Extreme weather events like heatwaves and droughts are becoming more common as the earth heats up. We need accurate predictions of these events to know how to allocate disaster resources and prepare for economic trouble.
This is difficult.
Scientists are quite good at calculating how a particular storm will develop and where it will go once it has formed, but knowing which regions of a country will experience extreme heat in the next decade? That’s much harder.
Professor Pedro Lind at OsloMet’s NordSTAR Centre and the OsloMet AI Lab has partnered with researchers at the University of Lisbon in Portugal to develop new artificial intelligence (AI) methods for predicting these events.
It’s a timely project and one that Lind says is very challenging from an AI and mathematical perspective. As part of the research group DHEFEUS, Lind and the other researchers are using their computer science and statistical analysis backgrounds to extract patterns from historical weather data and tease out the rules behind these events.
Before they can start making predictions though, they need to answer a fundamental question: What exactly counts as an ‘extreme weather event’?
Climate versus weather
“It’s difficult because climate is not defined locally or instantaneously; that’s the big difference between meteorology and climate”, says Lind.
Meteorology is what your weather app tells you. Depending on where you live, this can accurately predict what the weather will be like in your area over the next few days. In contrast, climate is the greater trend over a large area. This is much harder to know.
According to Lind, the challenge comes from the interplay between having a small amount of very accurate data for a limited location – as is the case with weather – and large amounts of data that are averaged over whole regions – what we know about climate.
He equates it to managing the COVID-19 pandemic. If you just looked at Norway’s average infection numbers during the height of the pandemic, you might get an impression that the country was doing fine and that we don’t need to take action.
However, at a finer level, we see that there were few infections in the countryside but a lot in the cities. We need to allocate resources and policy accordingly. Just looking at the big picture – the average – will not give us all the information we need.
Understanding trends over time and distance (like climate or infection rates) requires averaging, but we lose some fine detail when we do that. Specifically, we lose the outliers: the extreme events that Lind and his team are looking for.
Using AI
Fortunately, there are ways to cope with these modelling issues. Lind is using a method called ‘Isolated Forests’, a form of AI that can find outliers in the data. This lets him search through the data records and get an idea of what extreme events are happening now and how they have changed over the past decade.
"These are methods that use different sources of data to find patterns, what happened 10 years ago and what is happening now, so we can extrapolate what is going to happen in the future."
The goal is to find the risk factors for these extreme events.
Lind, along with his colleagues and two master’s students at OsloMet, is starting by combing through precipitation and temperature data from the Iberian Peninsula. Going through all that data trying to find events that only started to become common in the last decade by hand would be an impossible task.
But their AI can do it.
Within the ACIT lab, master’s student Marte Brekke is creating a kind of AI known as a ‘neural network’, so-called because it mimics how the human brain processes information. This type of AI is common in many applications today and has become possible in the past few years because of the immense computing power now available.
Lind begins by showing the AI a whole lot of data about weather in Portugal. This is called the ‘training data’. He is careful to ensure that the training data accurately represents the world and doesn’t contain any biases.
Once the AI has learned what causes extreme weather, Lind can test its prediction ability against real data.
The learning process develops a set of mathematical functions based on several combination of the input variables (precipitation and temperature in this case). “By repeating many functions and combinations, we get what is called a ‘deep’ neural network that should be able to reproduce any function in nature that maps a certain input to a certain output,” says Lind.
In other words, if they observe an extreme event in one place, the AI should be able to look at the date and predict where a similar event will happen in the future.
What is AI?
AI has been all over the news in the past couple of years. It seems like it’s just about everywhere. But what is it?
Remember in school when your teacher showed you a bunch of points on a graph and you had to come up with a line that fit the data best, and then you could use the equation of that line to predict where new points were going to be?
AI does that, just at a much bigger scale. Instead of an X-Y plot with a few points, AI algorithms can look at thousands of variables across terabytes of data. That’s why you could guess a point back in math class, but AI can predict what you are going to type next or even make new images in the style of old artists.
Predicting and planning
The next step for Lind and his team is to find the limits of their AI’s prediction ability and expand the data to include the rest of Europe, including Norway.
He cautions that making these predictions is more like trying to predict an earthquake than whether it will rain tomorrow.
The question is how well can we predict something like the risk that a particular town in Norway will experience another storm like extreme weather event Hans, a heatwave, or a drought.– Pedro Lind
They can’t predict the next storm yet, but by the end of the project they should be able to determine if an area will be in particular danger of extreme weather over the next decade.
As he and his team refine their AI model, Lind can draw on all the knowledge and experience of OsloMet’s AI lab. The methods that Lind uses to predict extreme weather overlap with their research in intelligent health, renewable energy, and even eye motion.
This will ultimately allow the researcher to classify different regions (Oslo, Bergen, etc.) for the likelihood of extreme events like heatwave or drought over the next several years and create a heatmap of the risk.
Lind hopes this information will give policy makers insight into extreme weather and give them time to effectively allocate resources.
