AI-based blending of conventional nowcasting with a convection-permitting NWP model

For hydrological runoff simulations in hydropower applications, accurate analyses and short-term forecasts of precipitation are of utmost importance. Traditionally, radar-based extrapolations are used for very short-term time scales (approx. 0 - 2 hours ahead). However, during recent years, convection-permitting NWP models have become better at very high spatial and temporal resolution forecasts (e.g. through radar assimilation, RUC configurations). Such models have the advantage of capturing the complex and non-linear evolution of precipitation systems like fronts or thunderstorms in a more physically accurate way than extrapolations, but they are also prone to inaccuracies in precipitation distribution. The aim of this paper is to employ machine learning to combine the strengths of the conventional radar extrapolation (localization and movement of existing storms) with the benefit of the model’s ability to predict storm evolution.  Results show that even a relatively simple sequential deep neural network is able to outperform both, the operational nowcasting and NWP model forecasts. However, the results are highly sensitive to variable selection, loss function, and localization features have a large impact on performance, which is also discussed.