Machine Learning Based Sub-Seasonal Forecasting
Abstract
Project AbstractApproved for Public ReleaseRecent advances in artificial intelligence and machine learning have allowed researchers to conquer seemingly insolvable problems. The recent success of our Navy-funded deep-learning based global weather forecast model su ggests that Machine-Learning-based General Circulation Models (MLCGMs) could be the key to improving sub-seasonal weather forecasts. Here we propose to develop a sub-seasonal forecast system based on ML-CGMs.Society and the Navy need better sub-seasonal forecasts. Such forecasts, which extend out to lead times of two to six weeks and predict anomalies averaged over periods of one or two weeks, are widely recognized as the most challenging and least skillful weather forecasts currently being attempted. Shorter-term forecast s gain predictability from our knowledge of the initial atmospheric state. Longer-term seasonal forecasts are aided by information a bout atmospheric forcing from phenomena such as El Nino and the circumstance that they are for anomalies averaged over months or se asons. Skillful sub-seasonal forecasts could improve planning for extreme events such as floods and droughts, for weather-related he alth effects such as heat waves, for energy demand, and for agriculture. From the perspective of naval operations, S2S forecasts can provide strategic guidance about likely weather conditions during multi-week campaigns.Prediction on sub-seasonal time scales unav oidably requires the estimation of probable atmospheric conditions by creating a suite of likely forecasts (an ensemble) that reflec ts our uncertainty about the initial state of the atmosphere and about the precise factors governing its evolution (such as uncertai nties in the representation of clouds). State-of-the-art Numerical Weather Prediction based General Circulation Models (NW- PGCMs) r equire too much computational time to allow the generation of ensembles of sub-seasonal forecasts with more than 50 members. Such en sembles are too small to adequately sample the full range of possible weather events. In particular, extreme weather is a low-probab ility event that is not be well-sampled in a 50-member ensemble. Our current ML- GCM is able to create probabalistic sub-seasonal fo recasts that are roughly comparable with those from the best NWPGCMs, yet can do so in at least two orders of magnitude less computi ng time. In the proposed research we will continue to refine our MLGCM and examine the improvements that can be made in sub-seasonal forecasting using ensembles of unprecedented sizewith thousands of members.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2021
- Source ID
- N000142112827
Entities
People
- Dale R. Durran
Organizations
- Office of Naval Research
- United States Navy
- University of Washington