Engineers at Google's DeepMind Have Made a Major Advance in Weather 'Nowcasting'

Meteorologists are not limited to forecasting tomorrow's weather.
Weather models are often asked to forecast the weather conditions for the next hour, also known as "nowcasting". DeepMind, a Google-backed artificial intelligence company, has made significant improvements in precipitation nowcasting accuracy.

You realize how crucial short-term weather forecasts can be. While you might just take your umbrella, they are vital for planning large events, managing transport services, dealing in natural disasters, maintaining farming, and many other tasks.

When compared to two other methods, the new model is ranked first in accuracy and usability in 89 percent. The new model uses generative modeling, a type that applies machine learning to create new data points from existing data.

The new model, called DGMR (Deep Generative Model Of Rainfall), has the primary task of predicting the probability of precipitation over the next one to 2 hours. It has received approval from more 50 meteorologists at UK's Met Office.

This system creates short "radar movies". (DeepMind).

DeepMind Nowcasting Team writes in a blog that "This collaboration between environmental sciences and AI focuses upon value for decision-makers opening up new avenues to nowcast rain and points out the opportunities for AI to support our response to challenges of decision making in an environment constantly changing."

Researchers behind DGMR described creating short "radar movies" to generate future radar patterns using past radar patterns. This is the type of extrapolation machine learning excels at (for more generative models see the creation fake faces).

Numerous current tools for nowcasting, including pySTEPS make use of numerical forecasting (NWP). These models use mathematical models to predict what the future will look like based on current conditions. These models are very powerful, but they are more accurate in the long term.

"These models are really impressive from six hours to about two weeks in weather prediction," Suman Ravuri (staff research scientist at DeepMind) told The Guardian.

DGMR aims at harnessing the enormous number-crunching power artificial intelligence and removing some bluriness or vagueness from existing machine-learning-driven forecasting models. This includes U-Net models that can struggle to maintain accuracy in all parts of the process.

The UK's actual meteorologists have given the DeepMind model their approval. Now, the researchers can begin to integrate it with existing weather prediction systems.

It's not yet possible to predict with 100% certainty whether you will be able stay dry for the next few hours. DeepMind is working to improve the accuracy of DGMR.

The team writes that "No method is perfect and there are limitations to it all. More work is needed in order to improve long-term prediction accuracy and accuracy for rare and intense events."

"Future work will require that we develop additional methods of evaluating performance and further specialize these methods for specific applications in real-world situations."

Nature published the research.