Increasing NWP thunderstorm predictability using ensemble data and machine learning

Forecasting thunderstorms several hours in advance remains challenging due to the increasing forecast uncertainty in numerical weather prediction (NWP) models with lead time. Ensemble systems address this limitation by providing multiple scenarios consistent with forecast uncertainty and are well known to enhance the skill of deterministic systems. This study introduces a simple yet novel analytic expression that quantifies the improvement in Brier Skill Score (BSS) achieved by averaging over binary classification predictions from multiple ensemble members. This is particularly relevant to severe weather forecasting, where the task often involves estimating the probability of events such as lightning, heavy rainfall, or hail. The derivation of the formula relies only on the assumption that ensemble members are indistinguishable. We validate this expression using SALAMA 1D, a recent machine learning (ML) model designed to predict thunderstorm occurrence from convection-permitting ICON-D2-EPS ensemble forecasts over Central Europe. Our formula accurately captures the impact of ensemble averaging on the ML model's performance, which, in this case, results in extending the model’s 5-hour deterministic skill out to 11-hour lead times. While we use an ML model to exemplify our formula, its validity extends also to traditional (non-ML) approaches for severe weather identification in NWP data. Furthermore, we show that ML models like SALAMA 1D, which are trained using observations as ground truth labels, can identify patterns in thunderstorm occurrence that remain predictable for longer lead times compared to raw NWP output. Our findings offer insight on the use of ensemble forecasts of thunderstorm occurrence and support the growing use of ML techniques in severe weather forecasting.