By Brent L. Shaw, Steven C. Albers, John A. McGinley, and Linda S. Wharton
IntroductionWeather plays a critical role in the planning and execution of spacecraft launch operations at both the Eastern Range, located at Cape Canaveral Air Force Station, Florida, and at the Western Range, located at Vandenberg Air Force Base, California. Each launch vehicle has specific tolerances for wind shear. Cloud cover, temperature, and lightning constraints are common for all launch vehicles. The presence of convective storm activity in the vicinity of the ranges can be significant due to the potential for lightning strikes and electrostatic discharge, which can damage or destroy the launch vehicle and/or its payload. In the event of a launch mishap, toxic clouds and falling debris have the potential to endanger ground personnel, thereby requiring high-resolution wind and stability information for input into atmospheric dispersion models used for risk mitigation prior to launch and as an emergency response capability. For example, Figure 1 shows high-resolution model output that compares an MM5 6-hour forecast from a 1.1-km domain over the Western Range of wind and temperature fields to a corresponding operational Eta forecast of surface ob-servations. See the "Results" section for more information.
Figure 1. MM5 6-hour forecast from 1.1-km domain over the Western Range, valid 1200 UTC on 11 January 2002. Image is MM5 1.1 km terrain heights. Surface meteorological aviation reports (METARs) are overlaid in cyan, and the Eta surface wind field is overlaid with green vectors (highlighted with the annotated white arrows).
RSA ProgramSeveral years ago the Air Force initiated the Range Standardization and Automation (RSA) program to modernize and standardize the command and control infrastructure of both ranges. This upgrade includes improvement of the systems used by the operational weather squadrons for monitoring and forecasting weather conditions in support of launch and emergency response activities.
FSL has been involved in the weather analysis and forecasting component of the RSA effort since 1996, when observation simulation and sensitivity experiments (OSSEs) were performed to determine optimum placement of various newly acquired observing systems.
Additionally, FSL was funded to develop a data assimilation and forecast system using the Local Analysis and Prediction System (LAPS). In the initial prototype, the value of an explicit mesoscale numerical weather prediction model coupled with LAPS was demonstrated using the Colorado State University Regional Atmospheric Modeling System (RAMS). The planned configuration consisted of a coarse outer grid with 16-km grid spacing and a fine nest within the coarse grid utilizing 4-km grid spacing. To meet the operational timelines, which require that a 24-hour forecast be produced within 6 hours of actual time, the domains were very limited in size, using 40 x 40 points horizontally. Since the initial design decision, the cost of computational power has dramatically decreased, and it is now feasible, within the budget of the program, to use a larger domain and add a third nest to approach the objective requirement of 1-km grid spacing.
To leverage other FSL research and development performed for the National Weather Service (NWS), Lockheed Martin Mission Systems (LMMS, the primary RSA contractor) selected the Advanced Weather Interactive Processing System (AWIPS) as the human-machine interface for interacting with all of the standard and range-unique meteorological data. To maximize performance and minimize cost, the new Linux version of AWIPS was selected and installed, allowing the use of relatively inexpensive commodity computers.
This article discusses the data assimilation and forecasting system, including recent improvements to the LAPS analysis, the selection and configuration of the forecast model, the revised domain configuration, and the integration of the data assimilation and forecast system within AWIPS. Also covered are issues related to running forecast models within the AWIPS environment and future plans for the RSA and similar assimilation systems.
LAPS OverviewLAPS Analysis LAPS has a rich history of operational application. It is available on all operational NWS AWIPS systems within the continental U.S., and has been used by other national centers, both foreign (e.g., the Taiwan Central Weather Bureau) and domestic (e.g., the U.S. Air Force Weather Agency). Initially, LAPS was designed to provide a computationally efficient method of combining all available sources of meteorological information into a single, coherent, three-dimensional depiction of the atmospheric state, with an emphasis on nowcasting. In recent years, LAPS has been increasingly used as a means to initialize numerical weather prediction models because of its robust data ingest, quality control, and fusion capabilities.
LAPS "Hot Start" Technique LAPS was recently modified to allow diabatic initialization of numerical weather prediction models, referred to as the "hot start" technique. This initialization method virtually eliminates the problem of model "spinup" by including cloud and precipitation fields in the initial condition in conjunction with a dynamic balance constraint. The specific enhancements that make this technique possible include an improved cloud analysis and a dynamic balance package. The moisture analysis has also been improved to take better advantage of the satellite, GPS, and ground-based moisture observations through the use of a variational technique.
