'Fengwu' meteorological model scientist team: Using AI to predict extreme weather in the future is not a dream!

[Global Times Report Reporter Zhao Yusha] Recently, a series of extreme weather events have occurred around the world, seriously endangering human production and life. Countries are increasingly demanding more accurate and efficient mid- and long-term weather forecasts. Recently, the Shanghai Artificial Intelligence Laboratory, in collaboration with the University of Science and Technology of China, Shanghai Jiao Tong University, Nanjing University of Information Science and Technology, the Institute of Atmospheric Physics of the Chinese Academy of Sciences and the Shanghai Central Meteorological Observatory, released the large-scale global medium-term weather forecast model "Fengwu". A "Global Times" reporter recently conducted an exclusive interview with the team of scientists behind the "Fengwu" model, and listened to their detailed description of the operating principles behind "Fengwu" and how the AI ​​model will change weather forecasting.

Leading in accuracy and timeliness

According to the "Global Times" reporter, the Shanghai Artificial Intelligence Laboratory AI for Earth joint team built a large model of "Fengwu" based on a multi-modal and multi-task deep learning method. Since it does not require complex physical system simulation, the AI ​​weather forecast model breaks through the computational bottleneck of traditional forecasting methods, so it can forecast and integrate efficiently. AI has a strong ability to adapt to meteorological data relationships and has the potential to break through the performance bottlenecks in traditional numerical model forecasting.

According to Ouyang Wanli, a leading scientist at the Shanghai Artificial Intelligence Laboratory, "Fengwu" is named after the "Xianfeng Tongwu" in the Qin and Han Dynasties, and is the world's earliest wind measurement equipment.

Weather forecasting is a very complex system, starting from the collection of data by the country's Fengyun satellites, weather stations, etc., to data sorting, quality control, and then to atmospheric assimilation, that is, processing the data and establishing the atmosphere according to the needs of the forecast model. status, and finally forecasting and post-processing.

Bai Lei, a young scientist at the Shanghai Artificial Intelligence Laboratory, told the "Global Times" reporter that the "Fengwu" model mainly focuses on the forecasting link, using data obtained from atmospheric reanalysis to train "Fengwu" and then obtain More accurate weather forecasts.

Use artificial intelligence to analyze the elements provided by atmospheric assimilation, including wind speed, temperature, humidity, etc., and create a large model called "Fengwu" for the purpose of predicting future weather. By utilizing past meteorological factors, such as temperature, artificial intelligence can make weather predictions and achieve good prediction results. "Ouyang Wanli explained.

How to improve the timeliness and accuracy of weather forecasts has always been a key topic in the industry. As global climate change continues to intensify, leading to a gradual increase in extreme weather events, expectations for the accuracy and timeliness of weather forecasts continue to increase. Among meteorological and climate forecast tasks, global medium-term weather forecast is one of the most important prediction tasks, and it is also relatively difficult. Due to the accuracy of meteorological observations, the complexity of physical processes in the atmospheric system, and the sheer scale of resources required to solve atmospheric models, the effectiveness of global medium-range weather forecasts has only improved by one day every decade over the past few decades.

Ouyang Wanli said that the advantage of the "Fengwu" large model is that it can reliably predict key meteorological elements for more than 10 days. According to Bai Lei, the effective forecast period of HRES, the best physical model in the world, is 8.5 days, while "Fengwu" reaches 10.75 days based on reanalysis data. "It can be said that the effective forecast period of 'Fengwu' has exceeded the previous longest period." Good physical model".

According to the Shanghai AI Laboratory, "Fengwu" surpassed GraphCast, a large meteorological model released by Google's DeepMind, in 80% of the evaluation indicators. In terms of forecast accuracy, compared with GraphCast, the 10-day forecast error of "Fengwu" is reduced by 10.87%, and compared with the traditional physical model, the error is reduced by 19.4%.

