Google-affiliated company DeepMind greatly improves the accuracy of weather forecasts up to 90 minutes ahead, the accuracy recognized by weather forecasters –GIGAZINE

By deep learning, with Alphabet similar to Google as parent companystructural analysis of proteinsOrThe strongest Go AI in the worldAn artificial intelligence company also known for developingdeepmindHowever, a deeper generative model that can predict with high accuracy the probability of precipitation up to 90 minutes ahead.DGMRannounced to be developed. Predicting climate change within two hours is considered the most difficult problem in weather forecasting, and it is expected that with the introduction of this model, the accuracy of weather forecasts will improve significantly.

Efficient Rain Now Casting Using Radar’s Deep Generative Model | Nature
https://www.nature.com/articles/s41586-021-03854-z

Nowcasting the next hour of rain | deepmind
https://deepmind.com/blog/article/nowcasting

In modern weather forecasting, the movement of fluids in the atmosphere is calculated numerically to predict future weather.numerical weather forecastIs used for. Numerical weather forecast is good for predicting weather from 6 hours to about 2 weeks later, but it is said to be less accurate when forecasting weather within 2 hours.

DeepMind co-leader Suman Ravuri and her colleagues have created a deep generative model “DGMR” to improve the accuracy of so-called “short-term forecasts” in under two hours. Used for image generation, etc.Generative Adversarial Network (GAN)The same algorithm as above was designed to learn the motion of rain clouds captured by weather radar and to estimate and generate the motion of rain clouds 5 to 90 minutes ahead.

To verify the accuracy of the prediction results generated by the DGMR, Ravuri et al designed two existing rainfall probability prediction models and asked 56 weather forecasters to evaluate the accuracy, hiding their names. As a result, the DGMR was rated as “most accurate and useful” in 89% of cases. Of the following, the upper left is the actual cloud movement, and the upper right is the DGMR. Unlike PySTEPS (bottom left), a convective approach where precipitation intensity is often very high, and deep learning UNet (bottom right), where simulation results can be blurry, DGMR balances the intensity and extent of precipitation potentials . , you can see that it is close to the actual observation record.

“We see this as an exciting area of ​​research, and our research will serve as the basis for new research, facilitate the integration of machine learning and environmental science, and make better decisions in weather forecasting,” the development team said. Hope it will be supported.”