Seamless global lightning forecasts based on convolutional graph neural networks

Global lightning forecasts are of particular interest to the aviation industry. However, global weather forecasts are typically based on numerical weather prediction model (NWP) runs with resolutions that cannot resolve deep convection, hence it needs to be parameterized. Although lightning may be diagnosed in these models, e.g. with the Lightning Potential Index (LPI,1), the quality of such forecasts is only mediocre. Successful attempts have been made to produce lightning probability forecasts with neural networks that use the LPI as input features and lighting observations as target (2,3). Moreover, high resolution global lightning data is available (GLD360, Vaisala, 4) which has been used for nowcasting of thunderstorms up to 2 hours lead time (5).

This work shows how the combination of global lightning observations (GLD360 from Vaisala, 4) with NWP can significantly improve the global lightning probability forecast for up to 8 hours. A graph neural network initialized with global lightning observations of the past hours has been developed. Observations are transferred into a latent space by a fully connected neural network and then integrated forward in time. Multiple loops using a convolutional graph neural network are employed to include spatial information. Here, NWP forecasts are fed in as feature and are the driver for the initiation of new thunderstorms, their propagation and decay.

The model is trained with one year ICON Reanalysis data and applied to global ICON ensemble forecasts. Preliminary verification results based on the resolution component of the Brier Score show that initializing the model with global lightning observations of the past hours significantly outperforms a model version without that initialization. This improvement remains for lead times up to 8 hours, then the verification scores of the two model versions converge. Furthermore, case studies show that the lightning is well resolved during the first forecast hours, i.e. beyond the typical 2 hour nowcasting horizon.

References:

(1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: https://doi.org/10.5676/dwd_pub/nwv/icon_010 .

(2) Baumgartner, M., Schröder, G., and Primo, C.: Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-34, https://doi.org/10.5194/ems2023-34, 2023.

(3) Schröder, G., Baumgartner, M., Primo, C., and Theis, S.: Challenges in training neural networks for global lightning probability forecasts, EMS Annual Meeting 2024, Barcelona, Spain, 1–6 Sep 2024, EMS2024-475, https://doi.org/10.5194/ems2024-475, 2024.

(4) https://www.vaisala.com/en/products/systems/lightning/gld360

(5) Müller, R., Barleben, A., Haussler, S., & Jerg, M. (2022). A Novel Approach for the Global Detection and Nowcasting of Deep Convection and Thunderstorms. Remote Sensing, 14(14), 3372. https://doi.org/10.3390/rs14143372