Huawei Cloud team says AI model produces faster, more accurate weather forecasts
- Pangu-Weather found to be ‘10,000 times faster’ than the most powerful forecasting tool currently in use, according to researchers
- It can come up with forecasts in seconds, including humidity, wind speed, temperature, sea level pressure and disaster warnings
They say the model, called Pangu-Weather, can produce forecasts in seconds, including humidity, wind speed, temperature, sea level pressure and even disaster warnings.
02:26
Record temperatures expected globally in 2023 as El Nino weather pattern returns
In a paper in peer-reviewed journal Nature on Wednesday, the team said their experiments had shown the model was “10,000 times faster” than the one used by the European Centre for Medium-Range Weather Forecasts (ECMWF) – the most powerful weather forecasting tool currently in use.
Pangu-Weather focuses on medium-range forecasting. Unlike nowcasting, which provides a detailed description of the current weather across local areas, medium-range forecasting covers a period of up to two weeks and plays an important role in disaster prevention.
Until now, AI-based forecasting has been inferior to numerical weather prediction for medium- and long-term forecasts. That method represents weather conditions as little squares on a map and uses maths to figure out how they change from one state to another. But it is computationally expensive and time consuming.
“When a problem involves complex mathematical mechanisms, large amounts of data and it is hard for people to find specific solutions, we think this problem is very suitable for AI,” Dr Tian Qi, lead scientist on the Huawei Cloud project, said in an earlier statement.
He said meteorological research was a great field for AI to play a role in.
World swelters to unofficial hottest day ever Monday then gets hotter Tuesday
One new algorithm developed for the model adapts to the Earth’s coordinate system to handle complex and uneven 3D weather data. The team also used new computational strategies with less iterations to cut cumulative errors.
To train the model for output at specific time intervals, the researchers used 100 epochs (cycles) with hourly samples of weather data from 1979 to 2021. Each of the sub-models had 16 days of training on 192 V100 graphics cards.
Huawei Cloud said in a statement on Thursday that Pangu-Weather could “now complete 24-hour global weather forecasts in just 1.4 seconds on a single V100 graphics card” – 10,000 times faster than traditional numerical prediction.
While other AI models have been mostly unable to predict extreme and unusual weather such as typhoons, Pangu-Weather achieved more accurate forecast results in testing compared with the ECMWF system.
“Our algorithm achieved high accuracy in tracking 88 named tropical cyclones in 2018, including some (for example, Typhoon Kong-rey and Typhoon Yutu) that remain a challenge for the world’s best tracking systems,” Tian said in the paper.
“This shows Pangu-Weather not only produces strong quantitative results but also preserves sufficient details for investigating certain extreme weather events.”
The China Meteorological Administration has confirmed that it used Pangu-Weather to successfully forecast, five days before, the path of Typhoon Mawar off the east of Taiwan in May. The typhoon, which moved over the Pacific, was the most powerful tropical storm seen so far this year.
Elon Musk says he has faith in China’s ability to develop artificial intelligence
Huawei Cloud unveiled a platform for AI models – called Pangu Models 3.0 – at a developer conference held by the tech giant in the southern Chinese city of Dongguan on Friday. It includes a natural language processing model that could be used for marketing, a computational vision model for image detection, as well as Pangu-Weather, a scientific computing model that the company says has applications in weather forecasting but also medicine design.
“We aim to use a large model as a foundation to grow other applications on top of it,” Tian said in the statement. “If we can solve one problem, we can quickly apply the solution to similar problems elsewhere.”
