This expository paper documents an experimental, real-time, 10-member, 3-km, convection-allowing ensemble prediction system (EPS) developed at the National Center for Atmospheric Research (NCAR) in spring 2015. The EPS is particularly unique in that continuously cycling, limited-area, mesoscale ensemble Kalman filter analyses provide diverse initial conditions. In addition to describing the EPS configurations, initial forecast assessments are presented that suggest the EPS can provide valuable severe weather guidance and skillful predictions of precipitation. The EPS output is available to operational forecasters, many of whom have incorporated the products into their toolboxes. Given such rapid embrace of an experimental system by the operational community, acceleration of convection-allowing EPS development is encouraged.
Since April 2015, the National Center for Atmospheric Research (NCAR) has been producing real-time, experimental, 10-member, 48-h ensemble forecasts with 3-km horizontal grid spacing over the conterminous United States (CONUS). Forecasts are initialized at 0000 UTC daily and will be produced in real time through at least July 2016.
While NCAR’s system is not the first real-time convection-allowing ensemble prediction system (EPS), its application of a limited-area, continuously cycling mesoscale ensemble Kalman filter (EnKF; Evensen 1994) data assimilation (DA) system to initialize the EPS contrasts the initialization approach of all other current real-time convection-allowing EPSs, which leverage external models to produce diverse initial conditions (ICs). This traditional reliance on external models for convection-allowing EPS ICs originated in spring 2007, when the Center for Analysis and Prediction of Storms at the University of Oklahoma produced pioneering real-time, 10-member, convection-allowing ensemble forecasts (Xue et al. 2007) that derived IC perturbations from NCEP Short Range Ensemble Forecast (SREF; Du et al. 2009) system members, and in the following years, other groups also chose to initialize real-time convection-allowing EPSs from external models, including the Air Force Weather Agency (Kuchera et al. 2014), Met Office (Tennant 2015), Deutscher Wetterdienst (Gebhardt et al. 2011; Peralta et al. 2012; Kühnlein et al. 2014), Météo-France (e.g., Bouttier et al. 2012), and NOAA’s National Severe Storms Laboratory.
Thus, NCAR’s use of a continuously cycling EnKF to initialize a pseudo-operational convection-allowing EPS for CONUS-scale prediction is unique,1 and there are several reasons for this choice. First, EnKFs allow a seamless integration of DA and ensemble forecasting that provides dynamically consistent initial ensembles, meaning the role of external models can be relegated solely to providing boundary conditions. This initialization approach may be preferable to those employing external analyses, which could have significant mismatches with respect to many aspects of the forecast model. Second, a continuously cycling EnKF configuration permits an assessment of model climatology and will expose model weaknesses. While a partial-cycling technique (e.g., Rogers et al. 2009) may produce better forecasts by effectively eliminating biases that can accumulate in continuously cycled systems (e.g., Hsiao et al. 2012; Romine et al. 2013), partial-cycling systems do not enable a robust assessment of model bias, as the forecasts partially reflect the quality of the external model periodically providing ICs. Thus, through DA statistics, such as innovation magnitude (i.e., difference between observations and model-simulated observations), continuously cycled EnKF systems can identify model biases such that they can be remedied (e.g., Romine et al. 2013).
Finally, NCAR scientists have performed several month-long trials in which continuously cycling, mesoscale EnKFs initialized convection-allowing ensemble forecasts (e.g., Romine et al. 2014; Schwartz et al. 2014), including a real-time demonstration of an EnKF-based convection-allowing EPS during the 2013 Mesoscale Prediction Experiment (Schwartz et al. 2015; Weisman et al. 2015). These experiments and other case studies (e.g., Jones and Stensrud 2012; Schumacher and Clark 2014; Jones et al. 2015) indicated that limited-area, continuously cycling EnKFs can be robust and initialize skillful and valuable convection-allowing forecasts, which inspired this year-round initiative at NCAR to demonstrate how a future operational EnKF-based convection-allowing EPS could be designed and implemented.
In light of this effort, this paper documents NCAR’s real-time convection-allowing EPS configuration and describes preliminary characteristics of the forecasts.
