Bringing Rain to the Subseasonal Forecasting Desert with Deep Learning Weather Prediction
, Professor, University of Washington, Department of Atmospheric Sciences
Sub-seasonal weather forecasts , at lead times of 2-6 weeks, are the least skillful forecasts currently being attempted. Improved sub-seasonal forecasts are needed for agricultural, energy, and water supply planning, and disaster preparedness.
Subseasonal prediction relies on the probabilistic estimation of a set of likely forecasts (an ensemble) reflecting our uncertainty about the initial state of the atmosphere and about the precise factors governing its evolution. Global Numerical Weather Prediction (NWP) models produce such forecasts by approximating solutions to mathematical equations. Yet, even using the most powerful computers, they require too much time to generate ensembles with more than 50 members. We recently developed deep-learning-based models that approach the accuracy of NWP. Using only a few GPUs, the deep-learning approach is orders of magnitude faster, allowing the use of much larger ensembles to sample future weather patterns.