Cloud Analysis Scheme The cloud analysis utilizes a wide variety of observational data, including GOES satellite imagery, aircraft observations, surface observations, and WSR-88D reflectivity data. It diagnoses the mixing ratios of each hydrometeor species, performs cloud typing based on stability and temperature information, and assigns appropriate vertical motion profiles to each cloud based on a steady state assumption.
We recently implemented improvements to the cloud analysis scheme. First, the RSA version integrated with the Linux AWIPS system can now ingest and use the narrowband radar reflectivity from multiple WSR-88D sites, available via the Satellite Broadcast Network (SBN) feed to AWIPS. Combined with the locally available wideband WSR-88D data, this allows for better coverage over the entire horizontal domain of the grid being used for RSA, which is spatially larger than a typical NWS Weather Forecast Office domain.
Second, the cloud analysis has been modified to use a climatological albedo field to increase the utility of the visible satellite imagery. This, combined with the recent integration of the GOES 3.9 micron channel, has greatly improved the system’s ability to detect low stratus clouds over the ocean, a problem faced particularly at the Western Range.
Third, for the purposes of initializing a numerical weather prediction model, the final three-dimensional concentrations of the hydrometeor species are scaled as a function of horizontal grid spacing to be consistent with typical values expected to be produced by a mesoscale numerical weather prediction model at the given grid spacing. In addition to the scaling factor, the values of cloud liquid and ice are limited to an amount less than or equal to the thresholds used within the model's microphysics scheme to trigger conversion of liquid to rain and ice to snow, respectively.
Finally, to ensure compatibility with current numerical weather prediction microphysics packages, any grid box volume containing cloud liquid or ice is raised to its saturation level with respect to the phase of the cloud species. This prevents rapid evaporation of the cloud, which often results in artificial cooling and spurious downdrafts within the first few time steps of model integration.
Dynamic Balance Package The other key component of the LAPS diabatic initialization is the dynamic balance package. The balance procedure uses a three-dimensional variational (3DVAR) formulation to combine the background "first-guess" fields with the initial LAPS univariate analyses of the mass and momentum fields while considering the cloud-derived vertical motions. A strong constraint of mass continuity is applied as part of the 3DVAR cost function, such that the horizontal wind field, is slightly adjusted to produce divergence fields consistent with cloud vertical motion fields. In addition, to prevent the introduction of excessive gravity waves during the initial few time steps of model integration, the cost function also minimizes the time tendencies of the horizontal wind fields.
The LAPS diabatic initialization technique has been tested in real time and undergone continuous improvement at FSL since late 2000. Statistical verification during the winter of 2000–2001 showed significant improvement in the 0–6 hour forecast compared to other more traditional initialization methods. For example, the 1-hour forecast of cloud cover exceeding 50 percent scored a 0.71 equitable threat score (ETS) using this method. In comparison, runs performed using a 3-hour preforecast analysis "nudging" period (where the model is nudged toward subsequent LAPS-analyzed state variables for a 3-hour period prior to the desired initial time) only scored a 0.56. Simulations for the same period initialized with only the 6-hour forecast of state variables from the NCEP (National Centers for Environmental Prediction) operational Eta model produced a much lower score of 0.29.
Field Evaluation The International H2O Project (IHOP) field experiment in the summer of 2002 provided another opportunity to evaluate the LAPS hot start in a quasi-operational setting and directly compare the resulting forecasts to operational numerical weather prediction guidance. Figure 2 shows the equitable skill score for 119 3-hour quantitative precipitation forecasts from three different models run over the IHOP verification area. The 12-km MM5 domain was run at FSL using the LAPS diabatic initialization, whereas the 12-km Eta grids were the operational forecasts from NCEP. These verification statistics were generated by FSL’s Real Time Verification System (RTVS) and show that forecasts produced using the LAPS diabatic initialization had significantly better skill and bias in forecasting precipitation in the 0 3 hour forecast period than either of the two operational models, particularly for the higher threshold amounts. While not as dramatic, the 12-km MM5 forecasts had better skill for most of the thresholds for the 0 6 hour and 0 2 hour quantitative precipitation forecasts as well.
