Machine learning for sub-seasonal weather forecasting


Sub-Seasonal to Seasonal (S2S) climate prediction has long been a gap in operational weather forecasts. The S2S timescale varies from two weeks to an entire season, although some have recently used the term more broadly to include seasonal forecasts up to 12 months ahead. S2S is considered more challenging than both numerical weather prediction (NWP) (1-15 days) and seasonal forecasts (2-6 months) due to the limited predictive information from land and ocean indicators and the weak predictive signal from the atmosphere. Improving S2S forecasts would significantly impact downstream applications such as streamflow forecasting, heatwave prediction, water resource management, and in-season climate-aware crop modeling on the sub-seasonal time scale. 

In this project, we propose to use a set of machine learning methods to improve the skill (or accuracy) and usability of S2S data products. We combine ensemble physics-based S2S forecasts with historical land and ocean variables. This includes things like soil moisture, sea surface temperature, El Niño indices, and other variables not usually incorporated in physics-based models, such as sea ice extent and snow coverage indices. Our goal is to improve the skill of forecasting temperature and total precipitation from 3 to 6 weeks ahead compared to current computational fluid dynamical forecasting models. Additionally, we produce extreme climate indicators to improve the communication of our forecasts.