12.3 EVALUATING NWP ENSEMBLE CONFIGURATIONS FOR AT&D APPLICATIONS Jared A. Lee*, Walter C. Kolczynski, Tyler C. McCandless, Kerrie J. Long, Sue Ellen Haupt, David R. Stauffer, and Aijun Deng The Pennsylvania State University, University Park, PA 1. INTRODUCTION Due to the chaotic nature of the atmosphere, there are inherent limitations to forecasting a single realization of the future atmospheric state. While numerical weather prediction (NWP) models have become more advanced in recent years in representing and predicting the atmospheric state, they are still limited because of imperfect model numerics, imperfect parameterizations of unresolved physical processes, and interpolations of input data that is sparsely located compared to current model grid resolutions. In recognition of these difficulties, contemporary NWP uses ensembles of simulations. Members in these ensembles often differ by imposed initial conditions (ICs), lateral and lower boundary conditions (LBCs), model physics parameterization schemes, and even the choice of NWP modeling system. Grimit and Mass (2002) state that there is a strong positive correlation between the ensemble spread and forecast error for short-range mesoscale NWP. For example, if the spread across the ensemble is low, the forecast error (and hence, the forecast uncertainty) is generally low, and vice-versa. There are a few different approaches to ensemble initialization that are documented in the literature. Each approach attempts to account for uncertainty in the forecast. Several approaches are designed specifically to account for uncertainty in the initial conditions. Some of these include using bred vectors, singular vectors, an ensemble Kalman filter, ensemble transform Kalman filter, or the ensemble transform approach. There are several studies in the literature that compare the relative performance of ensembles that are initialized with these different methods, but the ensemble Kalman filter has generally performed best (Wang and Bishop 2003; Buizza et al. 2005; Bowler 2006; Descamps and Talagrand 2006; Wei et al. 2006). Another source of uncertainty for regional NWP (or limited-area models) arises from the specification of the lateral boundary conditions. Warner et al. (1997) summarize many of the issues that modelers must consider with regard to LBCs. Notably, errors that arise due to the LBCs “sweep” across the model domain through the integration period, which can degrade model results and constrain ensemble dispersion (Nutter et al. 2004). Therefore, if there is a particular region that is of interest to the modeler, that region should be placed far enough away from the lateral boundaries so that errors from the boundary will not have advected into that region at the time of interest. * Corresponding author address: Jared A. Lee, The Pennsylvania State University, Department of Meteorology, 503 Walker Building, University Park, PA 16802; e-mail: jal488@meteo.psu.edu. In addition to ICs and LBCs, another source of uncertainty in NWP results from the model physics parameterizations. Because NWP models are unable to resolve many physical processes that occur, such as convection, cloud and ice microphysics, atmospheric radiation, as well as processes in the atmospheric boundary layer (ABL), parameterization schemes for these processes are necessary. The land surface must be represented by discrete categories of typical soil and vegetation types, as well as soil moisture profiles. All of these processes must be approximated. These approximations introduce some amount of error to NWP solutions that is unavoidable. Judging by the focus of the major national centers in the 1990s on creating IC ensembles (e.g., Toth and Kalnay 1993; Molteni et al. 1996), it was once thought that IC uncertainty dwarfed physics uncertainty in importance, at least for global forecast models. Many studies in recent years, however, have shown the importance of physics uncertainty in NWP, particularly in limited-area models. As an example, studies investigating the impacts of physics parameterization schemes on NWP forecasts have been conducted for a variety of situations, including for predictions of the southwest monsoon (Bright and Mullen 2002), mesoscale convective systems (Jankov et al. 2005), and the passage of a mid- latitude cyclone (Deng and Stauffer 2006), just to name a few. Several studies have also investigated the performance of ensembles that incorporate multiple sources of uncertainty. Warner et al. (2000) and Jones et al. (2007) showed that physics variability is important in cases with weak synoptic forcing, and IC variability is important in cases with strong synoptic forcing. Physics variability can also be more important than IC variability in increasing ensemble spread in the very short term (6- 12 h) (Stensrud et al. 2000), and can have a greater impact on near-surface meteorological parameters than IC variability (Eckel and Mass 2005). Fujita et al. (2007) found that spread was greater for dynamic variables in an IC ensemble, and greater for thermodynamic variables in a physics ensemble. They concluded that because the distribution of spread in the IC and physics ensembles they created was so different, the ensembles likely covered different portions of the probability density function (PDF) of the atmospheric state. They recommended using a combined IC/physics ensemble as the best choice for severe weather and boundary layer forecasting, since it incorporates variability from multiple sources of uncertainty. One of the main messages that can be gleaned from these studies is that it is critically important to include both IC/LBC and physics uncertainty in short- range (0-48 h) NWP ensembles, in order to sample the forecast PDF of the atmospheric state more accurately.