AI-powered forecasting proves a hit

Huawei's artificial intelligence-powered Pangu weather prediction system is making waves with its potential to revolutionize weather forecasting, with high-resolution global forecasts for locations roughly 27 kilometers apart generated in under 10 seconds.

Tian Qi, the leader of Pangu's research and development team, said it uses neural network models for weather forecasting systems and achieves higher prediction accuracy than the world's first similar AI weather forecasting model, Four-CastNet, which was released by Nvidia in 2022.

The World Bank says global weather forecasting may generate economic benefits worth $162 billion a year. Research from the China Meteorological Administration indicates that approximately 40 percent of China's GDP is related to weather and climate.

One example, Tian said, is wind power generation, where reducing wind speed forecast errors by half a meter per second could save the nation economic losses of 23.25 billion yuan ($3.22 billion) a year and reduce carbon dioxide emissions by 25 million metric tons.

"Accurate weather forecasting is of significant importance for wind power generation, precipitation forecasting, earthquake disaster reduction and guiding agricultural production," Tian said, adding that the breakthrough can become a crucial driving force for the advancement of new quality productive forces.

New quality productive forces refer to advanced productivity that breaks free from traditional economic growth modes and productivity development paths by using cutting-edge technology to improve efficiency and quality.

Tian said numerical weather forecasting theory originated in the early 20th century. It predicts weather by solving mathematical and physical equations that describe atmospheric motion.

The continuous improvement in computer processing speed in recent decades saw forecast lead times extended from one day to five to seven days, and forecast resolution decreased from several hundred kilometers to just a few kilometers.

However, in recent years, the traditional forecasting method has encountered bottlenecks. According to data from the European Centre for Medium-Range Weather Forecasts, three- to seven-day forecast errors for several meteorological elements decreased by less than 5 percent between 2012 and 2022.

Xie Lingxi, a senior researcher in Tian's team, said research in the field of AI weather forecasting began in both China and the United States around 2016.

"However, at that time, the resolution was still very low, probably only about one-tenth of what it is now," he said.

"The traditional weather forecasting method can extend the forecast time by about one day every 10 years. The Pangu model has extended the time by 0.6 days while achieving the same level of accuracy as traditional forecasts."

Accelerated improvement could be expected, Xie said, because the "evolution speed of artificial intelligence methods is often much faster than that of traditional methods".

"There are physical equations for atmospheric variables," he said.

"AI, on the other hand, does not rely on these equations but instead learns from massive historical data to construct deep neural networks for expressing a completely different complex mathematical function.

"Unlike traditional methods, the computational units in neural networks often do not have interpretable physical meanings."

Since the completion of the Pangu model in November 2022, it has garnered widespread attention from institutions such as the European Centre for Medium-Range Weather Forecasts and the China Meteorological Administration.

In the real-time forecasts for typhoons Doksuri and Saola last summer and autumn, Pangu played an important role because it was able to make judgments several days in advance regarding important events such as changes in direction and landfall.

Xie said that before Typhoon Saola was named, Pangu had already predicted that it would circle counterclockwise in the eastern waters of the Philippines and then make landfall on the southern coast of China.

"The path of Typhoon Doksuri was peculiar, skirting past the Philippines and the southern part of Taiwan before making landfall in Fujian province," he said. "Pangu made precise predictions for both."

In July, the European Centre for Medium-Range Weather Forecasts launched the Pangu weather prediction model on its website as part of its daily weather forecasting suite, providing it for free to the world.

"This means that weather enthusiasts worldwide as well as meteorological institutions in the least developed countries will benefit from the artificial intelligence weather forecasting model, obtaining more real-time and accurate weather forecasts and extreme weather warning results," the World Meteorological Organization said last year.