Building a National Operational Flood Forecast System for Denmark: Evaluating Top-Down vs. Bottom-Up and Process-Based vs. Data-Driven Modeling Strategies

The construction of three separate national-scale flood forecast models for the operational flood forecast system for Denmark presents a unique opportunity to compare “top-down” vs. “bottom-up” modeling approaches and “process-based” vs. “data-driven” model types. To implement an operational flood forecast model for Denmark as quickly as possible, a “top-down” process-based hydrological model (E-HYPE DK) was first extracted from the pan-European E-HYPE model developed from European and global data sources. A separate process-based model, DK-HYPE, as well as a data-driven model, DK-LSTM, were developed for Denmark from the “bottom-up” using national data sources combined with high-resolution catchment delineations and more detailed model process representations.

Evaluation of the two modeling approaches showed a trade-off between time invested and societal benefit. Overall, the top-down E-HYPE DK model provided benefit early in the project by providing rapid access to model results which could be used to guide the development of the entire forecast chain and warning system. In contrast, the bottom-up DK-HYPE model developed later in the project, provided better model performance and higher-resolution outputs than the top-down model but required longer time to develop and deploy. While the addition of local high resolution forcing data and hydrological properties in DK-HYPE certainly contributed to the improved performance, changing the representation of groundwater process better captured the importance of surface water-groundwater interactions in Danish river systems. 

Results from the project also highlighted trade-offs between the process-based and data-driven models. Compared to the process-based HYPE models, the data-driven DK-LSTM model required the shortest time for development and offered the best match between simulated and observed discharges. However, the data-driven model had difficulty in making predictions for events outside the training conditions (e.g. storms with unusually high precipitation) and did not provide information about internal variables that are provided by the process-based models (e.g. local runoff and soil moisture) which can be valuable for operational decision making.

The DK-HYPE model is now operational, providing public warnings for high river flows. The DK-LSTM is currently used as a supporting model during warning situations.