MIT News - Artificial intelligenceSource publication:
Loading content…
A machine-learning postprocessing layer that learns to correct errors in existing forecasting models doubles the accuracy of a gold-standard model over the two-to-six-week subseasonal range and, for the first time, consistently beats climatological averages.
Subseasonal forecasting now has a working correction: in the two-to-six-week window known as the predictability desert, forecasts beat climatology consistently for the first time. Previously, forecasts in that range were mostly worse than climatological averages, and people questioned whether investing in models for this time range was worthwhile at all. The team submitted three models with the postprocessing layer to the European Center for Medium-Range Weather Forecasts (ECMWF) seasonal forecasting competition; the entry combining ECMWF's model with another and then applying the correction placed first, doubling the original model's predictive power.
The approach wraps a learning layer around an existing forecasting model: it quantifies a system's systematic errors from historical data and shifts that system's forecasts accordingly. Other forecasting systems do simpler error correction, whereas this postprocessing layer adaptively improves its correction as more forecasts and data become available. Applied to other models, the correction improved them as well, putting even lesser models on par with ECMWF's forecasts; training used MIT's Engaging supercomputing cluster, but the resulting model does not require a supercomputer to run.
The correction also improves prediction of extreme weather, which may be the most socially impactful part. ECMWF's own non-AI error correction got worse at predicting extreme weather, because traditional bias correction always tries to reach the mean while extreme weather sits on the fringe of probabilities. The improvement in extreme-weather prediction is statistically significant, covering events such as flooding and cold spells.
The technique targets two-to-six-week subseasonal forecasting for energy, agriculture, and emergency management, and the work shows it improves land forecasts in most regions across the globe and can be applied as an add-on layer to other models. The meteorological office for the Republic of Kenya has reached out about implementing the model to improve monsoon-season forecasting and agricultural yields this winter.
The model was tested in the ECMWF seasonal forecasting competition, so its improvement is stated in terms of that competition's setup and the model combination used; the extreme-weather improvement is described as statistically significant, without a quantified range beyond the event types named.