Where exactly will Kanu go? Predicting Typhoon with AI, a headache for forecasters | Typhoon | Kanu
On August 1st, according to the Central Weather Bureau, Typhoon Kanu, the 6th typhoon of this year, will move northwest at a speed of about 20 kilometers per hour. It is expected to move into the East China Sea on August 2nd. On the afternoon of August 31st, Typhoon Kanu has strengthened to a super typhoon level.
Causing headaches for forecasters and confusion for "supercomputers", what concerns everyone is that there is still a great deal of uncertainty in the path of "Kanu", whether it will "move north" or not? Or come to Zhejiang? Recently, the reporter interviewed Zhang Junxi, a senior research specialist at the Intelligent Supercomputing Research Center of Zhijiang Laboratory, to explain how the typhoon path, which is currently focusing on global attention, has been analyzed? How to empower meteorological prediction with "supercomputing+AI" technology to make typhoon paths more accurate.
"Super Brain" calculates typhoon positioning
Typhoon is one of the most important extreme weather disasters affecting China, developed from low-pressure eddies on the tropical or subtropical ocean surface. I visit the coastal areas of China almost every year.
According to Zhang Junxi, the current mainstream typhoon prediction method has widely used supercomputers. By solving equations, the "eye positions" at different time points in the future are solved, connected into lines, and the commonly seen typhoon path prediction maps are drawn.
"Typhoon prediction requires collecting comprehensive meteorological data from specific regions, solving complex control equation systems, and determining the future atmospheric state for a period of time. It also combines meteorological parameters such as minimum sea level atmospheric pressure and wind speed to locate typhoon eye and predict typhoon path." Zhang Junxi introduced that effective forecasting requires a very high computational load, generally achieving floating-point calculations of over 100 billion times per second. For large-scale and high-precision calculations, it is also possible to reach the level of tens of megabytes per second.
"Supercomputing+AI", knowing the future weather in the blink of an eye
With the help of supercomputers, currently, it usually takes several hours of simulation and deduction to predict weather results for three days. With the empowerment of AI technology in weather prediction work, for a specific prediction requirement, after the calculation model is trained, AI can usually complete the inference task in a few seconds. Recently, a research paper published by Nature showed that the meteorological AI large model can predict key meteorological parameters for 5 days in just 1.4 seconds, and the weather for the next 5 days can be known in the blink of an eye.
What is the principle? Zhang Junxi told reporters that traditional technology converts physical laws that humans have already recognized into mathematical equations for solving, which is relatively slow. The AI model, on the other hand, can be completely based on existing observation data and reanalysis data established by combining the model without relying on physical laws. AI can search for patterns in massive big data by itself. Although the training speed of AI models is also relatively slow, the speed of further inference after training is extremely fast. Meanwhile, existing research has found that AI meteorological predictions may have slightly better accuracy in some scenarios.
"At present, AI has been involved in the prediction of typhoon paths. Generally, more than one path analysis is generated through technology. Finally, meteorological experts refer to various physical models and AI simulation results, and use their own experience to make comprehensive judgments before publishing." Zhang Junxi said.