2. Forecast model and analysis system configurations
NCAR’s real-time EPS consists of separate, but integrated, analysis and forecast components. In the analysis component, an ensemble adjustment Kalman filter (EAKF; Anderson 2001, 2003) implemented in the Data Assimilation Research Testbed (DART; Anderson et al. 2009) software updates ensembles of forecasts from the Advanced Research version of the Weather Research and Forecasting (WRF) Model (ARW; Skamarock et al. 2008). These EAKF analyses provide ICs for the forecast component. Aside from requiring external data from NCEP for lateral and lower boundary condition updates, the system is self-contained.
a. Analysis system configurations
Continuously cycling EAKF analyses update 50-member ensembles of WRF Model forecasts every 6 h. Table 1 lists various WRF Model and DA configurations, including which WRF Model variables are updated during each analysis. Additionally, DART’s forward operators use 10-m zonal and meridional wind components, 2-m water vapor mixing ratio and temperature, and surface pressure to compute model-simulated surface observations. Sea surface temperature and snow cover are updated from NCEP analyses daily at 0000 UTC.
To reduce spurious correlations due to sampling error, covariance localization forces EAKF analysis increments to zero a maximum of ~1280 km from an observation in the horizontal and 1.5 scale heights (log pressure coordinates) in the vertical using a piecewise polynomial function [Eq. (4.10) of Gaspari and Cohn (1999)]. For grid points where more than 2000 observations are contained within the ellipsoid defined by the vertical and horizontal localization radii, localization distances are reduced to approximate an ellipsoid that contains 2000 observations, assuming the observations are uniformly distributed. Sampling error correction (Anderson 2012) further reduces deleterious impacts from a limited ensemble size. Temporally and spatially evolving adaptive state-space inflation (Anderson 2009) is applied to prior (before assimilation) ensembles to maintain ensemble spread, and before forward operators are applied, inflation values are damped according to I* = (1 − α) + αI, where α = 0.9, and I and I* are the prior inflation values before and after damping, respectively.
A variety of surface and upper-air observations are assimilated (Table 1). Upper-level and surface moisture observations are assimilated as specific humidities and dewpoints, respectively. Observations taken within ±1.0 h of each analysis are assimilated, except the time window is shortened to ±0.5 h for surface observations. All observations are assumed valid at the analysis time, and for a typical 0000 UTC analysis, 66 000–70 000 observations are assimilated (Fig. 1a).
Observation errors generally follow NCEP’s specifications for the Global Forecast System (GFS; Romine et al. 2013, 2014). To enhance stability and minimize near-boundary magnitudes of analysis increments, observations within 75 km of lateral boundaries have their observation errors multiplicatively inflated by up to a factor of 5. Furthermore, surface observations are rejected if their station heights differ by more than 300 m compared to the model terrain. Aircraft (satellite wind) observations are superobbed in 60 km × 60 km × 25 hPa (90 km × 90 km × 25 hPa) boxes. Finally, an observation is rejected if the ensemble mean prior innovation exceeds 3 times the square root of the sum of the observation error variance and ensemble variance of the simulated observation.
b. WRF Model configurations
1) Configurations common to the analysis and forecast components
Version 3.6.1 of the ARW with 40 vertical levels and a 50-hPa top produces all weather forecasts in both the analysis (i.e., section 2a) and forecast components. Both components employ monotonic moisture advection and identical physical parameterizations (Table 1).
Ensemble forecasts are produced in the analysis and forecast components. Each ensemble member in both components uses identical physics and dynamics options. This single-physics EPS means each member is equally likely to represent the “truth,” which is not the case for multiphysics or multimodel EPSs. Therefore, while multiphysics ensembles may engender more spread than single-physics ensembles (e.g., Stensrud et al. 2000), the violation of equal likelihood muddles interpretation of probabilistic forecasts.
Unique lateral boundary condition (LBC) perturbations are derived for each member by taking random draws from global background error covariances included within the WRF Model DA system (Barker et al. 2012) and adding them to 0.25° GFS analyses and forecasts (Barker 2005; Torn et al. 2006). Although global EPSs are increasingly available, obtaining LBC perturbations from global ensembles is impractical in real time because of the absence of global EPSs with sufficient resolution, timeliness, and ensemble size (Romine et al. 2014).
