Machine Learning Improves Weather Forecasts Two to Six Weeks Ahead

Researchers have developed a machine-learning tool that improves weather forecasts two to six weeks ahead, a period when predictions typically become less reliable, according to MIT News and The Verge.

Both outlets reported that the approach offers a relatively low-cost way to improve forecasts and that the findings were published in Nature. Rather than replacing existing forecasting systems, the technique learns from their past errors and adjusts their predictions, MIT News explained.

MIT News reported that the correction method doubled forecasting accuracy when applied to a leading forecasting system. The winning entry in a European Center for Medium-Range Weather Forecasts competition combined the centre’s model with another model, then added the machine-learning correction.

According to MIT News, the resulting forecasts consistently outperformed long-term climate averages over the two-to-six-week period. The researchers also found a statistically significant improvement in predictions of extreme weather, including events such as flooding and cold spells.

The Verge highlighted the opportunity for farmers and emergency responders to plan further ahead. MIT News additionally described applications for energy operators trying to match supply with demand, alongside agricultural decisions about fertilising before rain and protecting crops from freezing conditions.

MIT Research Scientist Judah Cohen told MIT News that additional warning could give energy officials more time to prepare for an approaching ice storm. The outlet cited the February 2021 winter storms in Texas as an example of the consequences of inadequate preparation: disruptions caused shortages of food, water, heating and electricity, and at least 246 people died.

The forecasting method uses historical information to learn how an existing system makes mistakes. As further forecasts and observations become available, its corrections adapt, according to MIT News. The approach works with probabilistic forecasts, which describe the likelihood of different possible outcomes.

MIT News distinguished this technique from simpler forms of error correction. Cohen said conventional correction can pull predictions towards average conditions, making unusual events harder to anticipate. The new method improved predictions of extremes as well as average conditions.

The researchers demonstrated improvements in land forecasts across most regions worldwide, MIT News reported. They also tested the correction on other forecasting models, improving weaker systems to a level comparable with the European centre’s forecasts. The method improved models that already used another AI-based correction.

Both MIT News and The Verge described the software as open source. MIT News added that although the researchers used MIT’s Engaging supercomputing cluster to develop it, the finished tool does not require a supercomputer to operate.

The Verge focused on Microsoft researchers’ involvement and placed the work alongside AI weather-forecasting efforts by Google and Nvidia. MIT News identified collaborators from Microsoft, Harvard College, the University of Toronto, Brazil’s National Institute for Pure and Applied Mathematics, DTN and the European forecasting centre, alongside MIT.

MIT News also reported that the meteorological office for the Republic of Kenya had contacted the researchers about implementing the tool for forecasts of the annual monsoon season, with the aim of supporting agricultural yields.

The research received support in part from the U.S. National Science Foundation, the Carlos Chagas Filho Foundation for Research Support and the European Union, according to MIT News.

Sources

• MIT News: Researchers improve weather forecasts looking two to six weeks out
• The Verge: Microsoft has a new machine learning tool to improve weather forecasts