Huawei Cloud has developed the Pangu Weather weather forecasting system based on artificial intelligence. The meteorological device demonstrates high performance in terms of forecasting speed and accuracy compared to traditional solutions.
On July 5, 2023, the leading international scientific journal Nature published an article about Pangu Weather, a revolutionary weather forecasting device based on artificial intelligence.
The article describes how to develop an accurate and reliable international weather forecasting system based on artificial intelligence and deep learning using 43 years of data.
What is Pangu Weather?
Pangu Weather is the first artificial intelligence-based forecasting device that demonstrates better accuracy than traditional numerical methods. The invention allows to speed up forecasting by 10,000 times, which will reduce the time for weather forecasting around the world to a few seconds. The article, titled “Accurate Medium-Term International Weather Forecasting with 3D Neural Networks,” provides an independent assessment of these capabilities.
Pangu Weather challenges previous assumptions that AI-based weather forecasting is inferior in accuracy to traditional methods. Developed by the Huawei Cloud team, this first AI-powered forecasting model has higher accuracy than previous solutions.
Thanks to the rapid development of computing power over the past 30 years, the accuracy of numerical forecasting has improved significantly. Thus, it has become possible to provide warnings about natural disasters and climate change forecasts. However, this method is still labor-intensive.
To improve the speed of forecasting, researchers have been studying how to apply deep learning methods. However, the accuracy of AI-based forecasts for medium- and long-term scenarios remained lower than that of numerical forecasting. Artificial intelligence was usually unable to predict extreme and unusual weather conditions, such as typhoons.
There are about 80 typhoons in the world every year. According to the Chinese Ministry of Emergency Management, in 2022, direct economic losses due to typhoons in China alone reached 5.42 billion yuan. The earlier warnings of natural disasters can be received, the easier it will be to prepare for them.
Due to their speed, smart weather forecasting devices have become popular, but they lacked accuracy for two reasons.
First, existing AI-based meteorological forecasting systems were built on 2D neural networks, which cannot effectively process uneven 3D meteorological data. Secondly, medium-term weather forecasts can suffer from cumulative forecasting errors when the system is activated too often.
How does Pangu Weather overcome these challenges?
In scientific tests, Pangu Weather has demonstrated higher accuracy than traditional numerical methods for forecasts ranging from 1 hour to 7 days at 10,000 times the speed. The system can predict meteorological phenomena in seconds, including humidity, wind speed, temperature, and sea level pressure.
Pangu Weather uses a 3D Earth-Specific Transformer (3DEST) architecture to process complex, heterogeneous 3D meteorological data. Thanks to a hierarchical temporal aggregation strategy, the model was trained with different forecast intervals – 1-, 3-, 6-, and 24-hour. As a result, the number of iterations for predicting meteorological events at a specific time was significantly reduced, as well as the number of erroneous forecasts.
For forecasting in specific time intervals, the researchers studied 100 epochs (cycles) using hourly weather data samples for 1979-2021. Each of the submodels required 16 days of training on 192 V100 graphics cards. Currently, the Pangu Weather model can provide 24/7 worldwide weather forecasting in just 1.4 seconds on a V100 graphics card, which is 10,000 times better than traditional numerical methods.
Explaining why Huawei Cloud’s AI team decided to focus on weather forecasting, Dr. Tian Qi, Huawei Cloud’s Chief AI Scientist, IEEE Fellow, and lecturer at the International Eurasian Academy of Sciences, said:
“WEATHER FORECASTING IS ONE OF THE MOST IMPORTANT SCENARIOS IN THE FIELD OF SCIENTIFIC COMPUTING, AS METEOROLOGICAL FORECASTS ARE A HIGHLY COMPLEX SYSTEM. HOWEVER, IT IS DIFFICULT TO COVER ALL ASPECTS OF MATHEMATICAL AND PHYSICAL KNOWLEDGE.”
According to him, AI-based systems can analyze statistical patterns of atmospheric evolution based on big data. Pangu Weather is currently finalizing its forecasting system and is mainly focused on predicting the evolution of atmospheric phenomena.
“OUR ULTIMATE GOAL IS TO BUILD A NEXT-GENERATION WEATHER FORECASTING FRAMEWORK BY APPLYING ARTIFICIAL INTELLIGENCE-BASED TECHNOLOGIES TO IMPROVE EXISTING FORECASTING SYSTEMS,” THE SCIENTIST ADDED.
Fast operation
Commenting on the importance and quality of the Huawei Cloud team’s research, the scientific reviewers of the journal Nature explained that Pangu Weather is not only very easy to download and start using, it also runs very quickly from a desktop computer.
This means that every member of the meteorological community can use it if they wish. This opens up significant opportunities for research on how well the system predicts specific phenomena, and will also contribute to the development of the industry. In addition, the reviewers added that such results are certainly an improvement over previous performance. This article will allow us to imagine what forecasting models will look like in the future.
In May 2023, Typhoon Mawar attracted worldwide attention as the most powerful tropical cyclone of the year. According to the China Meteorological Administration, Pangu Weather accurately predicted Mawar’s trajectory five days before it changed course in Taiwan’s eastern waters.
Stable cloud environments, working sets, and appropriate operation and maintenance are essential for the continuous improvement of leading AI-based weather forecasting models.
