Exploring the portability of ML-based lightning nowcasting models

Significant damages caused by lightning can be avoided by predicting lightning and taking precautionary measures. Here, a machine learning model is developed to nowcast lightning flashes using dew point temperature, precipitation, wind speed, wind direction, and previous lightning flashes. The model is trained using data from seven weather stations in Switzerland. The model demonstrates promising performance, achieving an F1 score ranging from 0.70 to 0.76. The main objective is to assess the portability of nowcasting models, by training them in one location and evaluating them in another. The goal is to develop a Machine Learning model that can be applied in regions lacking historical atmospheric measurements. The study reveals that the models generally perform well, with only a minor drop in performance in most cases (6% drop in the F1 score). However, in two cases involving mountainous terrain and tall structures, a significant drop is observed when a dataset was tested with models trained on other regions. Based on the findings, we recommend that for regions with complex topography (e.g., mountainous terrain) and/or tall structures (e.g, wind turbine parks), lightning nowcasting models should be trained on region-specific data, rather than relying on general-purpose forecasters.