Experimental AI Hones Forecasts For Short-Term Weather Events

AI researchers have developed a system that predicts whether it will rain over the next two hours.

Google sister company DeepMind worked with the University of Exeter on the “nowcasting” technology.

It analyses the previous 20 minutes of high-resolution radar data and comes up with a preipitation forecast for the next 90 minutes, distinguishing between medium and heavy rainfall.

The technique differs from current methods that generally use numerical weather prediction (NWP), driven by mathematical equations that estimate weather based on the movement of fluids in the atmosphere.

Short-term accuracy

Current techniques are generally most accurate for periods between six hours and two weeks into the future.

DGMR (Deep Generative Model of Rainfall) instead uses generative modelling, which produces new data points after being trained on existing ones.

The model’s developers said it makes short “radar movies” predicting imminent weather patterns.

It could be more accurate for short-term weather patterns, including critical storms and floods.

The system was trained using UK radar maps from 2016 to 2018, and when tested on maps from 2019 was found by 50 Met Office meteorologists to be first for accuracy and usefulness in 89 percent of cases.

Weather adaptation

DeepMind said more work is needed to improve the accuracy of long-term predictions and on rare and intense events.

“It’s very early days but this trial shows that AI could be a powerful tool, enabling forecasters to spend less time trawling through ever-growing piles of prediction data and instead focus on better understanding the implications of their forecasts,” said DeepMind senior scientist Shakir Mohamed.

“This will be integral for mitigating the adverse effects of climate change today, supporting adaptation to changing weather patterns and potentially saving lives.”

The research has been published in Nature.