Ensemble flood forecasting in small catchment using AI-based and deterministic rainfall runoff models – a performance comparison

Reliable flood forecasting systems are an important prerequisite for local authorities and flood defence units to prepare for potential flooding at an early stage and to initiate the required measures. Small catchments in mountain ranges pose particular challenges in this regard, as they respond very quickly to rainfall. Furthermore, forecasts of rainfall in terms of their spatial and temporal extent and the associated impact on the areas are subject to a high degree of uncertainty. Ensemble rainfall forecasts as input data for rainfall-runoff (RR) models allow for the evaluation of the uncertainty of the resulting runoff. However, this requires fast-computing RR models to cope with the simulation effort, for which artificial intelligence methods are being used increasingly. Against this background, two hydrologic ensemble forecasting systems (EFS) are compared and evaluated that have been in operational use for small catchments in Saxony, Germany, for two years.

System 1, called EFS-howa, was developed in the HoWa-PRO research project, and its predictions can be tracked via the warning platform https://howapro.de/. As an RR model, it includes the event-based, deterministic hydrological model DeHM. DeHM covers the hydrologic processes for runoff formation and concentration, channel routing, and the simulation of flood retention dams. Measured discharge and water level data are assimilated within the forecasting process. For the hydrological ensemble forecast, rainfall data for observation and prediction from established products of the German Weather Service are used (radar-based QPE: RADOLAN-RW, radar-based nowcasting: RADOLAN-RV, ensemble QPF: ICON-D2-EPS). The runoff forecast lead time is 48 hours, and new forecasts are released every half hour if the QPF indicates a potential flood threat.

System 2, called EFS-kiwa, was developed in the KIWA research project, and its predictions can be tracked via the web demonstrator http://howa-innovativ.hydro.tu-dresden.de/WebDemoKiwa/. The RR model is a regional AI model (based on LSTMs) that was developed using measured RR data and hydrologic characteristics from 52 small and medium-sized catchments in Saxony, Germany. The current setup of the regional AI-RR model is based on hourly measurements of rainfall (using RADOLAN-RW), runoff, and rainfall forecasts. Thus, it achieves runoff forecasts with a lead time of up to 24 hours. The regional AI-RR model also allows for the fast and robust processing of ensemble rainfall forecasts from ICON-D2-EPS, enabling runoff forecasts with uncertainty/reliability information.

Both systems are evaluated in terms of their performance. Across various forecast lead times, different metrics such as KGE or percentage peak error, as well as threshold-based metrics such as false alarm ratio or area under the ROC curve (AUC), are calculated to explore the quality of both forecasting systems. The differences between the two demonstrators are highlighted by means of the selected metrics and specific simulation results. The associated benefits, advantages, and disadvantages for flood early warning are discussed.