AI-Based Meteorological Model Serves Accurate Global Weather Forecasts
The most accurate current forecast system is the numerical weather prediction (NWP) method, although it is computationally expensive. Daily weather forecasts, extreme disaster warnings, and climate change predictions are all realized by the NWP method that relies on high-performance computing and complex physical models.
The conventional NWP method requires four to five hours of calculation on a supercomputer cluster with 3,000 servers to forecast the next 10 days of global weather, said Tian Qi, the corresponding author of the article and chief scientist of AI with China's cloud service vendor Huawei Cloud.
Recently, AI-based methods have shown some potential in accelerating weather forecasting by orders of magnitude. But the forecast accuracy is still significantly lower than that of NWP methods, according to the article.
A large meteorological model research and development team from Huawei Cloud proposed a three-dimensional neural network adapted to the Earth's coordinate system to process complex and heterogeneous three-dimensional meteorological data.
Trained on nearly 40-year global data, the sizable meteorological model Pangu-Weather obtained 100-million-level parameters within two months.
It shows better deterministic forecast results on reanalysis data in all tested variables when compared with the NWP system, the operational integrated forecasting system of the European Centre for Medium-Range Weather Forecasts, according to the article.
The Pangu-Weather takes only 1.4 seconds to complete a 24-hour global weather forecast, including potential humidity, wind speed, temperature and sea level pressure, and other values. Its prediction speed is 10,000 times faster than the traditional numerical methods.
During super typhoon Mawar this May, the Pangu-Weather performed excellently by predicting the turning path five days in advance.
Bi Kaifeng, the first author of the research article, admits the shortcomings of AI-based weather forecasting, saying that it is still highly dependent on reanalysis data and needs to improve the capability of estimating extreme weather.
"We believe that the AI-based methods should coexist with conventional numerical methods to provide more accurate and reliable weather forecasting services," said Tian.
4155
