Weather forecasting is an important scientific tool to study the evolution of the atmospheric system, and it is also essential for social development and national strategies. The economic and social value it brings is huge. Worldwide, weather forecasting contributes $162 billion in economic benefits, or at least 0.185% of global GDP. In China, about 40% of GDP is affected by weather and climate, especially in agricultural production, where weather forecasting plays a more prominent role. In recent years, global warming has intensified and extreme meteorological disasters have occurred frequently, bringing great challenges to human production and life. Universal early warning mechanisms could not only prevent up to $13 billion in property damage each year, but also save countless lives.
History, principles and bottlenecks of traditional weather forecasting
The theory of numerical weather prediction originated at the beginning of the last century. In 1904, the Norwegian scientist Bjerknes proposed to predict the weather through mathematical and physical equations, but it was difficult to practice due to insufficient computing power. In 1922, British meteorologist Richardson mobilized a large number of manpower and completed the first 6-hour "forecast" with manual pen and paper after six weeks. In 1950, Charney, an American meteorologist, used an electronic computer for the first time to complete a 24-hour forecast in "only" 24 hours, causing a huge sensation in the meteorological community. With the improvement of computer performance, numerical weather prediction has gradually matured, the forecast duration has been extended to 5-7 days, and the resolution has been refined from hundreds of kilometers to several kilometers.
The basic principle of numerical weather prediction is to first form a grid of meteorological variables such as temperature, air pressure, humidity, wind speed and other meteorological variables from radar and satellite observation data, and to simulate their future evolution by solving atmospheric dynamics equations. Over the decades, scientists have refined complex systems of partial differential equations to more accurately describe the dynamics of atmospheric changes.
In recent years, numerical prediction methods have encountered bottlenecks in accuracy and speed. Accuracy has improved slowly, with timeliness improving by an average of only one day per decade. According to data from the European Centre for Medium-Range Weather Forecasts (ECMWF), between 2012 and 2022, the 3-7 day forecast error for many meteorological elements was reduced by less than 5%, due to the accumulation of errors in partial differential equations and the incomplete and inaccurate observation data. At the same time, weather forecasting consumes huge computing resources, and requires supercomputers to continuously calculate to meet actual needs, making it difficult for many underdeveloped countries to establish their own numerical prediction systems.
Large artificial intelligence models make weather forecasts more and more accurate
The rapid development of artificial intelligence technology has profoundly changed people's production and life. When traditional numerical forecasting is challenged, people wonder if artificial intelligence can be used for weather forecasting and surpass traditional methods. The Pangu Meteorological Model project carried out by Tian Qi's team at Huawei Cloud Computing Technology Co., Ltd. uses a 3D neural network adapted to the earth's coordinates and a hierarchical time-domain aggregation strategy to achieve accurate global medium-term weather forecasting. After being trained on global weather reanalysis data, the model can accurately forecast multi-layered meteorological elements within 7 days, which is about 0.6 days longer longer and reduces the forecast error of tropical cyclone track by 25% compared with the world's leading ECMWF system. The model can complete the global 7-day forecast in only 10 seconds, and the calculation speed is more than 10,000 times faster. This achievement was selected as one of the "Top Ten Advances in Chinese Science" in 2023. In a paper published in the journal Nature, they proposed the "Pangea Meteorological Model" that can be used for global medium-term weather forecasting, which not only surpasses the accuracy of traditional numerical prediction for the first time, but also performs well in typhoon track prediction. The Pangu meteorological model is based on the theory of deep learning, which is different from traditional numerical prediction: it does not rely on atmospheric dynamics equations, but builds a deep neural network, and uses historical meteorological data as training data to optimize the neural network. Tian Qi's team was the first to realize the great potential of this kind of method, and built a large 3D neural network model suitable for global high-resolution medium and long-term weather forecasting. The model applies the three-dimensional Transformer network architecture, adopts the classical encoder-decoder design pattern, and introduces the earth position prior and hierarchical time-domain aggregation strategies, which effectively improves the training efficiency and reduces the inference power consumption.
The team trained the Pangu meteorological model with global reanalysis data from 1979 to 2017 and conducted detailed tests on the 2018 data. The entire training process took about two months on 192 GPUs and more than 60 terabytes of training data. Experimental results show that Pangu has demonstrated amazing performance in the field of numerical weather prediction, and when using reanalysis data as input, its accuracy surpasses even the most accurate traditional numerical prediction model, the European Meteorological Centre's Integrated Forecasting System (IFS). It can complete the world's 7-day high-resolution numerical weather forecast in just 10 seconds, increasing the inference speed by more than 10,000 times, and reducing the computing power consumption by more than 100,000 times. The advantages of the Pangu meteorological model are mainly reflected in these aspects.
