The rainfall probability is predicted 90 minutes in advance with a GPU for one second, and the deepmind ml weather forecast is posted to nature

Heart of machine 2021-10-14 02:35:23
as everyone knows , English people like to talk about the weather very much . One important reason is that the weather in Britain is more changeable , You usually need an umbrella when you go out . There is also a lot of rain in China this summer , Rainfall has become an important factor affecting people's decision-making .

At the top of the UK AI Research Institute DeepMind Recently, it was written in cooperation with meteorological institutions and published in 《nature》 A paper on weather prediction was published on , It is pointed out that artificial intelligence is expected to help people cope with decision-making challenges in a changing environment .
Address of thesis :

Short term weather forecast

Throughout human social life , Weather forecasting has a long history . As early as the Eastern Han Dynasty , Zhang Heng invented the world's first anemometer —— Xiangfeng copper bird . Qin Jiushao, a mathematician in the Southern Song Dynasty, put forward the traditional rainfall test conversion formula . In the Yongle period of the Ming Dynasty , China's rain gauge has developed into a set of mature tools . In modern life, people usually use powerful numerical weather forecast (NWP) System to predict the weather .
Now? DeepMind The method of machine learning is introduced .

In the numerical weather forecast (numerical weather prediction, NWP) in , Basic planetary scale predictions can be provided several days in advance by solving physical equations , But it's hard to generate short-term ( For example, within two hours ) High resolution prediction . So we need real-time forecast (Nowcasting) To fill the performance gap .

Real time forecasting is important for water resources management 、 Agriculture 、 aviation 、 Areas such as emergency planning and outdoor activities are crucial . Advances in weather sensing have led to the use of high-resolution radar data at high frequencies ( Measure surface precipitation ) Become reality . However, the existing methods are difficult to make the best use of high-quality data , Therefore, this study proposes to use machine learning to improve real-time prediction .

Conditional generation model for real-time rainfall prediction DGMR

DeepMind This study focuses on real-time rainfall forecasting : Maximum advance 2 Hourly predicted rainfall 、 Time and place of rainfall . They use a method similar to GAN Depth generation model method DGMR, Based on the past radar data, the future weather information is predicted in detail and reasonably .

The model is trained on the corpus of large-scale rainfall events , And the study also uses an importance sampling scheme (importance-sampling scheme) To create a dataset more representative of heavy rainfall . In the whole process , All the models are in 2016 to 2018 Training on British radar observation data , And in 2019 It was evaluated on the test set in .

After training , The model can quickly generate full resolution real-time prediction : Just one NVIDIA V100 GPU You can generate a single prediction in about a second .

Based on the past 20 Data obtained by minute observation radar ,DGMR Be able to predict the future 90 The probability of rainfall in minutes .

In principle ,DGMR The real-time prediction algorithm is a conditional generation model , It is based on a given point in time T Use radar based ground rainfall estimates X_T, Based on the past M A radar field predicts the future radar field . The model contains potential random vectors Z And parameters θ, You can use the equation (1) Express :
as follows DGMR The schematic diagram of the model architecture shows the potential variables with space Z Radar generator (generator):
The figure below a Is the architecture diagram of radar generator ; chart b( Left ) From top to bottom is the time discriminator 、 Spatial discriminator and potential condition stack ,b( Right ) From top to bottom is G block 、D and 3D block 、L block .
Model to evaluate : The accuracy and practicability are the strongest

conceptually , This is a problem of generating radar images . at present , The research team used these methods , That is, large-scale rainfall events can be accurately captured , Many alternative rainfall scenarios can also be generated ( It is called ensemble forecasting ensemble prediction), To explore the uncertainty of rainfall .

The research team said , They are particularly interested in the ability of these models to predict rainfall events from moderate to heavy rain , These have a great impact on people's life and economy . therefore , They will DGMR And PySTEPS and UNet The two similar methods are compared . For the sake of fairness , The researchers hid the names of these models , And invited the National Weather Service 56 A meteorologist conducted a cognitive assessment .

Compared with the other two methods ,DGMR stay 1536×1280 A more realistic... Is achieved in an area of kilometers 、 More spatiotemporal consistent predictions , also DGMR Pay more attention to the future 5 To 90 Real time prediction in minutes .

The researcher explained in detail with two examples . First of all 2019 year 4 The extremely heavy rainfall in Britain in June ( Based on observation radar data ), This is shown in the following GIF ,DGMR Compared with the advection method PySTEPS It can better capture the circulation 、 Strength and structure , And it can more accurately predict the rainfall and movement in the Northeast . And deterministic deep learning methods UNet comparison ,DGMR The resulting forecast is clearer .
And then there was 2019 year 4 Heavy rainfall in the eastern United States in June ( Based on observation radar data ), This is shown in the following GIF , And PySTEPS comparison ,DGMR Can balance the intensity and range of rainfall , And not like UNet Then the prediction result is fuzzy .
Researchers say , Compared with other widely used real-time prediction methods ,56 A meteorologist is here 89% In our case DGMR As the preferred solution , It is proved that this method can provide strong support for real decision-making .

Continue to improve the accuracy of long-term prediction in the future

By using statistics 、 Economic and cognitive analysis ,DeepMind The researchers showed a new 、 Competitive radar rainfall real-time prediction method . Future researchers need to do more work to improve the accuracy of long-term prediction and prediction of rare and strong events .DeepMind Other methods for evaluating performance will also be developed , Further, these methods are specially used in specific practical applications .

Researchers think this is an exciting research field , They hope this paper can be used as the basis for their new work , Provide data and verification methods , Make it possible to provide competitive verification and operational utility , And promote greater integration of machine learning and environmental science , To better support decision-making to address climate change .

Link to the original text :
本文为[Heart of machine]所创,转载请带上原文链接,感谢