Photo: VCG

The world suffered record-breaking climate disasters in 2023. The frequent extreme weather-related disasters have highlighted an urgency for weather forecasts to be more precise. In China, where the artificial intelligence (AI) is booming fast, scientists are racing to develop models that provide accurate weather forecasts to reduce the cost of climate disasters.

When government officials, scientists, companies and climate organizations gathered at COP28 climate change conference in Dubai, United Arab Emirates at the end of last year, a new model for sub-seasonal forecasts using AI technology was unveiled by a Chinese scientist.

Important value

The model, "FuXi-Subseasonal," developed by scientists from the Shanghai Academy of Artificial Intelligence for Science (SAIS), Fudan University, and China's National Climate Center, has a thousand-fold increase in operational speed, and higher forecasting accuracy and longer forecasting period than existing international authoritative models, according to the team.

Climate disaster warning is another important value of this FuXi-Subseasonal model, Qi Yuan, who is in charge of the research team told media. He said his team has significantly increased the prediction period for extreme weather from 30 days to 36 days, predicting potential climate disaster events as early as possible, and gaining more time for response and mitigation measures.

The model of "FuXi-Subseasonal" represents one of China's mushrooming AI models used for predicting extreme weather.

When typhoon Doksuri hit China last July, Fengwu, a machine learning model developed by the Shanghai Artificial Intelligence Laboratory, surpassed European and American equivalents in predicting its moves.

Between July 21 and July 27 last year, Fengwu's forecast of the tropical cyclone's path was just off by 38.7 kilometers on average, whereas the corresponding data were 54.11 km for a model of the European Center for Medium-Range Weather Forecasts and 54.98 km for that of the US National Centers for Environmental Prediction, Global Times learned from the lab. 

Reducing error by one km in 24 hours can lower about 97 million yuan ($13.54 million) in direct economic loss, so accurate typhoon forecasting is vital in minimizing risks, researchers from the lab said.

Scientists told the Global Times that meteorological forecasting is a highly complex system that involves collecting data from national weather satellites, meteorological stations and other sources. The data is then sorted and subjected to quality control before undergoing atmospheric assimilation, where it is processed and used to establish the atmospheric state required by the forecasting model. Finally, the forecast is made and post-processing is conducted.

Bai Lei, a scientist at the Shanghai Artificial Intelligence Laboratory, explained that the "Fengwu" model focuses primarily on the forecasting stage. It utilizes data obtained from atmospheric reanalysis to train the model and obtain more accurate weather forecasts.

AI models such as "Fengwu" use artificial intelligence to analyze the elements provided by atmospheric data assimilation, such as wind speed, temperature and humidity in order to predict future weather. Artificial intelligence can utilize past meteorological elements, such as temperature, to forecast future weather and achieve more precise results, explained Ouyang Wanli, another scientist from the Shanghai lab.

Unlike the traditional physical models that mostly run on supercomputers, "Fengwu" only needs single graphics processing unit to generate high-precision global weather forecasts for the next 10 days in 30 seconds.

Chinese leading technology in using artificial intelligence to predict weather has also gained worldwide recognition. The European Centre for Medium-Range Weather Forecasts, last year launched Pangu Weather, a model developed by research team from Huawei, capable of predicting global weather on its website, South China Morning Post reported.

It said that the collaboration has seen the Chinese tech company translate the science behind the weather into practical applications, while the European weather agency has begun to embrace AI in its daily forecasts.

According to Huawei, the prediction accuracy of the Pangu model from 1 hour to 7 days has exceeded the prediction accuracy of some meteorological centers in Europe and the US in the same time span, Xinhua News Agency reported.

As Typhoon Doksuri approaches, dark clouds gather over the sky of Shanghai on July 26, 2023. Photo: IC

Provide an alternative

In July last year, the China Meteorological Administration issued a work plan aiming to accelerate the construction of artificial intelligence meteorological application technology system. The plan specifies that by 2025, a roadmap for artificial intelligence meteorological application development will be established and that by 2030, the development level of artificial intelligence meteorological applications will be at the forefront of the world.

Due to its own limitations and the uncertainty of weather, the traditional model of weather forecast still cannot meet the diverse and growing needs of today's users; whilst data-driven AI methods provide very useful tools to bridge this gap, Dai Kan, deputy head of China's National Meteorological Center told the Global Times in a previous interview. 

Local governments in China have already started exploring various ways of using artificial intelligence in weather forecasting.

Currently, the Guangdong Provincial Meteorological Bureau has gained good results in short-term precipitation forecasting based on deep learning using the Alibaba platform. The Beijing Municipal Meteorological Bureau has also applied machine learning methods to temperature forecasting. The Fujian Provincial Meteorological Bureau has promoted the application of machine learning-based objective correction methods for precipitation elements in multiple provincial meteorological bureaus.

However, scientists believe there is still room for training artificial intelligence to improve their weather forecasting abilities.

Dai noted that the current AI technology is mainly focused on short-term weather forecasting, but there is still a lack of support for the entire chain of weather forecasting business, including data quality control and multi-disaster weather early warning capabilities.

Moreover, the field is dominated by professionals with backgrounds in atmospheric science, yet scientists with backgrounds in statistics, computer science and big data mining are also needed to provide more comprehensive translation of the data, he said.