Extreme Precipitation and Flash Floods in Central Western Europe - Occurrences, Atmospheric Conditions, and Regionalized Modelling

The recent accumulation of extreme rainfall and subsequent flash flood events in Luxembourg in 2016 and 2018 has raised two major concerns:

1. How can the extreme rainfall and runoff events that recently occurred in Luxembourg and the Greater Region be conceptualized in an era of proven global change?
2. Can we use large-scale data sets and machine learning algorithms to simulate extreme and flash flood events?

To evaluate potential trends in the frequency and magnitude of extreme precipitation and flash flood events, we identified and analysed thunderstorm relevant weather patterns and atmospheric conditions. Therefore, we made use of the objective weather pattern classification by Hess & Brezowsky (1977) and proxy parameters from reanalysis data (ERA 5) representing abundant atmospheric moisture contents, sufficient atmospheric instability, and low wind speed and shear winds. These three categories typically characterize thunderstorm conditions with intensive and long-lasting rainfall, that have the potential to trigger flash floods. The occurrence of all other weather patterns fluctuated without clear trends between 1960 and 2018, whereas the occurrence of trough weather patterns, which can be relevant for meridional, thunderstorm causing conditions, has increased over time. However, between 1980 and 2020, we did find significant increases in the occurrence frequency and amount of abundant absolute atmospheric moisture contents along with increases in high atmospheric instability. These findings are the basis for potential increases in the occurrence and intensity of extreme precipitation and flash flood events.

Moreover, regarding the damage potential of flash floods, it is crucial to further develop the simulation of peak flows. As most hydrological models used in forecasting are exceeding their design limits during extreme summer peak flow events, we explore the potential added value of large hydrological data sets. Therefore, we physio-graphically characterized the entire region of the Luxembourgish stream network and collected hydro-meteorological data, which were compiled in the CAMELS-LUX data set. With this data we trained and tested a neural network (Long Short-Term Memory, LSTM). For the first time, we were simulating discharge in all 56 nested catchments achieving overall reasonable model results. However, also the LSTM model had difficulties simulating the extreme peak flows mentioned above. Attempts to improve the model performance during summer peak flows by adding the identified thunderstorm relevant atmospheric parameters to the model and increasing its temporal resolution, did not lead to significant improvements. Instead, we noticed the inability of the standard LSTM to extrapolate discharge simulations to values outside the normalized range of the training data. This shows the design limits of data-based models, that function very well only during conditions that are well represented in the training data set.

Overall, my research has contributed to a better understanding of the development of flash floods, that might not only be linked to atmospheric changes, but also to other determining driving factors. With the extensive description of hydrological data in and around Luxembourg, we pave the way for further exploitation of regional hydrological data. Delivering the basic LSTM model as a tool opens a range of possibilities for further model development and analyses.