Impact of formulation and resolution on ensemble forecast performance

A number of configurations of the Navy Operational Global Atmospheric Prediction System (NOGAPS) ensemble forecasting system were run for summer 2007 and 2008 seasons. Experiments were run testing ensembles with different initial perturbation formulations and different representations of model uncertainty. Ensembles were run at 110, 83, and 55km horizontal resolutions with 33, 17, and 9 members, respectively, to investigate resolution vs. number of member tradeoffs. Initial perturbations were formulated based on different implementations of the ensemble transform method. Model perturbations include stochastic forcing, parameter variations, and a simple model of the diurnal cycle in SST. In this talk, the focus will be on the impact on lower and upper tropospheric wind forecasts and tropical cyclone forecasts. Results indicate that probabilistic skill scores, such as the brier score, show significant sensitivity to ensemble design, while ensemble mean errors are less sensitive to ensemble design specifics. It is found that both stochastic perturbations as well as parameter variations can result in improved spread-skill relationships, and stochastic forcing has a significant impact on probabilistic forecasts of low-level winds exceeding a certain threshold. For tropical cyclone tracks, the increase in resolution from 110 km to 83 km gives a significant improvement in ensemble mean performance. An important result is that no particular formulation performs best under all the metrics considered, highlighting the fact that optimal ensemble design will be different for different sets of metrics.