Researchers improve weather forecasts looking two to six weeks out

In 2021, Texas experienced its worst-ever energy infrastructure failure due to three severe winter storms in February. The catastrophe resulted in over 246 deaths and shortages of food, water, heat, and electricity. Energy officials had limited time to prepare for the storms, which could have been mitigated by bringing in more natural gas.

Weather forecasts beyond two weeks are notoriously unreliable, creating a tight window for energy operators, farmers, and emergency response officials worldwide. A new technique involving researchers at MIT significantly improves weather forecasting beyond two weeks. The researchers created a machine-learning model that learns to correct errors in existing forecasting models, doubling their output over the two-to-six-week time frame.

This groundbreaking approach earned first place in a prominent international forecasting competition when applied to a gold-standard model. The research team's findings, published in Nature, demonstrate the potential for forecasts that surpass climatologic averages between two and six weeks. This development is crucial for farmers and energy industry professionals who need to balance supply and demand.

Co-author Judah Cohen, an MIT Research Scientist, highlights the significance of predicting both general weather patterns and extreme events like flooding or cold spells. The technique builds upon chaos theory, which Edward Lorenz developed in the 1970s, suggesting a two-week limit for weather prediction. Cohen argues that the atmosphere often falls into predictable patterns that can be forecasted.

Current weather forecasting relies on either dynamical models, which simulate the atmosphere and oceans, or artificial intelligence models that learn from historical data. While AI models have improved short-term forecasts, they perform poorly in longer ranges. Cohen focuses on subseasonal forecasting, the period between two and six weeks, known as the "predictability desert."

In this range, forecasters struggle to take advantage of the atmosphere's long-term forces. Cohen attributes the initial idea of using machine learning to correct forecast errors to his co-author Lester Mackey, with whom he has collaborated for over a decade. The researchers developed a postprocessing layer using AI to learn from forecasting systems' mistakes based on historical data and adjust their forecasts accordingly.

In their tests, the postprocessing layer doubled the predictive power of the original model in the subseasonal range, for the first time outperforming climatologic averages.