2) Unique analysis component configurations
Within the analysis component, the WRF Model produces 6-h, 50-member ensemble forecasts to advance the ensemble to the next analysis time, when it is updated by observations. These 6-h forecasts employ a single domain with 15-km horizontal grid spacing (415 × 325 grid points) centered at 39°N, 101°W (Fig. 1b). The time step is 60 s. Thus, the ratio of the time step [seconds (i.e., 60 s)] to the horizontal grid spacing [kilometers (i.e., 15 km)] is 4, smaller than that typically used in WRF Model simulations and chosen to maintain stability during continuous cycling. For each member, LBC tendencies are derived by differencing perturbed 6-h GFS forecasts and perturbed GFS analyses, and soil states evolve freely.
3) Unique forecast component configurations
The forecast component produces 48-h WRF Model forecasts with a 3-km horizontal grid-spacing nest (1581 × 986 grid points) embedded within the 15-km analysis domain (Fig. 1b). To lessen computational cost, the time step is 75 s in the 15-km domain (compared to 60 s in the analysis component) and 5 s in the 3-km nest. Cumulus parameterization is turned off on the 3-km grid, and two-way feedback links the 15- and 3-km grids.
The 15-km LBC tendencies are derived from perturbed GFS forecasts at 3-h intervals (e.g., the difference between perturbed 6- and 3-h GFS forecasts), rather than 6-h intervals, as in the analysis component’s WRF Model settings. Moreover, LBC perturbations are multiplicatively inflated as the forecast progresses (e.g., Torn 2010), where the LBC perturbation inflation factor increases linearly from 1.0 at the analysis time to 1.2 at 48 h, such that at 48 h, LBC perturbations are 20% larger than their initial values. The 15-km domain provides LBCs for the 3-km nest.
c. Initialization and output
Each 0000 UTC EAKF analysis initializes a 10-member ensemble of 48-h WRF Model forecasts over the nested domain. The 3-km ICs are provided by interpolating 15-km EAKF analysis ensembles onto the 3-km grid. While 50-member ensemble analyses are produced, computational constraints limit the 48-h forecasts to just 10 ensemble members, which is nonetheless sufficiently large to produce skillful and valuable forecasts (Clark et al. 2009, 2011; Schwartz et al. 2014). The 48-h, 10-member, 3-km forecasts are the end product of the system.
Many diagnostics, including hourly maximum (e.g., Kain et al. 2010) products and novel fields, such as 0–3 km AGL updraft helicity (UH), 0–1 km AGL relative vorticity, and hail size, are output during WRF Model integration. Additional fields, like simulated satellite imagery, are derived through postprocessing. These and other products, focusing on probabilistic interpretation of the 3-km ensemble guidance, are uploaded to the Internet in real time (http://ensemble.ucar.edu) and provided to collaborators.
3. Preliminary forecast evaluation
Rigorous verification activities assessing ensemble performance will occur in upcoming months and years as the dataset grows. However, in this section, using both subjective and objective techniques, some preliminary results focusing on accumulated 12–36-h forecasts are presented to highlight broad performance characteristics of next day predictions. The objective verification was achieved by comparing precipitation forecasts to NCEP stage IV analyses (Lin and Mitchell 2005) and computing areas under the relative operating characteristic (ROC) curve (Mason and Graham 2002), attributes statistics (Wilks 2006), and multiplicative frequency biases over the verification region (Fig. 1b), aggregated over all hourly 12–36-h forecasts initialized between 7 April and 5 July 2015. To compute these statistics, the model and observations must be on a common grid, so before computing these metrics, the 3-km precipitation fields were interpolated onto the stage IV grid using a budget interpolation algorithm (e.g., Accadia et al. 2003).
