10.1 Why does including a model-error representation improve probabilistic forecasts?

Berner, Judith, K. R. Smith, National Center for Atmospheric Research (NCAR), J. P. Hacker, Naval Postgraduate School, and C. Snyder, NCAR

The performance of five different model-error schemes and selected combinations is verified for probabilistic forecasts with the WRF-ARW model over the Contiguous United States. Including a model-error representation leads to more spread and small, but significant increases in forecast skill. In the free atmosphere, a stochastic kinetic-energy backscatter scheme performs best, while multiple-physics schemes tend to be superior near the surface. Combining stochastic and multiple deterministic parameterization results in the biggest improvement throughout the atmosphere.

To investigate which component of the Brier score is responsible for the improvement, it is decomposed into reliability, resolution and uncertainty for both

raw and de-biased ensemble forecasts.

We find that the different simulations have similar resolution, but can be discriminated with regard to reliability. All forecasts are calibrated to have the same variance as the observations in order to determine if the primary benefit of including a model-error scheme results from an increase in ensemble spread.