To maximize the operational utility of the findings, the ensemble system used for this research is configured to match that of the AFWA MEPS as closely as possible. The details of the MEPS configuration are based on work by Hacker et al. (2011a), hereafter H11, in which several methods of producing IC and physics perturbations were examined with a goal of finding “the most skillful ensemble, with the least degree of complexity” such that it would be operationally viable given typical computational restraints. As with most operational NWP models, incremental changes have since been made to the MEPS configuration, but the basic setup exists as it did when it was closely replicated to create the runs for this research in late 2010 (Kuchera 2011, personal communication). The configuration used for the runs is described below, with further details and justification available in H11.
The ensemble consists of 10 WRF (ARW version 3.2) members with 4-km horizontal grid spacing and 42 vertical sigma levels. This high-resolution domain is nested within a larger 12-km grid spacing middle nest, which in turn is nested within a larger 36-km grid spacing outer nest. Each member obtains its ICs and lateral boundary conditions (BC) from a different member of NCEP’s Global Ensemble Forecast System (GEFS, Wei et al. 2008). H11 found that this method of direct dynamical downscaling from a global NWP model to create ICs did not perform as well as when more advanced methods, such as an ensemble-transform Kalman filter, are used. However, given the low computational expense and implementation in MEPS, it is used here. For its part, GEFS is constructed from the Global Forecast System (GFS) NWP model using an ensemble transform (ET) technique (Bishop 1999) that accounts for regional differences in analysis error variance from the operational 3D-var scheme by including regional scaling of the initial perturbation (H11).
Certain properties of the lower boundary (land surface) are assigned a different value in each member based on random draws from -like distributions, with distribution
parameters selected based on physical arguments and empirical data. These properties are the albedo, soil moisture availability, and roughness length, and the values assigned to each member do not change throughout the experiment. This technique was described by Eckel and Mass (2005), and led to small error reductions in lower tropospheric predictions when tested by H11 compared to when they were not used.
NWP model uncertainty can be considered distinct from IC or BC uncertainty in that it arises from, among other things, imperfect parameterizations of subgrid-scale processes (microphysics, planetary boundary layer fluxes, deep convection), radiative forcing (shortwave and longwave), and land-surface fluxes. Running a unique combination of parameterizations for each member is one way to sample this uncertainty, ultimately resulting in more skillful predictions. This approach was promoted by Eckel and Mass (2005), and H11 demonstrated its importance for near-surface predictions, stating the technique “appears critical for probabilistic prediction in the PBL (planetary boundary layer).” The specific parameterization combinations (hereafter called “physics suites”) should not be selected arbitrarily because some suites that were not tuned together during their development can produce unreasonable and even unstable predictions (H11). The 10 suites used in this work are given in Table 1. They are the same as those used in H11, although they are numbered differently, which is explained as follows. During the testing of various suites, H11 initially identified 20 that appeared to be most viable (stable, and producing reasonable predictions), later selecting the best 10 for inclusion, which are the 10 used here. However, in this work, the member number, which has no meaning aside from identification purposes, is from its number in the original 20. References for the physics options are found in Skamarock et al. (2008).
The cumulus parameterization listed in Table 1 is used on the middle- (12-km grid spacing) and outer- (36-km grid spacing) nests only; no cumulus parameterization is used for the 4-km inner nest.
The period of the study is from 21 November 2008 through 21 February 2009, with NWP runs initialized every three or four days to minimize highly-correlated cases.
in each ensemble member is downscaled from its parent member from the global ensemble suite, solid and liquid water phases are not initialized.
Table 1. Summary of physics suite used for each member.
Member Microphysics PBL Shortwave Longwave Land Surface
Cumulus
(none on inner-most
nest)
1 Kessler YSU Dudhia RRTM Thermal KF
5 WSM6 MYJ CAM RRTM Thermal KF
7 Kessler MYJ Dudhia CAM Noah BM
8 Lin MYJ CAM CAM Noah Grell
10 WSM5 YSU Dudhia RRTM Noah KF
11 WSM5 MYJ Dudhia RRTM Noah Grell
15 Lin YSU Dudhia CAM RUC BM
16 Eta MYJ Dudhia RRTM RUC KF
17 Eta YSU CAM RRTM RUC BM
19 Thompson MYJ CAM CAM RUC Grell
Since cloud water is the primary field of interest in the study of fog, the first six hours of each case are evaluated with caution to account for the spin up of the field to a stable state, and these hours are not included in certain parts of the verification where noted. As previously discussed, given the NWP-only nature of this framework, skillful predictions during the first few hours are not an emphasis of this work, and so we mainly focus on the 6–20 h prediction timeframe (2200–1200 LT) representing short-term operational planning.
Figure 2 shows the domain of each of the three nests. Verification focuses on seven airfields (Figure 3) in California and Nevada representing three regions with distinct mesoscale influences: Crescent City (airport identifier KCEC, elevation 17 m) and Arcata (KACV, 66 m) represent a coastal region as both are less than 1 mile from the Pacific Ocean; Stockton (KSCK, 9 m), Modesto (KMOC, 29 m), and Merced (KMCE, 57 m) represent a valley region subject to frequent and heavy overnight radiation fog; and Emigrant Gap (KBLU, 1610 m) and Reno represent a mountainous region, with both
sites at relatively high elevations and surrounded by mountainous terrain. The NWP predictions for any given level at these seven sites are obtained by bi-linearly interpolating from the four grid points laterally surrounding each station. In most cases, NWP values from the lowest model layer or the 2-m level are of most interest. The lowest model layer (hereafter layer 1) exists at a height of 19–21 m above the model’s ground level. WRF post-processing computes 2-m values of temperature and water vapor from the heat and moisture fluxes provided by the PBL scheme using the flux- profile relationship (Stull 1988).
Figure 3. Location of verification sites (with elevation in meters). (Map background courtesy of Europa Technologies, Google, and INEGI 2011).