Floods are among the most destructive natural hazards. They affect thousands of people worldwide and cause damage for millions of euros. Structural measures such as retention basins, dikes and flood control areas, aim to minimize the flood risk. They have been installed along most rivers, and will continue to be installed and upgraded in the future. However, these measures cannot eliminate the flood risk completely. This might be more the case in the future than in the past, due to the potential current and future increase in flood risks as a result of urbanization trends and climate change. For these reasons, it is important that local authorities and the people at risk can be warned timely for the remaining floods. Flood forecasting systems are being developed for that purpose. These systems utilize rainfall observations and predictions to predict water levels and flows along the rivers through hydrological and hydrodynamic models. The applicability of these systems obviously depends strongly on their accuracy. Both models and models inputs are subject to uncertainties, which need to be taken into account in the forecasting. Quantification of these uncertainties therefore becomes mandatory. Another research question is how to reduce these uncertainties, hence improving the accuracy of the flood forecasts. Quantification and reduction of the uncertainties in river flood forecasting were the two main objectives of this doctoral research. Another (secondary) objective was the communication and visualization of these uncertainties.
To quantify the uncertainty in the flood forecasts a non-parametric data-based approach was developed. This approach takes into account the heteroscedastic behaviour of the forecast errors and can be applied on both water level and discharge predictions. The calculation time of the method is very limited, which is crucial for its application in an operational context.
The developed approach to quantify the uncertainties can also be used to identify the different sources of uncertainty and their contribution to the total flood forecast uncertainty. After simulating historical forecasts and eliminating individual sources of uncertainty (e.g. by making use of observations instead of forecasts), the total hydrological forecast uncertainty could be split in its sources by applying variance decomposition. Based on this improved insight in the uncertainty composition, targeted improvement actions could be defined, which aim to reduce the flood forecast uncertainty in the most efficient way.
The first improvement action studied was the hydrological model calibration. An advanced calibration methodology was applied that takes into account the performance of the hydrological model to make extrapolations beyond the range of historical conditions considered during model calibration. A novel method has been developed that studies the relationship between changes in peak flow versus changes in rainfall intensity. This relationship can be derived from the historical measurement series for different percentiles of flow and rainfall changes, and compared with the ones derived from the model simulation results. It is shown that by testing these relationships during the calibration of the hydrological model, more reliable model results are obtained for extremes.
Another large source of uncertainty in the flood forecasts is the uncertainty in the rainfall forecasts. The total uncertainty in the rainfall forecasts was quantified based on statistical analysis of forecasted versus observed rainfall. This rainfall forecast uncertainty was propagated to the total flood forecast uncertainty by means of Monte-Carlo simulation. This method was compared with another, more popular method, based on ensemble rainfall forecasts. This ensemble is based on the perturbation of the initial conditions of the numerical weather prediction model, allowing to generate multiple rainfall predictions. It is shown that the popular ensemble approach underestimates the total rainfall forecast uncertainty. Notwithstanding these improvements in the flood forecasting system, flood forecast uncertainties cannot be fully removed; some will remain. Therefore this work also addresses the visualization and communication of these uncertainties. Different uncertainty communication methods have been worked out, implemented and tested. The best method depends on the user of the forecast results. In addition, different tools have been developed to aid in the warning and decision making process. One promising tool is the calculation of the exceedance probability of a predefined (flood) discharge threshold, based on soil moisture estimates and precipitation data.