Extending Data-Driven Ensemble Forecasts to Subseasonal Timescales

Ensemble weather prediction systems have been the standard for decades. Deep learning approaches have shown promise for medium range weather forecasts, but currently published models are not more skillful than dynamical models such as ECMWF ENS at lead times beyond day 16. We discuss various challenges of extending these methods beyond week and we also present a novel modeling framework. The framework, HyDiff combines diffusion based models with ensemble forecasts generated by a statistical model in order to generate skillful ensemble forecasts out to 35 days. We have found that these models train far more efficiently than state of the art methods, while producing forecasts that are more skillful than ECMWF ENS and NOAA GEFS on subseasonal timescales.