Researchers improve weather forecasts looking two to six weeks out.
A team including MIT researchers built a machine-learning correction layer that improved forecasts two to six weeks ahead, doubling the accuracy of a leading model and winning an international forecasting competition.
A research team that includes scientists at MIT has developed a machine-learning method that improves weather forecasts in the two-to-six-week range. According to MIT News, the model learns to correct errors in existing forecasting systems. Applied to a leading forecasting model, it doubled accuracy over that time frame and placed first in an international forecasting competition. The researchers describe the work in a paper in Nature.
Forecasts beyond two weeks have long been unreliable, which leaves little lead time for energy operators, farmers and emergency officials. The article points to the February 2021 winter storms in Texas, which led to the worst energy infrastructure failure in state history and killed at least 246 people. Earlier warning could have allowed steps such as bringing more natural gas trucks into affected areas. MIT research scientist Judah Cohen, a co-author, calls the two-to-six-week range the "predictability desert."
The team built what Cohen described as a wrapper, a postprocessing layer that sits on top of an existing forecasting system. It uses artificial intelligence to learn from that system's past mistakes in historical data and then shifts its forecasts. The researchers say the correction keeps improving as more forecasts and data arrive. They built the model on an MIT supercomputing cluster, but running it does not require one.
Last year the team submitted three models with the postprocessing layer to a competition run by the European Center for Medium-Range Weather Forecasts. The winning entry combined that center's own model with another and applied the correction. Cohen said it was the first time a model beat long-term climatological averages in this range. "We consistently beat climatology averages with our models," he said. The model also showed a statistically significant improvement in predicting extreme weather, while the center's own non-AI correction made extreme-weather predictions worse.
The paper shows the technique improved land forecasts in most regions of the world and also helped other models, bringing weaker ones closer to the center's level. Cohen said it can be used with any model and is open source and low cost. He reported that Kenya's meteorological office has asked about using it to forecast the annual monsoon season for agriculture this winter. The work was supported in part by the U.S. National Science Foundation, the Carlos Chagas Filho Foundation for Research Support and the European Union.
At a Glance
- The method doubled the predictive power of a leading forecasting model over the two-to-six-week range, according to the researchers.
- It placed first in a competition run by the European Center for Medium-Range Weather Forecasts.
- The results appear in the journal Nature.
- At least 246 people died in the 2021 Texas winter storm failures cited as an example of the need for longer forecasts.
- Kenya's meteorological office has contacted the team about using the model for monsoon season forecasts.
Sources
This brief was written by Brightside Report from the reporting linked above. Spot an error? Request a correction.