Average 12–36-h ensemble mean and probability-matched mean (Ebert 2001) accumulated precipitation between 7 April and 5 July 2015 indicates that, climatologically, the ensemble generally places precipitation in correct locations with reasonable amplitudes (Fig. 2). Multiplicative biases (Fig. 3a) reveal overprediction for precipitation rates ≥5.0 mm h−1 and underprediction for lower rates, and ROC areas, computed both before and after “neighborhood approach” postprocessing (e.g., Schwartz et al. 2014), were >0.5 for all thresholds, indicating the EPS has discriminating ability (Fig. 3b). Although subjectively, the ensemble spread of precipitation appears larger than in previous, similarly configured NCAR systems (Romine et al. 2014; Schwartz et al. 2014), attributes statistics (Figs. 3c,d) indicate an overconfident ensemble and suggest underdispersion.2
Various surrogate severe weather diagnostic fields are derived from the EPS to produce probabilistic forecasts of severe weather hazards (e.g., Sobash et al. 2011). For example, Fig. 4 shows the union of smoothed ensemble probabilities (EPs) of hourly maximum 2–5-km UH ≥ 75 m2 s−2, surface wind ≥ 25 m s−1, and hail ≥ 2.54 cm within ~40 km of each point, accumulated over 12–36-h forecasts and overlaid with National Weather Service (NWS) severe weather warning polygons issued over the same periods.3 The cases in Fig. 4 were chosen to present forecasts of severe weather occurring across a variety of storm modes and geographic areas in both strongly and weakly forced convective regimes. Generally, the ensemble forecasts identified areas that experienced severe weather, although there were some misses and false alarms. Calibration of these forecasts is ongoing to identify optimal thresholds and generate reliable probabilistic guidance for severe weather forecasters.
Subjective forecast evaluation suggests the EPS has difficulty maintaining forward-propagating mesoscale convective systems (MCSs) that develop in weakly forced environments. For example, the 0000 UTC 27 April 2015 forecast correctly initialized convection over eastern Texas (Figs. 5a,e), but convection decayed with time, rather than accelerating eastward and organizing into a severe MCS over Louisiana (Figs. 5b,c,d,f–h). This behavior has also been identified on other occasions (e.g., in Fig. 4a, the NWS warnings in southeastern Texas are well east of the highest ensemble probabilities). Nonetheless, the ensemble has demonstrated valuable forecast guidance regarding strongly forced MCS events, although MCS propagation speeds often remain slower than those observed (e.g., Figs. 5i–p). Future work will aim to determine whether this behavior is attributable to WRF Model microphysics and attendant cold-pool strength, as underprediction at light precipitation thresholds (Fig. 3a) is consistent with insufficient stratiform precipitation and too-shallow cold pools.
NCAR’s convection-allowing EPS represents a unique approach for how future operational convection-allowing EPSs may be designed and is distinct from paradigms leveraging external operational analyses for convection-allowing EPS ICs. Preliminary results indicate NCAR’s system has discriminating ability across a range of precipitation intensities and generates useful guidance for high-impact convective weather events. Future work will directly compare NCAR’s EPS with other traditionally initialized high-resolution EPSs, which should be enlightening.
Even though NCAR’s EnKF-initialized forecasts are experimental, they have already proven useful to NWS offices around the CONUS, suggesting this ensemble system fills a current void. Given this rapid embrace by the forecasting community, we believe development activities regarding operational convection-allowing EPSs should be accelerated and hope this long-term, real-time demonstration spurs discussion and advancement.
In no particular order, thanks to NCAR’s Davide Del Vento, Al Kellie, Rich Rotunno, Carter Borst, David Edwards, Michael Moran, Wei Wang, Stan Trier, David Ahijevych, Jeff Anderson, Chris Snyder, Greg Thompson, and Chris Davis and Steve Weiss (SPC) for their support and collaboration. We appreciate the comments from two anonymous reviewers. Computing is performed on NCAR’s Yellowstone (ark:/85065/d7wd3xhc) supercomputer.
NCAR is sponsored by the National Science Foundation.
Not all areas under NWS severe weather warnings actually receive severe weather, but we believe subjective verification against NWS warnings is an efficient way of assessing model forecasts of severe weather, particularly over regions covered by NWS watches. Future work will investigate objective verification against NWS warnings.