Computer scientists from DeepMind and the University of Exeter in England have teamed up with meteorologists at the Met Office to create an AI model that can predict if it will rain up to 90 minutes in advance.
Traditional forecasting methods rely on solving complex equations that take into account various weather conditions, such as atmospheric pressure, humidity and temperature of the Earth’s atmosphere. The problem is, at least in Blighty, these systems tend to to predict what awaits us whole days or weeks to come.
Deep learning models are better suited for making shorter-term predictions – such as in the next few hours – according to an article published by the aforementioned boffins in Nature Wednesday. There are advantages to using AI algorithms; they do not have to solve thermodynamic equations and require fewer calculations than other predictive techniques.
The team led by DeepMind has formed a generative adversarial network (GAN) to produce a sequence of maps indicating where it will rain. Each of these precipitation maps shows where moisture accumulates and moves through the atmosphere, each covering an area measuring 1,536 × 1,280 km. Millions of examples of these maps were collected from radar observations from 2016 to 2018. Data collected in 2019 was for testing only.
The model was fed by a sequence of sample maps, each capturing weather data at five-minute intervals during the training phase. He learned to pick up common patterns depicting how clouds spread across the sky and whether or not they produced rain.
During the testing phase, the system was asked to generate the next set of maps to predict precipitation at intervals of five minutes up to 90 minutes from four previous examples. Essentially, it’s a bit like feeding a system a short video clip and training it to predict the next frames.
Its performance is evaluated using a number of factors, including the smoothness and gradation of changes between images or maps. Instead of calculating a simple method to determine the accuracy of the model, the team asked a panel of 50 expert meteorologists to rank the predictive maps produced by the GAN and compare them with maps produced from other types of weather. more traditional numerical prediction systems.
“With the help of a systematic evaluation carried out by more than 50 expert meteorologists, we show that our generative model ranked first for its accuracy and usefulness in 89% of cases compared to two competing methods”, indicates the document.
But how accurate is the model really? It’s hard to say, said Niall Robinson, co-author of the study and head of partnerships and product innovation at the Met Office. The register.
We decided that in some ways the purest way to see if our approach was worthwhile was to simply ask end users – our own meteorologists. In a blind study, they overwhelmingly preferred our new approach to other algorithms
“Machine learning algorithms generally try to optimize for a simple measure of the ‘good’ of the prediction. However, the weather forecast can be good or bad in many different ways; maybe one forecast gets precipitation in the right place but the wrong intensity, or another gets the right combination of intensities but in the wrong places, and so on. Ultimately, “good” depends on the quality that is useful for the problem encountered.
“We put a lot of effort into this research to evaluate our algorithm against a wide range of metrics, so we showed it to be ‘good’ in a number of different ways. Additionally, we decided that in a In some sense, the purest way to see if our approach was valuable was to simply ask end users – our own meteorologists. In a blind study, they overwhelmingly preferred our new approach to other algorithms. To our knowledge, this kind of really comprehensive assessment of several different types of ‘coupons’ It had never been done before to assess the use of AI in meteorology. I think this is quite unusual in the field. machine learning research. “
The research project is more of a proof-of-concept effort at the moment, and the Met Office won’t be using AI algorithms in real-world predictions anytime soon. The weather agency needs to create all kinds of other tools before something like a GAN can be used, Robinson said:
“Once useful new research is published, there is still a lot of work to be done to create a functioning department. For example, we need to take a close look at how new tools are deployed and maintained, the best user interfaces for our meteorologists, and how that fits into all the other forecasts we provide. This way, we can ensure we deliver the reliability that is at the heart of the Met Office goal. “
The Met Office said it is exploring solving other AI research problems with help from industry and academia. ®