![]() |
| Das Dokument ist frei verfügbar |
|
| | Nachweis | Kein Nachweis verfügbar |
|
| Floods are one of the most prevalent natural disasters impacting 55 millions lives yearly. Floods in small catchments have substantial implications yet their risks and impacts remain difficult to predict using the state-of-the-art GHMs and FEWS. This dissertation focuses on addressing these limitations by developing new methods to improve flood impact forecasting. Through the integration of improved large-scale streamflow modeling machine learning and real-time flood mapping the research aims to enhance FEWS capabilities and make them more responsive to localized flooding events. One key innovation explored in this research is the use of 2D hydrodynamic models to generate real-time flood inundation maps and impact indicators. Existing FEWS either interpolated flood hazard maps (e.g. GloFAS EFAS) or relay only the local rainfall depths or gauge levels to the users. This results in inaccuracies for unprecedented extreme events in the first case and misinformed actions in the second. We demonstrate the feasibility of the ICON-D2-EPS-mHM-RIM2D operational FEWS for the 2021 European Summer Flood in the Ahr valley. Parallelized RIM2D high resolution flood inundation ensemble runs on GPUs reduce the total forecast runtime of the FEWS to under three hours. The FEWS forecasts lead time to specified inundation thresholds and at-risk infrastructure factoring in forecast uncertainty which are crucial information for emergency response teams and policymakers. Representation of catchment shape presents a unique challenge for GHMs. D8 arguably the most widely used method of catchment upscaling in GHMs struggles to accurately represent catchment shapes smaller than 30 times the area of the grid pixel. This dissertation introduces a novel stream upscaling technique - SCC - as a solution. SCC improves the accuracy of streamflow simulations by allowing multiple downstream connections within a single grid cell thereby addressing the limitations of the widely-used D8 method which only permits a single outflow direction per cell. This advancement in stream network upscaling significantly enhances the precision of modelled catchments ensuring that even the smallest contributing areas are properly accounted for. The effectiveness of SCC is demonstrated through experiments in the Rhine basin and at the global scale. In these experiments SCC not only improves the accuracy of streamflow simulations but also offers up to a five-fold increase in computational efficiency compared to existing methods. This directly contributes to the real-time applications of FEWS where speed and accuracy are paramount. Furthermore SCC ensures accurate streamflow across modeling resolutions eliminating the need for GHMs to reach sub-kilometer scales for streamflow precision. This combined with the advantage that SCC can be integrated with any land surface or hydrological model significantly expands its use in global flood forecasting systems. Small catchments have historically been under-served by global models but are often the location for catastrophic flood events. This research makes an important contribution by addressing the "catchment size problem" a long-standing problem in GHMs where representation accuracy diminishes with smaller catchments. SCC improves the upscaling of catchment areas irrespective of catchment size solving the catchment size problem entirely and allowing GHMs to deliver locally relevant streamflow at given points of interest. This represents a major opportunity in flood forecasting technology allowing FEWS to rely on fluvial boundary conditions from a unified GHM setup for flood events at any scale whether local or regional eliminating the need for separate hydrological model setups and the associated resolution challenges. The SCC method also offers a novel solution to the issue of simulating streamflow at multiple points of interest within a single grid cell. Current methods (e.g. D8) are limited to providing a single streamflow value per grid cell. The multiple downstream connectivity of SCC allows for a grid to have multiple routing fractions with the corresponding values overcoming the limitation of single streamflow values within the same cell. This feature is especially important for regions with complex hydrological setting such as multiple tributaries intricate river networks dense networks of monitoring stations or high density of reservoirs where capturing the full scope of hydrological interactions is crucial for accurate streamflow predictions. In addition to these advancements in catchment representation the dissertation explores the use of ML to improve the simulation of streamflow downstream of regulated reservoirs. Reservoirs have the potential to introduce significant discontinuities in natural streamflow patterns. These discontinuities are often difficult to model using traditional hydrological approaches. The research presents a ML based method to predict non-consumptive demand at hydropower reservoirs based using downstream streamflow observations as control point. The ML demand model when fed to the hydrological model enables more precise simulation of daily streamflow downstream of 31 global reservoirs. This is particularly important for FEWS as regulated rivers often pose significant forecasting challenges due to the variability in reservoir operations based on water demand. The improved simulations would allow FEWS to generate more accurate predictions which can help mitigate flood risks in communities downstream of large/ disruptive reservoirs. While the current focus of the ML model is on non-consumptive reservoirs its methodology could be extended to consumptive uses like irrigation if reliable data is available. Another novel aspect of this research is the investigation into the role of reservoir bathymetry the underwater topography of reservoirs on lake surface evaporation. Reservoirs contribute substantially to global evaporation yet their shapes are often oversimplified in GHMs. The dissertation quantifies the impact of bathymetric assumptions on evaporation and streamflow predictions finding that oversimplifications can lead to significant overestimation of evaporation. A new function for estimating reservoir surface reflectivity based on latitude and the solar elevation angle is introduced offering a more physically accurate approach to modeling reservoir evaporation dynamics in GHMs. These findings enhance the accuracy of evaporation estimates and improve upstream fluvial boundary conditions for FEWS to forecast downstream flood risks. The final contribution of the dissertation addresses the computational challenges of large-scale hydrological modeling. Current global database includes 38 000 georeferenced dams. Simulating every reservoir in a large model domain is computationally expensive so this study introduces a prioritization method based on reservoir "disruptivity" - the degree to which a reservoir alters natural streamflow patterns. By establishing thresholds for excluding less disruptive reservoirs from simulations the dissertation offers a way to reduce computational costs while still maintaining accuracy in GHMs. This is particularly valuable for regional scale FEWS where real-time forecasting requires a balance between precision and computational efficiency. In summary this dissertation makes significant contributions to the development of more accurate efficient and reliable flood forecasting systems. Through the integration of fast 2D hydrodynamic models the introduction of the SCC method and investigation of reservoir representation methods the research enhances the capabilities of GHMs in generating fluvial boundary conditions in FEWS to forecast "flood impacts" in small catchments and regulated rivers. The advancements made in this dissertation offer transferable tools and methodologies that provide a foundation for future work in global hydrological modeling and flood forecasting and the opportunity to reduce flood impacts on communities worldwide. |
|
|