As climate change leads to extreme weather and wildfires, it has never been more important to forecast what lies ahead so that at-risk areas can be evacuated and the emergency services properly deployed.
In August, parts of the Middle East suffered scorching temperatures because of a heat dome, while this year the EU had its worst wildfire season.
But are the climatic factors that are causing these more severe conditions also making it harder to predict what will happen?
Forecasting in a warmer world
“The atmospheric state may become less predictable in hotter worlds than in colder worlds,” said Dr Simon Driscoll, a senior research associate in the Department of Applied Mathematics and Theoretical Physics at the University of Cambridge.
Forecasting models based on past data may be more likely to miss, for example, heatwaves, although balanced against this is the reality that improvements in computational power and satellites have, overall, resulted in “slow progress” in weather forecasting.
When it comes to wildfires, challenges are also emerging. Stefan Doerr, professor of wild land fire science at Swansea University in the UK, said that climate change had altered the relationship between seasons and the fire risk that they were associated with, making seasonal wildfire forecasting harder. The use of past data can complicate wildfire forecasting just as it can weather forecasting.
“There’s another problem: once fires become large, they generate their own extreme weather and even produce tornadoes,” Prof Doerr said.
“That’s still rare, but we’re seeing it [more] round the world over the last few years, making fires even more dangerous and unpredictable.”
Wildfires: New seasons, new risks
While climate change may complicate the forecasting of wildfires and the weather, improvements in computational power and satellite data are leading to progress in forecasts.
And a major development in recent years has been the use in forecasting of machine learning, a type of artificial intelligence (AI) in which machines learn from experience.
It is being used to predict the weather, wildfires, tsunamis and even earthquakes, and in some instances these forecasts can be faster and cheaper to generate, much less demanding in energy terms and potentially more accurate.
The rapidly growing field of machine learning-based forecasting involves multiple institutions, including the European Centre for Medium-Range Weather Forecasts (ECMWF), a pan-European organisation supported by 35 nations.
Their machine learning approach to wildfire forecasting, Probability of Fire, gives the likelihood of a fire developing at any given location up to 10 days ahead.
AI evolution and weather forecasting
Dr Joe McNorton, an ECMWF scientist, said that until AI’s influence, fire forecasting involved four key weather elements – temperature, humidity, wind speed and precipitation – and had evolved little in the past half century.
“The real contribution of AI and machine learning in the field has been the ability to just throw lots of odd data at it,” he said.
“Odd in the sense that in the modelling of weather or fire, it’s data that we wouldn’t have considered before, like population density or urban information – things that don’t offer a simple physical connection to fire, but somehow this black box of machine learning can interpretand produce something more meaningful.”
Given that past climatic or wildfire activity may be less relevant to the present day, an advantage of AI wildfire forecasting is that it does not necessarily depend on data going back a long way. AI wildfire models like those developed at the ECMWF are typically trained on information from just the past five or 10 years, which remains a reliable guide to present-day behaviour.
Dr McNorton said that the AI models were able to predict “quite well” the extreme wildfires that Canada has experienced since 2023, even though the data on which they were trained did not include such out-of-the-ordinary scenarios.
“So it seems relatively robust to record-breaking heatwaves and multiyear droughts and things like that, but we can constantly retrain it,” he said.
The approach uses “decision trees”, a series of yes and no questions, such as whether the temperature is above a certain level or whether the precipitation is below a particular threshold. It works through this chain of decision trees until it reaches a final probability.
Balancing energy use and AI
While there are multiple ways in which AI forecasting models can work, typically they require much less energy than conventional approaches.
The ECMWF has an AI-based weather forecasting model that, since early this year, has run alongside a conventional physics-based forecasting model. The AI model uses just a thousandth as much energy.
Energy use is so much less because when AI generates weather forecasts it tends not to need to model the myriad complex physical processes, such as cloud formation, that determine what the weather is going to be like.
According to Hannah Cloke, an ECMWF research fellow and professor of hydrology at the University of Reading in the UK, the organisation’s AI weather forecasting model has shown itself to be comparable to or, in some respects, better than traditional forecasting.
