Improved Event-Based Flood Warning System for Small Catchments Using a Lstm and an Event Catalogue

Abstract

The prediction of heavy flood events is heavily reliant on the available data and regional factors. In particular, explicit warnings are often hindered by the lack of data in small and urban catchments. Additionally, the impacts of urban flash floods caused by heavy precipitation events can be difficult to anticipate in small catchments, which increases the risk of damage in these areas. Warning systems for small catchments (< 200 km2) mostly rely on precipitation forecasting and soil water models. This research presents a methodology for improving the warning of flood events in small catchments using an artificial neural network.The German Weather Service's catalogue of heavy rainfall events from the last 20 years serves as the basis for identifying heavy precipitation events. However, the lack of morphometric parameters and discharge data for streams within the catchment hinders precise warning. The Long Short Term Memory neural network was used to model the missing discharge time series. A warning system was created by taking into account the catalogue.Catchments were determined based on a digital elevation model and assigned to heavy precipitation events. The maximum discharge to mean discharge ratio was determined for each selected catchment and gauge. Long Short Term Memory (LSTM) networks are used to simulate discharge time series in catchments without a gauging station. LSTMs have been successfully employed for this purpose due to their ability to model sequential data well (Kratzert et al., 2018).The warning system also takes into account the primary soil type, size, and topography of the catchment. This research describes the incorporation of information about the antecedent precipitation index, the magnitude of individual events, and other area parameters. It then explains how, in the event of heavy precipitation, it is possible to calculate whether critical discharge values were observed during similarly intense events in the past.

Keywords: Real time flood forecasting, Artificial intelligence, Long short-term memory, Discharge simulation, Prediction of ungauged basins