A weather forecasting system based on artificial intelligence from DeepMind has managed to do accurate rainfall forecasts with a period of one or two hours in advance rather than a few days as current numerical prediction models do.
The numerical weather prediction models are currently able to offer planetary.scale predictions several days in advance, but they fail to make predictions in short periods of time, of a few hours.
Knowing precisely if it is going to rain in an hour or two could help to better organize day.to.day life, and even has a direct impact on water management, agriculture, aviation, emergency planning and outdoor events.
This obstacle is the one that the DeepMind team, in collaboration with the UK National Weather Service, seeks to improve to offer accurate predictions in less than two hours.
Immediate forecasts are possible thanks to radar data and in combination with the machine learning. The adopted approach employs statistical, economic and cognitive measures in a generative model that makes predictions based on previous radars.
“Our model produces realistic and consistent predictions spatiotemporally on regions of up to 1,536 km x 1,280 kilometers and with delivery times between 5 and 90 minutes in advance”, they explain in the text published in Nature.
The researchers say that the generative model has been ranked first in the evaluations of more than 50 meteorological experts for its “accuracy and usefulness in 89 percent of the cases versus two competitive methods “.