UvA researchers can better predict extreme weather with AI model
Weather forecasts have improved significantly in recent years with the help of AI. UvA researchers have contributed to an AI model that can now predict extreme weather more accurately than traditional weather stations. “If you see extreme weather coming, you have more time to make evacuation plans.”
Long periods of drought, extreme rainfall and severe tropical storms: the likelihood of extreme weather is increasing due to climate change. These extreme weather conditions are more difficult to predict than “normal” weather conditions because they have occurred less frequently and therefore less data is available for weather models.
When extreme weather can be predicted, this is more likely in wealthy Western countries where most weather stations are located. In the global south, meteorological institutes are fewer and farther between, while the most serious consequences of extreme weather occur more frequently.
Weather forecast within minutes
UvA researchers, together with an international team of data scientists, meteorologists and climate experts, have come up with a solution. They developed within Microsoft Research, the research subsidiary of Microsoft, an AI model, called Aurora, that can predict extreme weather conditions with greater accuracy than most traditional weather models.
“The idea is that if you see extreme weather conditions coming earlier, you have more time for mitigation strategies or evacuation plans,” explains Ana Lucic, UvA researcher in artificial intelligence who collaborated on the research.
In addition, Aurora uses less computing power, which means that making a forecast is not dependent on the supercomputers of meteorological institutes. What’s more, Aurora works faster: unlike the traditional method, a ten-day weather forecast takes only one minute of computing time instead of hours. The researchers have published their results in the journal Nature, and the model’s code is available online for anyone who wants to use it.
How do you predict the weather?
Weather forecasts are currently still made by large meteorological institutes such as the KNMI in the Netherlands. They measure weather conditions via ground stations, weather balloons and satellites. This data is fed into a traditional weather model based on physical laws that describe how weather systems work based on air and ocean currents.
By solving mathematical equations, the Navier-Stokes equation to be precise, a forecast of the weather at a point in the future is produced. Solving this equation takes a lot of time and computing power and can therefore only be done by the supercomputers of the major weather stations. Moreover, for each new point in time, the calculation must be performed again based on the newly obtained data.
Artificial intelligence can speed up that process. Researchers trained a machine learning model, a form of AI, on millions of hours of weather data from the classic model. Lucic: “Training an AI model is time-consuming, but once it works, it is much faster and cheaper than traditional models.”
Artificial intelligence can speed up that process. Researchers trained a machine learning model, a form of AI, on millions of hours of weather data from the classic model. Lucic: “Training an AI model is time-consuming, but once it works, it is much faster and cheaper than traditional models.”
AI predicts weather better
The use of artificial intelligence for weather forecasting is not new. Scientists have previously demonstrated that when you train an AI model on millions of hours of historical weather data, the predictions are as good as those made by supercomputers based on the laws of physics.
New to Aurora is that the model can not only predict the weather but also perform other tasks. “Our model has been trained on different types of data,” explains Lucic. “A mix of analyses, weather forecasts and climate projections.” This enables the model to predict air pollution, ocean wave heights and tropical storms. Lucic: “Previously, it was not possible to predict all this in a single model.”
The researchers see Aurora as a basic model that can be used for many more tasks. “It’s a matter of adding more data to Aurora and further fine-tuning it,” says Lucic. In the future, the model could also be used to predict flood risks, forest fires, agricultural yields or the yield of solar and wind energy.
What’s more, in many cases, the model is also more reliable. For example, the air quality was more accurate than traditional models in 74 per cent of cases, for ocean wave heights this was 86 per cent, and Aurora predicted tropical storms often better than the seven international weather services. The researchers were surprised by this. Lucic: “We did not specifically train the AI model on tropical storms, but the machine learning algorithm learned this from other data.”
Chaotic system
However, there are also limits to AI’s weather predictions. Lucic: “The weather is and always will be a chaotic system. This means that it is inherently hard to predict because small changes can lead to enormous differences, also known as the butterfly effect. It is therefore very difficult to predict the weather with 100 percent reliability for the long term, we can only ensure greater accuracy.”
