Google’s DeepMind applies machine learning to weather forecasts, defeating traditional forecasting methods

Since observing the stars, humans have been trying to predict the weather. The English classes in primary and secondary schools also let us know that most of the British people’s greetings start with the weather. Should I bring an umbrella? How to arrange the route of vehicles in heavy rain? What safety measures need to be taken during outdoor activities? The forecast of the weather is of great significance to daily life.

"Short-term weather forecast" is a forecast for the weather trend in the next 0-12 hours, and "nowcasting" is a type of short-term weather forecast, which specifically refers to the weather forecast for the next 0-2 hours, which is an energy source. Management, maritime services, flood warning systems, air traffic control, etc. provide key decision-making information.

▲ Picture from: Getty Image

Recently, DeepMind, a Google-owned AI laboratory, published a paper in Nature. The research content is to apply machine learning to rainfall nowcasting, and create a deep generative model (Deep Generative Model, hereinafter referred to as DGM) .

The alliance of environmental science and artificial intelligence has opened up a new path for nowcasting. DeepMind believes that the current impending forecast has two problems.

▲ Picture from: Getty Image

On the one hand, today’s weather forecast is mainly driven by the Numerical Weather Prediction System (NWP), but it is difficult for NWP to generate high-resolution forecasts for the near time within 2 hours. Nowcasting fills this critical interval. However, mainstream nowcasting methods also have shortcomings-it is not easy to capture important non-linear events .

On the other hand, several climate prediction methods based on machine learning have been developed in recent years. Although these methods can accurately predict low-intensity rainfall, they do not perform well in rare moderate to heavy rain events.

▲ Observation radars in the past 20 minutes provide probabilistic predictions for the next 90 minutes. Picture from: DeepMind

In short, DeepMind believes that in order to make nowcasting more valuable, it is necessary to provide accurate forecasts, fully consider uncertainty, and make statistically significant improvements in heavy rain forecasts.

At the same time, advances in weather sensing have allowed high-resolution radars to be used at high frequencies, often every 5 minutes with a resolution of 1 km. These high-quality data provide opportunities for the intervention of machine learning technology.

DeepMind's DGM learned the probability distribution of the data and was trained based on a large number of precipitation event data sets recorded by the British radar from 2016 to 2018. After training, it can provide nowcasts after running on a single NVIDIA V100 GPU for just over a second. DeepMind asserts that DGM can predict weather events that are difficult to track under potential randomness and accurately predict the location of precipitation.

▲ Compared with the other two methods, DeepMind's prediction (upper right) is more accurate and clear. Picture from: DeepMind

Judged by 56 meteorologists, compared with mainstream nowcasting and other machine learning models, DGM has a more realistic and consistent nowcasting in an area of ​​1536 km×1280 km, compared to the other two in 89% of the cases The method is more accurate and practical, and the approach time is 5 to 90 minutes.

Artificial intelligence has more uses in the field of climate change. In October 2019, researchers used artificial intelligence to generate extreme weather images to visualize climate change. It is difficult for climate issues to evoke collective mobilization. One reason is that people believe that these changes usually occur in distant time and space. Therefore, only personally relevant and even emotional information can produce truly effective communication.

▲ The generated image on the right. Picture from: venturebeat

Researchers input images of different locations and building types (such as houses, farms, streets, cities) to form more than a dozen artificial intelligence synthesis patterns, and then ask evaluators to choose between real images and semi-generated images to calculate the average error rate . The ultimate vision of this work is to create a machine learning architecture to generate the most realistic images under extreme weather, including floods, wildfires, tropical cyclones, and even more catastrophic events based on the location selected by the user.

▲ Picture from: Getty Image

"Climate change" is the key word of this year. In 2021, the Nobel Prize in Physics was awarded to three scientists, two of whom were awarded for research on "building a physical model of the earth's climate, quantifying its variability, and reliably predicting global warming" . According to the international non-profit organization CDP, the world's 500 largest companies will need to pay about $1 trillion in the next few decades to cover the costs associated with climate change, unless they take positive measures in advance.

DeepMind senior researcher Shakir Mohamed believes:

The ability to model complex phenomena, make rapid predictions, and express uncertainties makes artificial intelligence a powerful tool for environmental scientists.

In line with this situation, DeepMind's model and other similar models may have a wide range of applications, helping forecasters spend less time browsing the ever-growing pile of forecasting data, so as to focus on the meaning behind the forecast.

Grapes are not the only fruit.

#Welcome to follow Aifaner's official WeChat account: Aifaner (WeChat ID: ifanr), more exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo