Towards a machine-learning enhanced nowcasting tool for storm severity analysis and prediction

At the German Weather Service (DWD), an important tool for issuing severe weather warnings is the NowCastMIX system (James et al., 2018), which provides warnings for the next hour based on radar, lightning, NWP and other data. In its analysis step, NowCastMIX detects thunderstorm cells and assigns them a severity level based on remote sensing and numerical model data at their locations. In the prediction step, the thunderstorm cells are clustered and then spatially extrapolated according to the estimated motion vectors of the cells to provide a simple temporal evolution for the next hour. The extrapolated cell clusters retain the severity assigned to them in the analysis step.

One of the focal points of the work presented is the analysis step, in particular the assignment of severity to the detected storm cells. This is currently based on fuzzy logic, which allows estimation of severity based on a set of thresholds for measured or predicted meteorological variables. This approach works well, but systematic changes to the input systems (e.g. radar upgrades, improvements to radar products, major model upgrades) always require re-tuning of the fuzzy functions, making the adaptation of NowCastMIX relatively complex and costly. As an alternative, we present our attempts to replace fuzzy logic with machine learning. Although this approach is very attractive due to its simplicity and flexibility, it also has its limitations, which we will present in detail.

The other focus of the presentation is the prediction step. As mentioned before, this is based on linear advection in the current version of NowCastMIX. To improve this, we present an approach that incorporates the convective life cycle of cell clusters by predicting the temporal evolution of storm severity using machine learning.