Tsinghua University and China Meteorological Administration large model appears in Nature: Solving world-class problems, 'ghost weather' forecast timeliness reaches 3 hours for the first time

Really "prepare for a rainy day", Tsinghua University's "ghost weather" forecast model is here!

It is the kind that can solve the world's unsolved problems -

It can predict extreme precipitation from 0 to 3 hours at the kilometer scale.

Extreme precipitation weather, including short-term heavy precipitation, storms, blizzards, hail, etc., can provide early warning.

Picture

Completing this research was not easy.

The School of Software of Tsinghua University cooperated with the National Meteorological Center and the National Meteorological Information Center. It took three years of joint research to propose this large extreme precipitation nowcasting model called NowcastNet, and it took nearly six years of radar observation data to complete it. Model training.

In the process test of 62 weather forecast experts across the country, this method is significantly ahead of similar methods in the world, and the research results have now been published in Nature.

Picture

Currently, NowcastNet has been deployed and launched on the National Meteorological Center’s short-term forecast business platform (SWAN 3.0), which will provide information for the national extreme precipitation weather short-term Provide support for forecasting operations.

So why is nowcasting of extreme precipitation so difficult? How did the Tsinghua team solve this problem?

Why is it listed as a scientific problem?

In recent years, due to the influence of global climate change, extreme precipitation weather has occurred frequently. Precipitation nowcasting that is more accurate, more precise and has longer warning lead time has become a focus of people's attention.

Since most extreme precipitation weather processes only last for tens of minutes and have a spatial scale of several kilometers, they are severely affected by complex processes such as convection, cyclones, topography, and chaotic effects of the atmospheric system.

However, numerical prediction technology based on physical equation simulation is difficult to effectively predict kilometer-scale extreme precipitation.

Therefore, at the World Meteorological Organization Summit on May 27 this year, precipitation nowcasting within three hours was listed as one of the important unsolved scientific problems.

△Precipitation nowcasting based on radar observations is one of the world's difficult problems

There have been methods for predicting extreme precipitation weather before.

Numerical calculation and deep learning are the two mainstream methods of precipitation nowcasting, but both have obvious flaws:

Numerical calculation methods are difficult to effectively model the spatiotemporal multi-scale characteristics of the precipitation process. Restricted by cumulative forecast errors, the forecast timeliness is often within one hour.

Although deep learning methods are good at modeling nonlinear systems, statistical models have inherent small-sample over-smoothing problems. The prediction solution process lacks physical conservation law constraints, and the generated numerical fields have severe blur and distortion, making it difficult to provide services. Value of extreme precipitation forecasts.

Nowcasting large model NowcastNet

In response to the above challenges, since 2017, the team of Professor Wang Jianmin and Associate Professor Long Mingsheng of the School of Software of Tsinghua University has established a research team with the National Meteorological Center and the National Meteorological Information Center. Cooperate on the application of artificial intelligence technology in meteorological big data.

After three years of joint research, the nowcasting large model NowcastNet was proposed and trained on radar observation data from the United States and China in the past six years.

The core of this model is the neural evolution operator for end-to-end modeling of the physical process of precipitation, which achieves the seamless integration of deep learning and physical laws.

△NowcastNet, a large nowcasting model that integrates physical modeling and deep learning

Specifically, the research team first designed a mesoscale evolution network, using In order to model mesoscale precipitation processes with more significant physical properties such as advection motion, and based on the material continuity equation (i.e., the law of conservation of mass), a neuroevolution operator was designed to simulate the ten-kilometer scale motion in the precipitation process end-to-end, and through Backpropagation minimizes the cumulative forecast error.

Secondly, the research team proposed a convective scale generation network. Based on the prediction results of the mesoscale evolution network, the probabilistic generation model is used to further capture the kilometer-scale precipitation process in which chaotic effects such as convection generation and dissipation are more significant.

Thanks to the above-mentioned fusion design, this model combines the advantages of deep learning and physical modeling, and is the first time in the world to extend the timeliness of precipitation nowcasting to 3 hours (mentioned above, the previous numerical calculation method Usually within 1 hour) and make up for the shortcomings of extreme precipitation forecasting.

