Implementation of a FourCastNet v2 Ensemble System to Improve Consensus Weather Forecasting at The Weather Company

A number of global deep learning numerical weather prediction (DLNWP) systems have shown skill metrics for prediction of key surface and atmospheric weather parameters that are improved over those of traditional NWP systems. At The Weather Company, we use a consensus weather forecasting system that optimally combines forecasts from different systems to create a forecast blend that generally outperforms any individual system. Recently, we have incorporated two DLNWP models in our consensus weather forecasting system and realized improved skill for surface temperature and wind speed forecasts.

In this work, we wish to evaluate the impact to our consensus forecasting system by incorporating NVIDIA’s FourCastNet v2 (FCNv2). FCNv2 is designed to be a deterministic weather forecast emulator. An ensemble version of FCNv2 has recently been developed and calibrated by NVIDIA and Lawrence Berkeley Lab, using both optimal initial condition perturbations and a multi-model ensemble with multiple trained model checkpoints.

The Weather Company is collaborating with NVIDIA to implement the FCNv2 ensemble into TWC's consensus weather forecasting system. In this talk we will describe the FCNv2 configuration and illustrate output for a number of key weather parameters. We will review verification metrics, both for the FCNv2 ensemble output as well as the impact of the ensemble output on the TWC forecast blend, and discuss implications for the future of NWP using deep learning models.