Issuing flood warnings in Denmark based on an ensemble of hydrological forecast models.

While Denmark has a long history of coastal floods and a robust storm flood warning system, forecasting fluvial flooding is a new challenge for the country presented by climate change. In 2022 the Danish Meteorological Institute (DMI) was assigned the responsibility for issuing national flood warnings in Denmark. DMI is mandated to send public flood warnings based on an ensemble of hydrological forecasting models from summer 2025.

This work presents the Danish experience in defining combined model ensemble criteria and warning levels approach for issuing flood warnings.

This study evaluates the criteria for determining the most reliable flood warnings using variants of two physically based classical hydrological models and one data-driven machine learning model. The models included in the ensemble are: (a) a physically based flood forecasting model based on the Hydrological Predictions for the Environment (HYPE) model (b) a variant of the HYPE model incorporating data assimilation, (c) a physically based MIKE-SHE model with detailed representation of groundwater, (d) a data-driven Long Short-Term Memory (LSTM) machine learning model using catchment characteristics and hydrological variables, (d) a hybrid model coupling of the LSTM model to the HYPE model.

While the HYPE and LSTM models have been developed by DMI, the MIKE-SHE model is developed by the Geological Survey of Denmark and Greenland (GEUS).

The performance of the ensemble model criteria is evaluated over the period 2011-2022 for warning stations across Denmark. We define warning levels based on extreme value statistics for observed discharge data to identify several return periods, issuing warnings when forecasted flows exceed these thresholds.

The study assesses the ensemble-based criterion that offers the best performance for issuing flood warnings across Denmark. We evaluate the performance using metrics such as Critical Success Index (CSI) and the Equitable Threat Score (ETS). Additionally, we examine trade-offs between the Success Rate and False Alarms and analyze spatial trends in model performance. We also reflect on the ease-of-use, scalability across Denmark and efficiency of the warning criteria because the criteria will ultimately be adopted for operational flood warning in real time in Denmark.

We find that it is important to consider the interplay between limitations in individual models in our ensemble and the choice of warning criteria to select a combination that provides a robust basis for issuing useful flood warnings.

Our experience in implementing Denmark's first flood warning system combining ensemble models with warning criteria offers valuable insights for countries where flooding is emerging as a new challenge brought by climate change.