Machine-learning fix doubles accuracy of weather forecasts two to six weeks out

A fix for the 'predictability desert'

Weather forecasts beyond two weeks are notoriously unreliable, which leaves energy operators, farmers and emergency officials with little time to prepare. MIT News 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. Had energy officials had more warning, for example to bring in more trucks of natural gas, damage might have been reduced.

The new technique targets what is known as subseasonal forecasting, the range of two to six weeks. MIT Research Scientist Judah Cohen, a co-author, calls this range the "predictability desert". In the short range, forecasters can use the initial conditions of the atmosphere, and over longer terms they can use slow-varying boundary forces such as El Niño. The subseasonal range benefits from neither. "In most cases forecasting in the subseasonal ranges is no different than going to a casino and rolling the dice," Cohen says.

How the postprocessing layer works

Today's best forecasts come from dynamical models, which numerically simulate how the atmosphere and oceans evolve, or from AI models trained on historical weather data. According to the article, AI models have improved short-range prediction but do worse than dynamical models at longer ranges.

The researchers built a machine-learning model that wraps around an existing forecasting system. Cohen says the idea came from co-author Lester Mackey, who calls it postprocessing. The layer uses AI to learn from a system's past mistakes in historical data and then shifts its forecasts. Other systems apply simpler error correction. The forecasts are probabilistic, meaning they give the likelihood of a range of possibilities. "As more forecasts and data become available, our model adaptively improves its correction," Cohen says.

The team trained the model on MIT Engaging, a supercomputing cluster, but the resulting model does not need a supercomputer to run.

Results in an international competition

Last year the team submitted three models with the postprocessing layer to the European Center for Medium-Range Weather Forecasts (ECMWF), which the article describes as considered the best forecasting center in the world. The first-place submission combined the ECMWF forecasting model with another model and applied the correction on top. This doubled the original model's predictive power in the subseasonal range.

It was also the first time a model improved on relying on climatological averages. "Up until now, no one had been able to show a model predicts weather better than long term averages for these subseasonal forecasts," Cohen says.

The model also showed a statistically significant improvement in predicting extreme weather such as flooding or cold spells. Cohen notes that when ECMWF built its own non-AI error correction, the model got worse at predicting extremes. He suggests this is because traditional bias correction pulls toward the mean, while extreme weather sits at the edge of probabilities.

Broad use and what comes next

In the Nature paper, the team showed the technique could improve land forecasts in most regions across the globe. Applied to other models, it improved those as well, bringing lesser models on par with ECMWF forecasts. Cohen describes it as an add-on wrapper that can be used with any model, including those already bias corrected with another AI algorithm, and says it is open source and low cost.

Cohen says the meteorological office of the Republic of Kenya has already asked about using the model to improve forecasts of the annual monsoon season and agricultural yields this winter. He sees uses in deciding when to fertilize around rain, protecting crops from freezes and avoiding droughts, as well as in energy and emergency management. He quotes energy officials who say two days of warning for an ice storm is not enough, while two weeks could be.

The lead authors are Hannah Guan, Soukayna Mouatadid, Genevieve Flaspohler, Paulo Orenstein, Jeremy Berman, Jakob Schloer and Joshua Talib, with co-authors from Microsoft and MIT. 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.