Improving medium-range extreme weather predictions using AI NWP models and AI-based postprocessing
Drs.Yingkai Sha and Ryan Sobash
NCAR
4 June, 2pm
Abstract:
This presentation will cover two different projects both related to improving medium-range
high-impact weather predictions using AI-based post-processing methods. First, we will highlight the development of a generative-AI based system to improve probabilistic forecasts of extreme precipitation events across the conterminous United States from Days 1 – 8. The method utilizes a Latent Diffusion Model to post-process 6-hourly Global Ensemble Forecast System (GEFS) precipitation forecasts and produce an expanded generative ensemble up to 500 members in size. This generative ensemble is found to improve the characterization of extreme precipitation events up to day 6, and outperforms the raw operational GEFS as well as a second non-AI statistical baseline based on analogs. Additional work was undertaken to investigate the decision-making process of the method, including performing explainability studies which provide insight into forecast performance. Second, we will highlight work using data-driven AI NWP models to improve forecasts of convective hazards from Days 1–8. Specifically, we used the Pangu-Weather AI NWP models to produce synoptic and meso-scale predictions out to Day 8 with both GFS and ECMWF initial conditions. A dataset of 5 years of forecasts (2019–2023) was used to train decoder-only transformers to produce probabilistic predictions of convective hazards, such as tornadoes. Validation of these forecasts during 2024 showed that they outperformed forecasts based on the operational GFS, especially at medium-range lead-times (e.g., Days 3–7). Together, these two studies highlight the potential to extend the practical predictability of rare, high-impact weather events into the medium-range lead-times using AI-based post-processing and data-driven NWP techniques.