After playing go and StarCraft, deepmind AI learned to predict the weather


Tencent technology news on September 30, after playing go and StarCraft, deepmind has brought its artificial intelligence (AI) into another challenging field, that is, helping to predict the weather. In the past few years, the AI company of alphabet, the parent company of Google, has been quietly cooperating with the Met Office. They recently reported their research results in the journal Nature. In short, deepmind has designed a new machine learning model to predict whether it will rain in the next few hours.
This type of weather forecast is called “precipitation approaching forecast”, that is, the forecast is given on a very short time scale before rainfall, sometimes only two hours in advance. Today’s weather forecast can well predict the rainfall in the next six hours to a few weeks, but there will be a blind spot before that, which is where machine learning can help bridge the gap.
The researchers wrote in their paper: “Proximity forecasting, that is, high-resolution precipitation forecasting two hours in advance, supports the real social and economic needs of many industries that rely on weather for decision-making. For these industries, accurate proximity forecasting is an important challenge for a long time. We use the method of deep generation model to directly solve this important problem, improve the existing solutions, and provide a reference for the real world The world’s decision makers provide the required insight. ”
Proximity forecasting is the key to making weather related decisions because it provides information for emergency services, energy management, retail, flood early warning systems, air traffic control, marine services, etc. However, in order for proximity forecasting to be useful, forecasting must provide accurate forecasts and explain the uncertainties, including things that may greatly affect human life Pieces.
In addition, the increasing climate crisis also means that extreme weather events such as rainstorms or floods will only become more frequent. The ability to predict rainfall better and faster is very important for making rapid decisions in these situations, such as stopping trains or evacuating people.
The Met Office relies on radar images to predict the weather. The working principle of radar is to send a beam into the atmosphere and then time the reflection time, which will tell us how much water is in the atmosphere. The more water, the more rainfall. Then, these data are sent to the head office of the met office, where they are processed to obtain the airborne water cloud map of the UK. Deepmind’s model Based on the UK radar images from 2016 to 2018, they were trained to reliably predict what will happen in the next hour or two.
In recent years, several machine learning based methods have been developed. They are trained on large data sets observed by radar in order to better simulate heavy precipitation and other unpredictable precipitation phenomena. For example, Google and the National Oceanic and Atmospheric Administration (NOAA) Cooperate, research and develop machine learning systems that may be injected into NOAA business. Microsoft also funds to identify repeated weather and climate patterns from historical data to improve the sub seasonal prediction model.
However, deepmind points out that AI proximity forecasting models do not always include small-scale weather models or provide forecasts for the whole region. As an alternative, the company has created the so-called depth generation model (DGM) For prediction. Deepmind claims that DGM can predict weather events, which are difficult to track due to potential randomness. In addition, they can predict the location of precipitation as accurately as the system of tuning tasks, while retaining attributes useful for decision-making.
More than 50 meteorologists verified the model. In the research, meteorologists were asked to compare deepmind’s DGM method with another temporary prediction method called pysteps and a different depth learning method. In nearly 90% of cases, deepmind’s model performed best in accuracy and practicability.
However, other scientists do not approve of deepmind’s research. Peter Clark, a meteorologist at Reading University, said: “I don’t see any changes in forecasting here. It’s puzzling to rely on this indicator to prove whether their model is useful. I’m still surprised that they didn’t choose to use a more appropriate objective score. In addition, there is little detail on how these forecasts are carried out and even what is actually evaluated.”
Deepmind did not give specific figures to show how accurate its model is compared with other existing models. “We want to take a more cautious approach rather than report simple figures,” said Shakir Mohamed, a senior scientist at deepmind and author of the paper
Deepmind’s research may just provide a different method, rather than completely subvert the rainfall prediction technology we know. Rob Thompson, a meteorologist at Reading University, said: “it is similar to other cutting-edge models, but it is not far ahead, but it may be a little better.”
David Schultz, a meteorologist at the University of Manchester, agreed. “The proposed scheme seems to have defeated the existing one. I don’t know if it has changed the rules of the game. However, some of the existing methods have been greatly improved in their paper,” he said.
Mohammed said there are no plans to put the model into use, but the research team hopes to eventually use it to support future real-time weather forecasting. However, the researchers said that AI will not completely replace weather forecasters in the future. Suman Ravuri, a deepmind research scientist and author of the paper, said: “This will require the participation of experts and human beings to ensure that their understanding of the forecast is reasonable and need to communicate with the public.” (reviewed by Tencent technology / Jinlu)