Introduction

From 13 May to 25 June 2002, FSL scientists were involved in the International H2O Project (IHOP), an extensive field study covering the Southern Plains and based in Oklahoma. More than 100 people participated in the campaign to determine primarily how to improve the characterization of the four-dimensional distribution of water vapor and its application to better understand and predict convection. The four main components of the program included quantitative precipitation forecasts (QPFs), convective initiation (CI), atmospheric boundary layer processes, and instrumentation. FSL ran experimental versions of local and national scale models during IHOP to assist with nowcasting and short-range forecasting and to determine whether such models could provide useful forecast and nowcast guidance for convective weather.

FSL's Rapid Update Cycle (RUC) model, a national scale model, and the Local Analysis and Prediction System (LAPS), a smaller scale model, were used during IHOP. LAPS was designed to run onsite at a National Weather Service Forecast Office (WFO), using local analyses that take advantage of locally available data. LAPS is currently running in AWIPS at WFOs on an hourly cycle with a 10-km grid spacing. LAPS was run at a 12- and 4-km horizontal grid resolution and used to initialize some of the models that FSL ran during IHOP.

One goal of a short-range model is to provide better prediction of precipitation without a spin-up period. To aid in this goal a "Hot Start" scheme was developed using the LAPS cloud analysis to prescribe a vertical velocity profile where sufficiently deep clouds are present at initialization time. The three-dimensional dynamical relationship between mass and momentum is adjusted by the LAPS balance algorithm to force consistency with the diagnosed cloud vertical motions and allow for a smooth model start. During IHOP, a 12-km horizontal resolution MM5 hotstart initialized with LAPS was run, with a nested 4-km version covering the IHOP experimental domain (Figure 1). LAPS also was used to initialize a similar 12-km setup for the Weather Research and Forecast (WRF) model. In addition to these models initialized with LAPS, FSL ran a 10-km version of the RUC model. The RUC model employs a three-dimensional variational (3DVAR) analysis for the mass fields, and initial RUC hydrometeor fields are adjusted to correspond to base scan reflectivity patterns at the initial time, but without any modification of the initial vertical velocity field (in contrast to the hotstart method). The models run by FSL for IHOP are summarized in Table 1. (These experimental model runs are archived by UCAR at http://www.joss.ucar.edu/ihop.)

Evaluation of Model Performance

One of our goals during IHOP was to compile a fairly extensive subjective evaluation of the various models in real time. Objective model evaluation was done within FSL (see Brent Shaw's article WRF). In order to perform subjective evaluation, an online evaluation form was designed that allowed the forecaster/nowcaster to document what the model was forecasting, the relationship of various forcing features to the subsequent convection forecast by the model, and the forecaster's confidence in the forecast. The forecasters also made many freeform comments that give insight into how the models performed with the various short-range forecast problems during IHOP. A summary of these comments follows:

• The MM5 models using LAPS for the initial state did an excellent job of initializing ongoing convection, but often this convection was lost in the first hour of simulation. Adjustments were made to the Hot Start scheme for a set of post-IHOP reruns of both MM5 and the WRF model, and our preliminary evaluation of some of these reruns indicates some improvement with this problem. The most easily "lost" convection involved elevated storms (nonsurface-based convection), while very strong individual storms and lines were much better retained from the initialization.
• Outflows tended to be easily produced from convection in the MM5 model, especially so in the 4-km run, whereas the RUC tended to be more conservative in producing outflows but was able to do so.
• The most difficult storms to forecast were elevated convection, which usually formed in the very early morning (presunrise) hours and could persist for up to 6 hours after sunrise. Coincidentally, this type of convection is also among the most challenging for forecasters, as it can occur without any obvious surface forcing feature present. Though seldom producing severe storms (at least during IHOP), elevated convective events were often of the "surprise" category. There were often indications in the model of possible activity, for example in the form of midlevel echo but without precipitation reaching the surface, so an underforecasting of the convection. Convection associated with a warm front (on the cool side of the warm front) also tended to be an area where the models were deficient. This type of convection often was not surface-based, sharing that characteristic with the elevated storms noted above.
• Some of the forecasts of convective initiation along drylines were quite good. For a few cases the model beat the forecasters, particularly when temperatures both at the surface and aloft were quite hot. In these cases, forecasters overestimated the time it would take to break the cap, while the model more correctly forecast convective initiation earlier.
• Other good forecasts occurred with well-defined surface-based forcing features, such as cold fronts.
• As noted earlier, there was some skill in the model's ability to forecast storm type and evolution, with several events during IHOP that featured upscale growth into organized lines that often accelerated much faster than indicated in the precipitation fields from conventional models (for example, the Eta model).

