Accurate weather predictions are essential for modern economies, such as farming and shipping, as well as being required by governments (such as issuing flood warnings or alerting the populace to a forthcoming hurricane), as well as being convenient for the everyday person.
Consequently, considerable research goes into meteorology and this research is being advanced through the application of deep learning.
In a new study, researchers from Rice University have shown how capsule neural network technology can be used to predicts extreme weather with analog forecasts (perhaps a little ironic in the digital age, but it worked).
With this experiment, by applying deep learning scientists developed a system that was taught how to predict extreme weather events, such as heat waves, up to five days in advance.
These predictions were made using minimal information about current weather conditions. To train the algorithm, the system processed hundreds of pairs of maps. Each map indicated the surface temperature and air pressure at five-kilometers height. With the pairs concept, each map was of an identical area, however the temperature and pressure conditions were several days apart.
The exercise included weather patterns that were associated with extended hot and cold spells (of the types associated with extreme and often dangerous weather scenarios, like heat waves and winter storms).
The objective was to understand more about the physics and precursor conditions that lead to extreme-causing weather patterns, and which are a limiting factor with current processes for weather forecasting.
Following the training, the algorithm could examine maps it had not previously been shown and could produce five-day forecasts to indicate the likelihood of an extreme weather event, with up to with 85 percent accuracy.
The development of the deep learning tool has been reported to the Journal of Advances in Modeling Earth Systems. The research paper is titled “Analog forecasting of extreme‐causing weather patterns using deep learning.”