Hail, tornadoes and high winds can become some of the main dangers of a storm , as they can cause serious damage, coupled with an even more important problem: they are difficult to predict accurately for different models existing weather conditions.
However, thanks to the work carried out by the National Center for Atmospheric Research (NCAR), this problem could have come to an end. And it is that the researchers are using artificial intelligence for the execution of much more effective and, above all, much more precise experimental forecasts.
It would be a neural network capable of forecasting meteorological hazards by convection, providing the probability of tornadoes, hail or strong winds for a particular area at a given time, whose data is updated twice a day and, in addition, is available from completely free online.
As part of an experiment conducted at the National Oceanic and Atmospheric Administration (NOAA), NCAR scientists began running these weather research forecasts in the spring of last year (2020). Then, after the severe spring weather temperature ended in late June, they analyzed the forecasts, and then compared them with more traditional techniques used to forecast hazards associated with storms.
The researchers found that the neural network was able to significantly improve the accuracy of forecasts based on the output of the traditional model , especially in situations where such traditional forecasts tended to perform worse, even for night storms. and for the eastern and western parts of the United States.
According to the experts, “the neural network not only predicted more skillfully when and where severe storms were most likely to occur, but was also able to better predict whether the dangerous event would be dominated by winds or hail.” What’s more, the system made it possible to achieve a significant improvement in most situations .
For a given meteorological model to be capable of generating a thunderstorm, it needs to be run at a resolution high enough to capture fine-scale atmospheric phenomena, including updrafts and downdrafts, which act to drive the creation of the storm. storm. This generally requires the existence of a space of 4 kilometers or less between the grid points within the model, allowing it to be able to simulate the storm itself, but it cannot produce many of the associated hazards. such as tornadoes and hail, which occur on even smaller scales.
Until now, meteorologists relied on particular results in the model data, or surrogates, in order to determine the probability that a given severe storm could produce such hazards.
But the developed neural network can absorb around 40 different factors , including storm location, dew point, weather, wind speed, surface pressure and, most notably, the helicity of the storm. the updraft.
The neural network then uses patterns in the way those predictors are related to each other , after the system was trained with a large data set (specifically, 500 previous forecasts obtained from the high-resolution fast update model of the network itself. NOAA).
The success of this tool suggests that machine learning could become a truly useful tool for operational forecasting , positively helping to better predict hazards associated with straight-line wind storms or improve forecasts for those. regions where supercells are less likely to form.