One is in terms of deterministic weather forecasting. In the 2018 ERA5 data test, the Pangu meteorological model surpassed IFS in forecast accuracy in meteorological variables such as temperature, air pressure, humidity, and wind speed, increasing the forecast time by 0.6 days, becoming the first artificial intelligence model to surpass IFS.
The second is in terms of extreme weather forecasting. Based on its predicted mean sea surface pressure (MSLP) variables, the Pangu meteorological model can accurately predict the position of the typhoon eye every 6 hours in the future through an iterative algorithm, and then calculate the typhoon path. In 2018, the test results of 88 named typhoons around the world showed that the absolute position error of the 3-day and 5-day predictions for the eye of the typhoon was more than 25% lower than that of the high-resolution system of the European Meteorological Center.
The third is in terms of integrated weather forecasting. The inference speed of the Pangu meteorological model is extremely fast, which can greatly reduce the computational overhead of integrated weather forecasting. In the ensemble forecast with 100 member variables, the accuracy of long-term forecasting has been significantly improved, and the uncertainty of the forecast results can be quantitatively analyzed.
Practical application of artificial intelligence large model in weather forecasting scenarios
Based on the above research results, the Pangu Meteorological Model team actively collaborated with partners such as the China Meteorological Administration, the Hong Kong Observatory, the European Meteorological Centre and the World Meteorological Organization to translate scientific research results into practical applications. So far, these efforts have yielded substantial results.
During the 2023 flood season, Pangu Meteorological Model cooperated with the China Meteorological Administration (CMA) to successfully track multiple typhoons and incorporate the relevant results into the regular consultation mechanism. Taking Typhoon Mawa No. 2302 and Typhoon Doksuri No. 2305 as examples, the Pangu meteorological model accurately predicts their paths and provides key support for mainland typhoon defense. The Hong Kong Observatory has also verified that the Pangu Meteorological Model outperforms traditional methods including IFS in the forecast of Typhoon Sula No. 2309, demonstrating excellent performance.
In addition, the European Meteorological Centre (EUMC) has also conducted practical tests on the Pangu Meteorological Model, covering a number of extreme weather events such as winter storm Otto and tropical cyclone Freddy in the Southern Hemisphere in February 2023. In July of the same year, the European Meteorological Center officially incorporated the forecast results of the Pangu meteorological model into its official website, providing real-time and accurate reference data for users around the world.
After listening to the technical report of the team, the World Meteorological Organization spoke highly of the Pangu meteorological model, believing that its low inference and computational overhead makes it have broad application prospects in developing countries. The Pangu team will work closely with the World Meteorological Organization (WMO) to introduce the Pangu Meteorological Model to 30 least developed countries in the world to provide early warning capabilities for these regions.
Possible risks and challenges to the development of numerical meteorological prediction on the mainland
Although the artificial intelligence methods represented by Pangu have shown great potential, the mainland still faces many risks and challenges in the field of numerical weather prediction, which need to be solved urgently.
First, there is an external dependency on data. At present, the mainland mainly relies on external data sources, such as internationally shared meteorological data, in numerical meteorological forecasting, which has potential risks. Once there is a problem with an external data source or access is restricted, the accuracy and timeliness of the continental numerical weather forecast will be seriously affected. Therefore, it is an urgent need to build a comprehensive and independent weather data system. This requires the mainland to speed up the independent research and development and deployment of meteorological observation equipment, integrate multi-source data such as satellites and ground observations, and form a complete and reliable data chain to provide solid data support for numerical meteorological forecasting.
Second, the bottleneck of computing power is a challenge. Weather forecasting involves a lot of computational and analytical work, and the demand for computing resources is extremely high. However, at present, there are still shortcomings in the field of high-performance computing, and it is difficult to meet the demand for computing power. This may lead to limited training and optimization of forecasting models, affecting the accuracy and efficiency of forecasting. Therefore, the mainland should increase investment, research and development of independent and controllable high-performance computing hardware and software, strengthen the construction of artificial intelligence computing power clusters, and provide a strong underlying computing power guarantee for numerical weather forecasting.
Finally, interdisciplinary research is insufficient. Meteorological forecasting involves many disciplines such as mathematics, physics, and computer science, but there are still shortcomings in the mainland such as the lack of interdisciplinary research teams and cooperation mechanisms. Therefore, the mainland should strengthen its support for innovative theoretical research and practical exploration, promote the cross-integration of different disciplines, cultivate and attract top talents, and reserve strong intellectual resources for future technological breakthroughs.