Improved event-based flood warning system for small catchments using artificial intelligence and the CatRaRe catalog

The prediction, warning, and impact of heavy precipitation events are highly dependent on the available data basis and regional factors. Especially in small catchments, explicit warning is often hampered by the lack of runoff data. The effects of urban flash floods triggered by heavy rain events can also be difficult to predict in catchments of small streams. This greatly increases the risk of unexpected damage in these areas. Prediction systems for small catchments (up to about 200 km²), so far mostly rely only on precipitation forecasting and simple soil water balancing. With this research, a methodology improved by artificial intelligence for the prediction of flood events in small catchments is presented.

The CatRaRe catalog with heavy rainfall events of the last 20 years from the German Weather Service is used as a basis for the investigations of heavy rainfall events in small catchments. However, stationary area parameters and runoff data of streams within the catchment are further missing for a precise area-based prediction. The latter were modeled by a recurrent neural network (RNN) in the form of runoff ratio values. Thus, with additional consideration of the CatRaRe catalog, a step prediction system of selected, small catchments in North Rhine-Westphalia (NRW) is created.

Based on the Digital Elevation Model of NRW (DEM 50), catchment areas are determined and assigned to the events of the CatRaRe catalog. For each selected catchment, the maximum discharge values of a gauging station, within a catchment and given time window after the precipitation event, are investigated. As an additional factor, the ratio between the maximum discharge as well as the mean discharge after a corresponding precipitation event is determined. In catchments without a gauging station and therefore without a runoff time series, an RNN is used to fill data gaps. Recurrent networks using the LSTM (long term short memory) method have already been successfully used to simulate time series, since LSTM networks can model temporal and spatial variability well^1,2.

Further input variables for the RNN are the primary soil type and the size and topography of the respective catchment. Additional information about the pre-rainfall index and the magnitude of individual events is also incorporated as differently (area-dependent) weighted measures. Thus, in the case of a precipitation event, it can be calculated whether critical runoff values or runoff ratio values were observed during similarly intense events in the past. This information is regionalized by means of the RNN and can be transferred to non-gauged catchments.

As the data basis grows, events that have already occurred - in particular the July 2021 event - will be evaluated accordingly in further analyses and the sensitivity of the respective influencing variables for the forecast will be adjusted.

 

¹Kratzert, F., Klotz, D., Brenner, C., Schulz, K., and Herrnegger, M.: Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks, Hydrol. Earth Syst. Sci., 22, 6005–6022, https://doi.org/10.5194/hess-22-6005-2018, 2018

²Hu, C.; Wu, Q.; Li, H.; Jian, S.; Li, N.; Lou, Z. Deep Learning with a Long Short-Term Memory Networks Approach for Rainfall-Runoff Simulation. Water 2018, 10, 1543, https://doi.org/10.3390/w10111543