forces and by optimally combining the resulting multi-model stream-flow forecasts based on the quality of the forecast from previous days. The total PU is finally derived by integrating the hydrological and the input uncertainty. The sharpness and accuracy of the total PU is evaluated by means of the Continuous Rank Probability Score (CRPS) and shows some significant improvements in the quality of the probabilistic flood forecasting system.