The Norwegian Water Resources and Energy Directorate (NVE) is responsible for national flood warning in Norway. Flood warnings rely on forecasts generated by hydrological models for 145 flood forecasting catchments across the country. Currently, two models are operational: DDD and HBV. In addition, a machine-learning model based on Long Short-Term Memory (LSTM) networks is under development. Reliable and robust flood warnings depend on producing the best possible forecasts. The following are examples of master’s projects that contribute to improving flood forecasting models.
Improved calibration of flood forecasting models
Hydrological models used for flood forecasting are currently calibrated and evaluated using all simulated discharge values, typically with performance criteria such as the Kling–Gupta Efficiency (KGE) and Nash–Sutcliffe Efficiency (NSE). However, this approach may produce calibrated models that do not perform optimally during actual flood events.
This project will investigate whether calibration approaches that give greater priority to high discharge values or flood events can improve flood forecasts without compromising the representation of other important hydrological processes.
Precipitation correction for flood forecasting
Hydrological forecasts are highly sensitive to the quality of precipitation data products. NVE currently uses precipitation products provided by the Norwegian Meteorological Institute: SeNorge at daily resolution, disaggregated SeNorge at 3-hour resolution, and MetNordic at 1-hour resolution. In many catchments, measured runoff exceeds the precipitation estimated by these products.
To balance the water budget, a precipitation correction factor must therefore be applied during model calibration. This correction factor may vary over time because the data underlying precipitation interpolation also change. At present, the hydrological models used for flood forecasting apply different precipitation correction factors. The nationwide models use a dedicated precipitation correction map that improves the water balance in most locations. However, because each model applies its own correction, substantial discrepancies may arise between model estimates—for example, in simulated snow storage during the melt season.
This project will examine possibilities for improving precipitation correction methods and increasing consistency between models used in operational flood forecasting.
LSTM models for flood forecasting
LSTM models have been tested for the 145 flood forecasting catchments in Norway and have outperformed traditional hydrological models in simulations of streamflow and floods. However, these models have not yet been evaluated in an operational forecasting setting.
This project will focus on: (i) developing a model configuration suitable for flood forecasting, and (ii) evaluating forecast performance for lead times of up to nine days. The work may contribute to the future operational use of machine-learning models in NVE’s flood forecasting system.
Regional information in catchment model calibration
The key motivation for this project is to improve streamflow forecasting at ungauged locations, where local discharge observations are not available for calibration. This is challenging because traditional hydrological models can have a wide range of parameter sets that produce similarly good simulations at gauged locations. Consequently, it is uncertain which parameter values will perform well in catchments without observations.
Calibrating multiple catchments simultaneously can help identify parameter values that are robust across different locations. In the most extreme case, this results in one global parameter set, as used in NVE’s gridded SeNorge model. However, SeNorge performs less well than the same model calibrated individually for each catchment, demonstrating that local differences remain important.
At the same time, NVE experiments show that globally trained LSTM models can perform very well because they use catchment characteristics to represent local behaviour. This project will investigate whether traditional hydrological models can apply a similar approach by including catchment information directly in multi-catchment calibration. The aim is to develop a model that can be calibrated globally while still adapting to local conditions, thereby supporting improved forecasts at ungauged locations.
Updating a hydrological model to improve forecasts
Hydrological model forecasts may be inaccurate for several reasons. For example, the groundwater reservoir at the time a forecast is issued may contain too little or too much water. In such cases, even an accurate precipitation forecast may result in incorrect streamflow estimates.
A routine has been developed for the DDD (Distance Distribution Dynamics) model that adjusts, or updates, the model’s groundwater reservoir based on discrepancies between simulated and observed discharge. By accounting for these discrepancies, the routine can improve subsequent forecasts. Because of the model’s hydrological “memory,” an update made when the forecast is issued may have a positive effect throughout the forecast period.
This master’s project will analyse the catchment characteristics and hydrological conditions under which the update routine is effective, as well as situations in which it is less useful. The student should be comfortable becoming familiar with model code, with guidance, and analysing large datasets.