Using a Neural Network as an Improved Method of Predicting Wind Gusts in the HRRR

The High-Resolution Rapid Refresh (HRRR) numerical weather prediction model is run operationally by the National Weather Service to provide 1 to 36 hour forecasts for a range of different variables. One of these variables is wind gust across the conterminous United States. The current approach for predicting wind gusts in the HRRR uses an algorithm that weights the wind speed excess at different levels in the boundary layer relative to the surface to estimate the wind gusts at the 10-m level. While this approach considers the vertical wind profile in the boundary layer, it does not take into account any of the local features such as terrain. When compared to the measurements on site, the current HRRR algorithm tends to skew toward overprediction and has a RMSE of approximately 5 m/s. The aim of this project was to improve the wind gust predictions using a neural network. To do this, a two-layer regression network was trained on METAR gust measurements and the measurements of other involved variables: temperature, pressure, precipitation, wind speed, terrain features, etc. These training points were taken at a variety of seasons and times of day. The neural network was then tested using the complementary variables in the HRRR model, as if the neural network were simply part of the model. The gust predictions were evaluated against the METAR measurements, and have returned lower rates of error.