A team including MIT researchers has developed a machine learning method to improve weather forecasts two to six weeks ahead. MIT says the method helped a forecasting system outperform long-term climate averages in that range and improved its predictions of extreme weather.
The method works on forecasts made by existing models. It learns from their errors in historical data, then adjusts the probabilities they assign to possible weather outcomes. According to MIT, the corrections can adapt as more forecasts and data become available.
The method doubled a model's predictive power
In a paper published in Nature, the researchers tested the approach in a forecasting competition run by the European Center for Medium-Range Weather Forecasts. Their first-place submission combined the center’s model with another forecasting model, then applied the correction. MIT says it doubled the original model’s predictive power over the two-to-six-week range.
The team also reported a statistically significant improvement in predicting extreme weather. Tests showed better land forecasts in most regions worldwide, according to MIT. The correction improved other forecasting models the researchers tried as well.
Better forecasts could give officials time to plan
More reliable forecasts weeks ahead could give farmers, energy operators and emergency officials more time to plan. MIT research scientist and co-author Judah Cohen cited decisions such as when to fertilize crops before rain or prepare for a freeze. Those are potential uses, not reported outcomes of the study.
Cohen says Kenya’s meteorological office has contacted the team about using the method for monsoon forecasts. He also says the model is open source and low cost, and does not need a supercomputer to run.
