AI Tool Doubles Accuracy of Two-to-Six-Week Forecasts
Weather forecasts are good for the next few days, but they become much less reliable beyond two weeks. A team including researchers at MIT has now found a way to improve forecasts for two to six weeks ahead, a period scientists call the subseasonal range. The work was published in Nature and shared by MIT News on 14 October.
The team did not build a new weather model. Instead, it created a machine-learning 'wrapper' that learns from the past mistakes of an existing model and corrects its output. When the team applied it to the model of the European Centre for Medium-Range Weather Forecasts (ECMWF), it doubled forecasting accuracy in this time frame. A submission using the method placed first in an ECMWF forecasting competition.
For the first time, forecasts in this range beat climatological averages, the long-term weather patterns usually used as a baseline. MIT research scientist Judah Cohen calls the two-to-six-week window the 'predictability desert', because it falls between short-range forecasts and long-range seasonal outlooks.
The method also gave a statistically significant improvement in predicting extreme weather, such as flooding or cold spells. That matters, because older error-correction methods often made extreme-weather forecasts worse.
The tool can be added to other models too. It is open source, low cost and does not need a supercomputer to run. Kenya's meteorological office has already contacted the team about using it to forecast its annual monsoon season.
Cohen says longer warning times could help farmers decide when to fertilise or how to protect crops from a coming freeze. Energy and emergency teams could also get weeks, instead of days, to prepare for a big storm.