Figure 2. The Real-Time Verification System equitable skill score (top) and frequency bias (bottom) for 3-hour quantitative precipitation forecasts over the IHOP domain for 119 LAPS 12-km MM5 (blue bars and line) and NCEP operational 12-km Eta (red bars and line) cases. A perfect equitable skill score is 1.0, and a perfect frequency bias is 1.0.
In addition to the objective verification results, real-time runs of MM5 using the LAPS diabatic initialization have been provided to the NWS Weather Forecast Offices located in Boulder and Pueblo, Colorado, for operational evaluation. Feedback from the forecasters has been very positive. Based on this experience and the objective verification, the LAPS diabatic initialization technique has also been implemented for the RSA projects.
LAPS Enhancements The LAPS data ingest capability is also being significantly enhanced to support the RSA project. In addition to the large suite of meteorological data LAPS already utilizes, the capability to ingest and use the local range-unique data sources is being developed and tested at FSL. These data include 915-MHz and 50-MHz wind profilers, mini-SODAR wind profiles, tower observations, and temperature profiles from the Radio Acoustic Sounding System and the Automated Meteorological Profiling System. The addition of these high-resolution (temporal and spatial) data sources should improve the atmospheric analyses and subsequent forecasts of boundary layer winds and stability, both of which are critical for the atmospheric dispersion modeling applications used by the range safety offices.
Other software improvements to LAPS directly related to the RSA project include the capability to specify a non-uniform distribution of vertical pressure levels on which to perform the analysis, allowing for higher resolution in and near the boundary layer or other layers of significant interest. Further, a new LAPS postprocessor that supports MM5, RAMS, and the new Weather Research and Forecast (WRF) model has been developed to provide additional flexibility in the choice of forecast models coupled to LAPS. This postprocessor could be easily extended to support other models (e.g., the Advanced Regional Prediction System, workstation Eta) as the need arises.
NWP Forecast Component In principle, LAPS can be coupled with any mesoscale numerical weather prediction model. The NCAR/PSU fifth-generation mesoscale model has been selected for use as the forecast component within the RSA project. FSL’s close working relationship with NCAR and extensive experience with the MM5 model, combined with its public domain nature, make it a logical choice. Also, the selection of MM5 as the forecast model places the RSA program on a direct track for future scientific improvements in the area of mesoscale numerical weather prediction, including an upgrade to the WRF model in the future.
MM5 version 3, release 5 (MM5v3–5), is the version that has been implemented for RSA. It is a nonhydrostatic model using a terrain-following pressure coordinate and offers a wide variety of physics options. Minor modifications have been made to accommodate the LAPS diabatic initialization technique. Additionally, the model, its preprocessing packages, and an FSL-developed postprocessor have been placed into a directory structure more suitable for operational applications, and various Perl scripts to configure and run the forecasts in real-time have been developed.
The postprocessor and its associated scripts support a wide variety of output formats, including AWIPS netCDF, LAPS netCDF, Vis5D, GRIB, and tabular text point forecasts. The postprocessor is able to take advantage of the buffered output option supported in MM5v3, so that forecasters at the ranges can view the model output as it is produced rather than waiting for the entire run to complete. Like the LAPS analysis software, this custom version of MM5 is maintained at FSL using revision control tools.
RSA Domain ConfigurationThe domain configuration for both ranges (Figure 3) utilizes a triple-nested domain, with each of the three nests centered upon the range location. The outer grid has a horizontal grid spacing of 10 km, and each subsequent nest has a grid spacing three times smaller than its parent. Thus, the innermost grid has a grid spacing of 1.1 km, and each grid has 97 x 97 horizontal grid points.
Figure 3. Analysis and forecast domain configuration for Western Range (left) and Eastern Range (right).
Vertically, the LAPS grid has been configured to use 41 pressure levels, with more levels concentrated in the lower portion of the domain. The MM5 model uses 41 terrain-following sigma levels, with the highest resolution also located within the boundary layer. The number of vertical levels selected for the forecast model is a trade-off between resolution and operational efficiency. Doubling the number of vertical levels has the net effect of quadrupling the number of computations required to complete a forecast. The factor of four increase is because twice as many time steps are needed, and they have to be computed on twice as many grid points.
The physics options selected for the MM5 forecast model represent a blend of performance and applicability to the project. Since the diabatic initialization provides five classes of hydrometeors (cloud water, cloud ice, rain, snow, and graupel), the microphysics scheme developed by Paul Schultz of FSL is employed for its computational efficiency and ability to deal with all five species. The dispersion models to which the output will be coupled at the ranges require forecasts of turbulent kinetic energy, so the Burk-Thompson formulation has been chosen for its efficiency and applicability to the high-resolution boundary layer. Although it can be argued that the 10-km grid requires a parameterization to handle convection, no parameterization is employed. This decision is based on FSL’s experience and forecaster feedback with the hot start technique over the last three years.
The LAPS analyses are produced on all three grids each hour. The first-guess field currently comes from a 1-hour MM5 "update" forecast on the 10-km grid statically initialized using NCEP Eta forecast fields for the initial and lateral boundary conditions. The system can optionally be run in a full four-dimensional data assimilation (4DDA) cycle; however, preliminary results show decreased forecast quality. Possible reasons for this are discussed later.
A new hot-started MM5 forecast is created every six hours, with initial times at 0300, 0900, 1500, and 2100 UTC. The model runs using two-way feedback to provide consistency between the three grids, as it allows features generated on the fine-scale grids to be represented in their parent grids. The current hardware utilized is not sufficient to run a full 24-hour forecast (as required) for all three nests. Given the current state of mesoscale numerical weather prediction, this may not be scientifically justified anyway. Thus, the system is configured to run all three domains for the first 9 hours of simulation time. At that point, the 1.1-km grid is deactivated and only the 10 and 3.3-km are allowed to run from the 9 12 hour forecast period, after which the 3.3-km grid is deactivated and the 10-km runs the rest of the forecast out to 24 hours. This configuration provides a good balance of short-range, high resolution forecast products for nowcasting along with weather planning for the next day’s activities.
Hardware ConfigurationLAPS and MM5 share a Linux cluster modeling server, which is fully integrated within the local AWIPS network at the ranges. The RSA program is the first known operational application of affordable Linux clustering technology in the United States for NWP. The cluster consists of nine, dual-processor nodes containing Intel Pentium-III processors running with 1 GHz clock speed. One node serves as a front-end node, and this is where the LAPS analyses, the 1-hour MM5 update forecasts, and all preprocessing and reformatting for AWIPS occurs. The remaining 8 nodes are used as "compute" nodes for executing the parallelized version of MM5v3–5 along with simultaneously running the model postprocessor.
Three of these clusters were procured by LMMS during the winter of 2001 2002. One was installed at each range, and the developmental system was installed at FSL. A photograph of the cluster in its equipment rack at FSL is shown in Figure 4. This replica of the operational configuration at FSL has greatly improved the ability to transmit improvements in the data assimilation system to the ranges, because developmental code can be rigorously tested in an environment identical to the operational configuration. Furthermore, when problems occur at the ranges, FSL is able to respond much more rapidly with an identical system for troubleshooting problems.
Figure 4. Photograph of the Linux cluster used for development at FSL.
The clusters are integrated with AWIPS in a modular fashion. The only interaction with AWIPS is through a sharing of the primary AWIPS data tree located on the AWIPS data server. The only load presented to the AWIPS system by the modeling server is the additional network traffic between the cluster’s master node and the AWIPS data server. This configuration makes it easy to add a modeling server to any AWIPS system.
Model Results at the RangesThe RSA LAPS-MM5 system has been running in real-time at FSL since late 2001. The system was installed first at the Western Range in December 2001 and at the Eastern Range in September 2002. During this period, many of the improvements discussed earlier were made and delivered incrementally to the LMMS for installation at the ranges. The modular design of the system and the ability to perform rigorous testing in Boulder allows for relatively easy updates for LMMS. New versions are delivered on CD-ROM complete with an installation script that simply requires inserting the CD in the computer and running one script. Typically, this takes less than five minutes and no further actions are required.
The final RSA LAPS-MM5 configuration is (at this time) more than a year from being completed. However, both ranges have the benefit of receiving incremental deliveries of the system as new capabilities and improvements are implemented. The current system makes use of more sources of meteorological data than any other known operational system. Thus far, results from the basic capability running at the ranges have been proving useful since they are able to resolve features not possible in the operational models as depicted on AWIPS. Additionally, the temporal resolution of the forecast model output (hourly) and the ability to provide many more 3D fields to the forecaster (e.g., the hydrometeor fields) provide a capability not normally available to a typical Weather Forecast Office environment.
A comparison of an MM5 1.1-km forecast wind and temperature field to surface observations and the corresponding operational Eta forecast was shown in Figure 1. The MM5 forecast is able to produce realistic flows around the terrain features (image) that cannot be resolved by the 40-km Eta data provided on AWIPS. This resolution, along with the AWIPS capability to generate cross sections and time-height series of this high-resolution data, is critical to range launch and safety operations.
Current Work and ConclusionsFSL's support for the RSA program is ongoing. New required capabilities are being added to the LAPS-MM5 system, weaknesses of the existing system are being addressed, and planning for the future has begun. The range-unique datasets have recently been made routinely available to the FSL developmental system, and FSL staff are developing and testing the ingest and proper usage of these data.
As mentioned previously, preliminary testing showed no added benefit to using a rapidly updating 4DDA cycle as originally planned. In fact, running a 4DDA cycle generally hurts forecast quality. Two potential primary reasons are being investigated to address this: (1) the effects of improperly forecast convective storms in the first-guess field from a diabatically initialized numerical weather prediction forecast, and (2) improper specification of the 3D errors associated with the first-guess forecast.
Despite the increased horizontal resolution and elimination of the spinup problem, convective storms continue to have low predictability in terms of exact location and evolution. When the short-range forecast improperly forecasts a convective cell in an incorrect position (or one that does not even exist), there is currently no easy way to remove the cell and its effects (downdrafts, cold pools, etc.) from the first guess. To address this, the dynamic balance package is being modified to allow a true thermodynamic constraint to allow removal of anomalous storms.
While the 3DVAR balance package accommodates 3D error terms for both the observations (univariate analyses in this case) and the first guess, these have been estimated rather than explicitly computed. Furthermore, the same estimates of model error are used regardless of which model provided the first-guess field. Thus, when using MM5 as the first guess, it is assumed to have errors equivalent to the national Eta or Rapid Update Cycle (RUC) grid. Since MM5 is run at a higher resolution, resolves mesoscale features, and uses a very short-range (1-hour) forecast, it should likely be given more weight in the balance package than is allowed by the current estimates of error, based on national models. Also, since the LAPS balance package already accounts for a large portion of the necessary balance that 4DDA systems are designed to produce, we hypothesize that proper specification of the background error terms are increasingly important if we are to realize the advantages of 4DDA. To address this need, LMMS has funded the project to develop an online verification system that will not only provide feedback to the forecasters, but will allow better specification of the error arrays.
During 2003, FSL will continue improvements to the LAPS-RSA system already underway. Proposed tasks include testing with and potential operational migration to the new Weather Research and Forecasting (WRF) model, which is showing promise in other FSL programs. The success of the program to date has opened the door to emerging markets for a capability similar to that designed for the RSA program. For example, two NWS initiatives involve setting up a similar system. The Coastal Storms Initiative involves setting up a real-time, diabatically initialized WRF model run at the Jacksonville, Florida, Weather Forecast Office. A fire weather forecasting initiative involves a similar setup at the NWS Western Region headquarters.
This project marks the first known operational (in the strictist interpretation) implementation of locally run data assimilation and forecasting system on relatively inexpensive Linux clusters within the U.S. While done on behalf of the Air Force, the synergistic nature of the program implies that benefits from this project can be reaped within other components of the national meteorological infrastructure. The use of a local, diabatically initialized mesoscale numerical weather prediction model does not replace the necessity for high quality national products. Rather, it has the potential to complement those products by filling the niche between nowcasting techniques, which work well for the 0 2 hour period, and the national numerical weather prediction guidance, which performs well for the "day 1" and beyond forecast. Moreover, a locally run system takes advantage of high spatial and temporal resolution observational data not available at the national centers. It also provides users with flexibility to tailor the operational products and run environment to their specific needs. This project may well be a "pathfinder" for the local forecast office of the future.
Note: A list of references on this and related topics are available at the FSL Website (http://www.fsl.noaa.gov) by clicking on "Publications" and then "Research Articles."
(Brent Shaw is a researcher in FSL’s Local Analysis and Prediction Branch, headed by Dr. John McGinley. Mr. Shaw is also affiliated with the Cooperative Institute for Research in the Atmosphere (CIRA), Colorado State University, Fort Collins, CO. He can be reached at Brent.Shaw@noaa.gov or by phone, 303-497-6100.)