J8.5 Using Machine Learning to Improve 10-m Wind Forecasts in the Winter – Toward an Improved WSSI Forecast

Tuesday, 30 January 2024: 5:30 PM
Johnson AB (Hilton Baltimore Inner Harbor)
Marshall Romanick Baldwin, CAPS, Norman, OK; and N. A. Snook, K. A. Brewster, and P. Spencer

The Winter Storm Severity Index (WSSI) describes anticipated overall impacts to society due to winter weather. Forecasting this index has the potential to assist both emergency managers and the general public in prioritizing response efforts, minimizing risks, and making informed decisions ahead of winter hazards. The NOAA WPC WSSI is calculated by taking the maximum of 6 sub-indices that quantify the expected severity of several winter weather hazards. Of these sub-indices, the blowing snow, ground blizzard, and ice accumulation indices all consider wind speed. Therefore, improving the accuracy of wind speed forecasts associated with winter weather events could result in improved WSSI forecasts.

We will present preliminary results from an ensemble of machine learning (ML) wind speed forecasts generated from a CAM ensemble. The ML is trained on and uses 2-dimensional inputs from each member of a 12-member combined NWP ensemble consisting of 4 FV3-LAM member forecasts from an ensemble produced by the Center for Analysis and Prediction of Storms (CAPS) for the Winter Weather Experiment and 8 members of the operational HREFv3 from the winters of 2020-21 and 2021-22. NWP inputs used include horizontal 2D fields of 10m and isobaric wind speeds, surface pressure, 10m temperature, isobaric geopotential heights, and simulated reflectivity. Additional derived features, such as gradients of the aforementioned fields, are also included. Static fields, including land use, time of day, and orography, are also used as inputs for the ML ensemble. Labels for ML training are obtained from the NOAA Unrestricted Mesoscale Analysis (URMA) 10m wind speed field.

Given the spatial structure of this data, we chose to use a U-Net architecture for ML wind speed forecasting; a U-Net is a deep learning approach utilizing a convolutional neural network designed to identify spatial patterns in images. The ensemble of U-Nets is trained to predict wind speed on 64x64 pixel patches for each ensemble member. The patches are then stitched together to create the forecast over the entire Contiguous United States (CONUS) domain. WSSI forecasts generated using U-Net predicted wind speed are compared to those generated using raw wind speed predictions from the corresponding NWP ensemble members to measure any improvement in forecasting skill as verified by URMA winds.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner