Weather forecasts that look several weeks ahead have long struggled to beat historical averages. A new machine-learning approach developed by a research team including MIT researchers aims to improve those predictions by learning how existing models tend to go wrong.
The technique focuses on the two-to-six-week window, a period when forecasts have been especially unreliable. Better guidance at that range could give energy operators, farmers, and emergency officials more time to prepare for conditions that affect supply, demand, and public safety.
Learning from forecast errors
The team’s method works as an added correction layer on top of an existing forecasting model. It studies that model’s past forecasts alongside historical data, learns its recurring errors, and adjusts its output accordingly.
Forecasts in this period are probabilistic: they describe the likelihood of a range of possible conditions rather than giving one certain outcome. As fresh forecasts and data become available, the correction model can adapt and improve how it adjusts predictions.
The researchers tested the approach with multiple forecasting systems. Their strongest competition entry combined the European Center for Medium-Range Weather Forecasts model with another model, then applied the learned correction. That submission doubled the original model’s predictive power in the subseasonal range and placed first in a forecasting competition held by the center.
Why weeks two to six are difficult
Short-range forecasts can use information about the atmosphere’s current state. Longer-term outlooks can draw on slowly changing influences. The weeks between those horizons have been called a “predictability desert” because they have less access to either kind of advantage.
The challenge also reflects the atmosphere’s complexity. In the 1970s, MIT Professor Edward Lorenz described how small differences in initial weather conditions can lead to very different outcomes. His work led to a theoretical two-week limit on how far detailed weather prediction could reach.
The research team’s approach does not remove that underlying uncertainty. Instead, it uses patterns in earlier forecasting errors to extract more useful information from models in a range where forecasts have often performed no better than long-term averages.
Extreme weather is part of the test
Improving average forecasts is only one goal. The researchers also reported a statistically significant improvement in predicting extreme weather, including events such as flooding and cold spells.
That result matters because conventional error correction can pull predictions toward the average, potentially making unusual conditions harder to capture. The team’s machine-learning layer improved extreme-event predictions while also improving time-averaged forecasts.
The potential value of more lead time is illustrated by Texas’s 2021 winter storms. The resulting energy infrastructure failure led to shortages of food, water, heat, and electricity, and at least 246 people died. The article notes that earlier preparation, such as bringing more trucks of natural gas into disrupted areas, could have mitigated damage.
A layer that can travel across models
The researchers found that their correction technique improved land forecasts across most regions of the globe. They also applied it to other forecasting models, including models that already used a different AI-based correction, and saw improvements there too.
That flexibility comes from its design as a wrapper around a forecasting system. The team says it can be applied beyond the model used in the competition and can bring less capable systems closer to the performance of the European Center for Medium-Range Weather Forecasts forecasts.
The model was built using MIT Engaging, a supercomputing cluster hosted by MIT’s Office of Research Computing and Data, but running the finished model does not require a supercomputer. The researchers describe the work in a paper in Nature and say the tool is open source and low cost.
Subseasonal forecasts are not guarantees, but more accurate probabilities could help organizations make decisions earlier. The study’s results suggest that learning systematically from model errors may make the two-to-six-week outlook more useful, including when the forecast concerns uncommon and consequential weather.