Can be used to predict extreme weather

For extreme weather, do AI large models have better solutions? According to the Global Times, the Shanghai AI Laboratory is exploring the use of "Fengwu" to predict extreme weather such as typhoons, and has achieved certain results so far. At the just past 2023 World Artificial Intelligence Conference, Bai Lei said that "Fengwu" made an accurate trajectory forecast for this year's "Mawa" typhoon based on the initial business field.

"This shows that although extreme weather has occurred frequently in recent years, the underlying laws of the entire atmosphere are still similar, and some of these common laws can be discovered and learned by AI." Bai Lei said. However, it is still necessary to fine-tune the model in response to meteorological changes, continuously use more data to improve the effectiveness of the algorithm, and improve the model in a targeted manner to further improve its ability to predict extreme weather.

Bai Lei believes that there is still a lot of room for improvement in the use of AI for weather forecasting. “The resolution can still be improved. For example, in the past, we could predict the weather in a district. In the future, we hope to be accurate to a street. We are currently moving towards a higher resolution. Work in a precise and more precise direction."

Broad application prospects

According to meteorological experts, although there are currently some products on the market that can provide weather forecast services for the next 15 days or even longer, the forecast performance of more than 10 days is still very uncertain and cannot meet the standards of effective forecasts. Practice has shown that combining observations, numerical predictions and artificial intelligence can help significantly improve the accuracy of numerical predictions. "Fengwu" has achieved a breakthrough in global weather forecast effectiveness of 10.75 days and has huge commercial application potential.

When asked about which aspects of production and life the "Fengwu" large model will be applied to, Ouyang Wanli said that in addition to the weather forecasts that everyone can usually see on their mobile phones, the large model can also be applied to industrial-level weather forecasts and services. In agriculture, marine, electric power and other industries.

Take the electric power industry as an example. Wind power generation depends on the strength of the wind, and solar power generation depends on whether there is sufficient sunlight. The "Fengwu" large model can assist in more accurate predictions of wind speed, sunshine, etc. In addition, Fengwu’s prediction function can affect grid consumption and power dispatch, and is also affected by meteorological changes.

According to the team, in the future, the "Fengwu" AI meteorological model can complement the traditional physical model. With its excellent performance and accuracy, it can provide more accurate and practical weather forecast information for production and life, helping Digitized weather forecasts provide strong support for agriculture, forestry, animal husbandry, fishery, aviation, navigation and other industries as well as public safety.

Ouyang Wanli added that sometimes meteorological experts may get different predictions from the two models, so "Fengwu" can complement the physical model. The more information provided, the more conducive to accurate meteorological predictions. . He explained that for example, the physical model may predict that heavy rainfall will occur in two certain areas, while the artificial intelligence model predicts that heavy rainfall will cover a third area. At this time, the scope of disaster prevention and reduction can be expanded. If preventive measures are not taken, the losses will be huge, but in comparison, the cost of disaster prevention and mitigation is very low. ”

Ma Zhuguo, a researcher at the Institute of Atmospheric Physics, Chinese Academy of Sciences, has long been focused on research in the field of climate change. As a climate model expert, Ma Zhuguo is well aware of the close relationship between computing power, algorithms and information data processing. In his view, artificial intelligence entering weather forecasting and atmospheric physics application scenarios essentially brings about the integration of big data and other information through computing power and algorithms, providing more new technical support and methods for model models, and then Improve forecast accuracy and efficiency, "but we cannot expect or simply say that artificial intelligence will replace traditional mainstream forecast methods, at least not yet."

Taking the familiar satellite cloud images as an example, Ma Zhuguo explained that in weather forecasting, satellite monitoring is very effective and can intuitively see the trajectory and speed of clouds. Once digital model algorithms are used to replace manual intuitive observations, in essence, The changes brought about are also obvious.

The above is the detailed content of 'Fengwu' meteorological model scientist team: Using AI to predict extreme weather in the future is not a dream!. For more information, please follow other related articles on the PHP Chinese website!