Towards the improvement of discharge forecasts: an analysis of nowcasting using radar precipitation for water management

The timely publication of accurate discharge forecasts for rapidly responding catchments is critical in terms of optimal water management and flood warnings. Discharge forecasts are usually produced using hydrological models, which are enormously dependent on the underlying precipitation forecasts. A new study by Deltares, Wageningen University and Research Centre, and Delft University of Technology was published recently in Water Resources Research. It examines the added value and possible improvements of nowcasting of discharge forecasts in twelve Dutch catchments and polders.

Forecasting continues to be challenging

Despite ongoing improvements in weather models in recent decades, predicting the exact location and amount of precipitation remains challenging, even several hours ahead. And the smaller the catchment, the more important the exact location and timing of precipitation becomes. Nowcasting, predicting precipitation up to a few hours in advance using rain radar products, may represent a solution.

650 precipitation events

The fact that an accurate discharge forecasting system can be of enormous value became clear again last summer during the floods in Limburg, Ardennes and Eiffel regions. Precipitation forecasts with nowcasting could be useful in this respect. Before a nowcasting technique can be applied in water management, it will be necessary to determine the added value and possible weaknesses of the technique for producing discharge forecasts. And that is precisely where the problem lies, because no large-scale analysis had been conducted of nowcasting for discharge forecasts, either in the Netherlands or internationally. On the basis of a large-scale analysis of nowcasting, this study provides an initial insight for more than 650 individual precipitation events in the Netherlands. Ruben Imhoff (Deltares): “We hope that this analysis will provide Dutch water managers with a picture of what they can expect from nowcasting in their forecasting systems. However, internationally as well, this is also the first study of this size in this domain. It is enormously important to identify not only the added value but also possible areas for improvement.”

Advanced techniques

Several nowcasting techniques were tested in the analysis. Within a given margin of error, the methods tested provide a peak discharge forecast, and the timing and height of the highest discharge wave, about two to three hours earlier on average than methods without nowcasting. More advanced nowcasting techniques often produce better results than the other tested techniques, and they also significantly reduce the number of false alarms, in other words the forecasts that generate a warning of an event that does not ultimately materialise. An important area requiring attention continues to be that all the nowcasting techniques tested underestimate the future precipitation volume for the catchments, particularly during rainfall events with high precipitation intensities. All in all, nowcasting may certainly provide added value for water management.

Machine learning and deep learning

The analysis makes clear the strengths of nowcasting for producing discharge forecasts but it also immediately shows where further research will be needed. In terms of water management, these are the underestimates of the precipitation volumes, particularly during thunderstorms, which are extremely difficult to predict with the current numerical weather models. Machine and deep learning techniques may help here. The research now being conducted by Deltares and colleagues from Wageningen and Delft universities focuses on the optimal combination of nowcasting techniques with current numerical weather models – blending – in order to get the best of both worlds. The ultimate goal is to extend the forecast horizon referred to here from about 2 to 3 hours to about 6 hours.