Quantification of the practical predictability of thunderstorm occurrence using machine learning

While the statistical post-processing of numerical weather prediction (NWP) data constitutes a powerful ingredient of many forecasting suites of severe weather, post-processing for thunderstorm occurrence becomes ever more difficult as the lead time of the NWP forecast increases. In terms of identifying thunderstorm occurrence as a function of lead time, this increased difficulty is reflected in a decay of skill for which even sophisticated machine learning (ML) models cannot fully compensate. In this work, we propose how the time scale of skill decay of supervised ML models can be studied as a function of the spatiotemporal label resolution used for training. If the label is constructed from lightning observations, label resolution is modified by varying the time and radius thresholds by which strokes of lightning are associated with NWP data. We exemplify our method using SALAMA, a feedforward neural network model which we have developed for identifying the probability of thunderstorm occurrence in NWP data. The model has been trained on convection-resolving ensemble forecasts over Central Europe and lightning observations. We show for SALAMA that the time scale for skillful thunderstorm predictions increases linearly with label resolution, which underlines the practical ability of our method to quantify the predictability of thunderstorm occurrence.