Deep Learning-Based Nowcasting of Damaging Winds

Background:

Each year in the United States, damaging wind events (e.g., derechos, microbursts, tornadoes) cause an exorbitant loss of lives and properties (including power grids). Due to the proliferation of renewable energy developments in recent years, these events are also causing considerable disruption to energy production. Thus, there is an ever-growing need for better forecasting damaging wind events. Currently, mesoscale meteorological models (MMMs) are widely used for short-term wind forecasting. It is well-known in the literature that these models often do not faithfully represent the exact magnitude and precise location of damaging wind events. Some improvements can be achieved using finer spatial resolution and more advanced physical parameterizations [1-2]. These numerical strategies are computationally very expensive and require considerable high-performance computing resources. With the advent of advanced deep learning (DL) approaches (e.g., Fourier neural operators, vision transformers, graph neural networks), we envisage that the time is opportune to develop Next-Gen MMMs, which are orders of magnitude faster than contemporary MMMs (e.g., the WRF model) and offer at-par or better accuracy. Our proposed project is a first step in this direction.

Industry Relevance:

If successfully developed, in the near future, it may be possible to run a trained, DL-based MMM on a basic workstation instead of a high-performance computer. If our speculation becomes a reality, then an engineer working for an electric utility company in the northeast can accurately forecast highly localized extreme wind conditions and ensure grid reliability. We believe that such democratization of MMMs will aid in partially mitigating the imminent threats of climate change and associated perils.

Objectives:

In the proposed WISER project, we will develop a novel DL-based extreme wind nowcasting model (henceforth DeepExWind). This model will produce reliable damaging wind forecasts for a temporal horizon of up to 3 hours. Even though the training of the DeepExWind model will be computationally expensive, the actual forecasts will be performed on a basic workstation in a couple of minutes time.