Estimation of storm life-cycle by machine learning to enhance ensemble nowcasting of cell objects

Object-based cell detection and tracking algorithms are widely used to assist meteorological forecasters in their decision-making regarding severe weather warnings. In many approaches, the extrapolation of current storm movement delivers a forecast of future cell position for lead times from five to 120 minutes. However, most operational nowcasting algorithms do not make predictions about the life cycle of storms, although information about cell weakening, intensification, or the general lifetime might be beneficial with regard to meteorological warnings. Prediction of storm life-cycle is not straightforward and involves large uncertainties, as recent research has shown. We aim at using probabilistic machine learning approaches to learn life-cycle information from past storm data and incorporate their prediction and its corresponding uncertainty into an object-based ensemble prediction system.

In a first concept, we identified maximum future storm severity as well as storm lifetime as the important characteristics to be included in the object-based ensemble nowcasting approach. We carried out a cross validation based on a 5-year recalculation of DWD’s object-based nowcasting system KONRAD3D including atmospheric storm environment from NWP model data. Cross validation shows that prediction techniques based on decision trees, i.e. random forest (RF) and gradient boosting (GB) are able to deliver a reasonable performance. For maximum storm severity, both methods do outperform a persistence forecast assuming no future change of current cell severity. Due to the sparse occurrence of long-lasting storms, prediction of cell lifetime is more difficult and the corresponding uncertainty is larger. Quantile regression variants of RF and GB are useful to quantify uncertainty based on current storm attributes and atmospheric environmental conditions and might be a useful base for the generation of an object-based ensemble nowcasting system.

KONRAD3D-EPS, the current ensemble prediction system under development within the project SINFONY (Seamless INtegrated FOrecastiNg sYstem) at the German Meteorological Service, uses a horizontally flipped parabola to model the life-cycle of storm cells in terms of their severity. Each member is initialized by drawing from parameterized distributions of storm lifetime and maximum severity. The improvement presented here uses quantile regression GB predictions to initialize the ensemble in order to adjust the spread depending on the given meteorological situation. In addition to machine learning results, we will present basic functionalities of KONRAD3D-EPS and show the conceptual overview on how to combine it with machine learning predictions.