Start
February 23, 2022 - 3:30 pm
End
February 23, 2022 - 4:30 pm
Categories
Weather and Climate SystemsWeather and Climate Systems
Forecasting the Forecast with Deep Learning
Chris Rattray
Wednesday, February 23
3:30 PM
Global numerical weather prediction systems produce occasional significant drops in the accuracy of their forecasts often referred to as “forecast busts†in the literature. Despite recent efforts, understanding the causes of these forecast failures remains challenging, but essential. Ensemble spread is often used as a measure of forecast reliability with large spread expected to be associated with poor forecasts. However, ensemble models are often overconfident when forecast busts occur in deterministic modeling systems. Therefore, new techniques would be helpful to predict the occurrence of forecast busts and to guide forecast centers toward paths for model improvement.
This work explores using neural networks to predict model forecast accuracy and heat map visualization techniques to provide information on which atmospheric flow characteristics will result in forecast busts. Our preliminary efforts have concentrated on using convolution neural networks to detect medium range (i.e., 6-day) forecast busts over Europe using the initial conditions and model associated with ERA-interim reanalysis. The neural network model was trained using the u-wind, v-wind and geopotential heights at the 850, 700, 500 and 200 hPa levels. Our results show promise in detecting forecasts likely to be associated with forecast busts, with accuracy as high as 88%. Early insight into the flow characteristics suggest that an upper-level trough over the western United States and a ridging pattern over the western Canadian Arctic are associated with forecast busts during the warm season. Due to our limited sample size, future efforts will be focused on improving reliability of these results.