In order to fully test the operational guidance value of the nowcasting large model NowcastNet for typical weather processes, the National Meteorological Center invited 62 frontline forecasting experts from 23 provincial and municipal meteorological stations to conduct a survey on 2,400 extreme precipitation processes in China and the United States. A posteriori and a priori tests are performed and compared with methods currently used in business.

The forecast system currently widely used by meteorological centers around the world includes the advection-based pySTEPS method. PredRNN is a data-driven neural network that has been deployed at the China Meteorological Administration. The DGMR model was proposed by Google DeepMind in cooperation with the British Met Office.

All models are trained and tested on a large radar dataset of precipitation events in the United States and China.

Picture

△Meteorological expert inspection results and numerical index evaluation results, CSI is used to measure the position accuracy of the forecast; PSD is used to measure the spectrum of the forecast Comparison between characteristics and radar observed precipitation variability.

As shown in the figure above, NowcastNet comprehensively surpasses existing technologies in numerical indicators such as Critical Success Index (CSI) and Power Spectral Density (PSD), and is considered to have the highest forecast in 71% of weather processes. value.

In the process of extreme precipitation, NowcastNet is the only nowcasting technology that shows strong business value.

Take the typical extreme weather processes in China and the United States as an example:

At 23:40 on May 14, 2021, a heavy precipitation process occurred in the Jianghuai region of China, Hubei, Anhui and other places The region has issued a red warning for heavy rainfall, and NowcastNet can accurately predict the changing process of three heavy precipitation supercells.

Picture

△a. Predicted geographical information, b. Prediction results of different models at T 1 hour, T 2 hours and T 3 hours , c. CSI is an indicator used to evaluate the accuracy of predictions

At 9:30 on December 11, 2021, a tornado disaster occurred in the central United States, killing 89 people and injuring 676 people. NowcastNet It can provide clearer and more accurate forecast results for the intensity, landing area and movement pattern of heavy precipitation.

Picture

The inspection shows that NowcastNet has good guiding significance for the precise prevention and control of extreme disaster weather.

Currently, the research results are published in Nature under the title "Skilful Nowcasting of Extreme Precipitation with NowcastNet", and are also published by Nature News and Viewpoint" reported on "The Outlook for AI Weather Prediction".

The researchers believe:

This research explores a new paradigm of data-driven and physics-driven "scientific learning" and proposes a method for modeling and predicting the space-time material field under the constraints of physical conservation. The general method also has application prospects for other problems with multi-scale physical properties.

They also said:

In the future, the application of this solution will be further promoted in scenarios such as physical problem solving, atmosphere and ocean simulation, and industrial design simulation.

Team Information

Professor Wang Jianmin and Associate Professor Long Mingsheng of the School of Software, Tsinghua University, as well as machine learning expert Michael I. Jordan, professor at the University of California, Berkeley, and honorary professor at Tsinghua University, are the authors of the paper. Corresponding Author.

Ph.D. students Zhang Yuchen and Associate Professor Long Mingsheng from the School of Software, Tsinghua University are the first authors of the paper. Master students Chen Kaiyuan and Xing Lanxiang participated in the research work.

Researcher Jin Ronghua of the National Meteorological Center provided meteorological knowledge and data support and presided over the inspection work of meteorological experts nationwide. Experts such as Luo Bing, Zhang Xiaoling, Xue Feng, Sheng Jie, Han Feng, and Zhang Xiaowen provided support for the research work. Guidance, advice and assistance.

This research was supported by the National Natural Science Foundation of China’s Innovative Research Group Project, the Outstanding Youth Science Fund Project, and the National Engineering Research Center for Big Data System Software.

Paper link: https://www.nature.com/articles/s41586-023-06184-4

The above is the detailed content of Tsinghua University and China Meteorological Administration large model appears in Nature: Solving world-class problems, 'ghost weather' forecast timeliness reaches 3 hours for the first time. For more information, please follow other related articles on the PHP Chinese website!