In February 2021, three winter storms in Texas left millions of people without electricity and heat; at least 246 people died. Power engineers could have prepared for the frost, but for that they needed a reliable forecast two weeks ahead, not two days. A new development by researchers from the Massachusetts Institute of Technology and Microsoft may change this situation.
Meteorologists call forecasts for a period of two to six weeks a "desert of predictability." Back in the 1970s, the mathematician Edward Lorenz showed that, owing to the chaotic nature of the atmosphere, predicting the weather beyond two weeks is practically impossible. Over short periods, data on the current state of the atmosphere help; over long ones, data on slow processes in the ocean. But in the interval between them, these methods do not work.
The American researchers did not set out to build a new model from scratch. Instead, they developed an algorithm that analyzes past errors of existing forecasting systems and makes corrections to their calculations. This system, based on machine learning, works as an add-on to any meteorological model and becomes more accurate over time as it accumulates more data on the forecasts made and the actual weather.
To test their development, the team entered it in a competition run by the European Centre for Medium-Range Weather Forecasts, one of the world's leading centers. The system that combined the center's model with the new correction algorithm posted the best result: the accuracy of forecasts for a period of two to six weeks doubled. For the first time in a long while, a model predicted the weather over this interval better than calculations based on multi-year averages for the specific season.
Until now, no one could show that a model predicts the weather better than multi-year averages over this interval. We consistently beat climatology.
— Judd Cohen, one of the study's authors
According to the scientists, the algorithm predicts extreme events especially well: heat waves, frosts or floods, which conventional correction methods often smooth over. More accurate forecasts could benefit entire industries. Farmers will be able to better plan irrigation, and energy companies will be able to prepare grids for peak loads. Referring to conversations with representatives of the energy industry, Judd Cohen explains that two weeks would be enough for them to prepare for an ice storm.
The algorithm has been published openly. Kenya's meteorological service has already shown interest in the development, where an accurate forecast of the rainy season will help raise crop yields. Before the system becomes a standard in meteorology, it must be tested in different climate zones, from the tropics to the polar regions.
Letters