Deep Learning Based Nowcasting of Damaging Winds, Phase II

Background:

Damaging wind events such as derechos, microbursts, and tornadoes cause significant loss of life and property, including disruptions to power grids and renewable energy production. Current forecasting models, primarily mesoscale meteorological models (MMMs), often struggle to accurately predict the magnitude and location of these events. While finer spatial resolution and advanced physical parameterizations can improve accuracy, they are computationally expensive. Recent advancements in deep learning (DL), including Fourier neural operators, vision transformers, and graph neural networks, present an opportunity to develop next-generation MMMs that are both faster and more accurate. This project aims to take the first step toward achieving that goal.

Industry Need:

A successful deep learning-based MMM could run on a basic workstation rather than requiring high-performance computing resources. This democratization of forecasting models would allow utility engineers and grid operators to make highly localized, real-time extreme wind predictions, improving grid reliability and resilience against climate-related disruptions. By reducing computational costs and making advanced forecasting tools more accessible, this project aligns with the broader goal of mitigating climate change risks and enhancing disaster preparedness.

Objectives:

The project aims to develop deepExWind, a novel deep learning-based extreme wind nowcasting model capable of providing reliable damaging wind forecasts for up to three hours. While training the model will be computationally intensive, once trained, it will be able to generate forecasts in just minutes on a standard workstation.

Methodology:

Inspired by recent successes in deep learning for global weather forecasting (e.g., FourCastNet, Pangu-Weather, GraphCast), this project will adapt these techniques to the more complex problem of mesoscale short-term forecasting. Since periodic boundary conditions do not apply at the mesoscale, Wavelet Neural Operators will be used. Training data will include high-resolution simulated datasets (RTMA and CONUS404), gridded radar data (GridRad.org), and observational data from the New York State Mesonet (NYSM). Model performance will be compared against NOAA’s HRRR forecasts. The PI has access to the necessary computational resources (NVIDIA A6000 GPUs) to execute these tasks.

Deliverables:

The deepExWind model will be made publicly available via an open-access repository.

Research findings will also be shared through conference presentations and peer-reviewed journal publications.