Machine learning could improve extreme weather warnings
A deep learning technique could reduce the error in 10-day weather forecasts by more than 90%, allowing communities to better prepare for extreme events such as heat waves.
Because small changes in atmospheric and surface conditions can have large, difficult-to-predict effects on future weather, traditional weather forecasts are released only about 10 days in advance. A longer lead time could help communities better prepare for what's to come, especially extreme events such as the record-breaking , which melted train power lines, destroyed crops, and caused hundreds of deaths.
Meteorologists commonly use adjoint models to determine how sensitive a forecast is to inaccuracies in initial conditions. These models help determine how small changes in temperature or atmospheric water vapor, for example, can affect the accuracy of conditions forecast for a few days later. Understanding the relationship between the initial conditions and the amount of error in the forecast allows scientists to make changes until they find the set of initial conditions that produces the most accurate forecast.
However, running adjoint models requires significant financial and computing resources, and the models can measure these sensitivities only up to 5 days in advance. tested whether a deep learning approach could provide an easier and more accurate way to determine the optimal set of initial conditions for a 10-day forecast.
To test their approach, they created forecasts of the June 2021 Pacific Northwest using two different models: the GraphCast model, developed by Google DeepMind, and the Pangu-Weather model, developed by Huawei Cloud. They compared the results to see whether the models behaved similarly, then compared the forecasts to what actually happened during the heat wave. (To avoid influencing the results, data from the heat wave were not included in the dataset used to train the forecasting models.)
The team found that using the deep learning method to identify optimal initial conditions led to a roughly 94% reduction in 10-day forecast errors in the GraphCast model. The approach resulted in a similar reduction in errors when used with the Pangu-Weather model. The team noted that the new approach improved forecasting as far as 23 days in advance.