Selected Case Studies From IHOP

Two cases are examined using both model runs during IHOP and the reruns that occurred after IHOP. The reruns provided a series of model runs from the WRF model, which was actually run during IHOP but not able to display in real time. As a result of the real-time experiences during IHOP, we decided to apply improvements to the hotstart method to the reruns. A significant improvement involved removal of a warm bias that existed in the LAPS initialization, and was likely responsible for overprediction of convective precipitation during IHOP from the MM5.

2 June 2002: Dryline Case – On this day the western half of the IHOP domain was dominated by very hot temperatures, reaching the low 100s (oF) during the afternoon. A well-defined dryline was not seen initially, but was sharpened in the early afternoon, as shown by the LAPS analysis of wind and dewpoint along with low-level reflectivity in Figure 2. This sharpening first appeared as a surge of westerly surface flow that emerged out of eastern Colorado that then pushed into Kansas. On this day, the IHOP forecasters predicted that a dryline would become better defined during the afternoon (somewhat later than what occurred) but they thought convective initiation along it would occur fairly late in the afternoon as they waited for the dryline to sharpen and temperatures to break the significant cap that was in place. As it turned out, the presence of the very hot surface temperatures and a somewhat stronger and earlier dryline push than expected allowed the cap to be broken and convective initiation to occur over 2 hours ahead of the IHOP forecasters' prediction. It was noted in IHOP that the MM5 model did a good job indicating this convection earlier than expected, particularly the run initialized at 1500 UTC. Some of the runs from that day are examined and the somewhat different forecasts are contrasted for this relatively "tricky" case. With the expense of some of the resources in IHOP, in particular the aircraft, timing of convective initiation was a critical forecast issue. In this case, even a forecast error of 2 hours for convective initiation from a midmorning forecast was critical and resulted in an aborted mission, since convection was well underway before the aircraft (leaving Norman, Oklahoma) could reach the dryline target.

In summary, results are mixed for this aspect of 15 June; on one hand the 0600 UTC runs indicate some fairly impressive predictability, but inability to repeat this predictability for the IHOP runs closer to the event is curious. We hope to compare MM5 and WRF reruns from 1200 UTC to the 0600 UTC reruns to see if the storm was still forecast for these later model runs.

The final portion of the 15 June 2002 case that is examined involves the organization of three areas of convection into a squall line by 0000 UTC on 16 June that then accelerates southward out of the IHOP domain by 0600 UTC. A radar overview of this evolution is shown in Figure 14. At 1800 UTC the organized storm discussed earlier is just crossing into western Kansas, and at 2100 UTC is at the western end of the line segment located in southwestern Kansas. By 0000 UTC a line extends from northern Oklahoma west-southwest into the Texas Panhandle, with the eastern portion developed from the area of cells that moved south out of Nebraska. After 0000 UTC, the line organizes and accelerates as it bows over western to south-central Oklahoma. The model forecasts from the 0600 UTC reruns of the MM5 12-km are shown in Figure 15 and the WRF 12-km forecasts are shown in Figure 16. The MM5 organizes a group of cells in Kansas at 2100 UTC into a line segment close to where it is actually found at 0000 UTC, then accelerates the line southward. Although the actual line moves faster than the forecast, the track is similar and the model forecast includes a bowing line as observed. Considering that the later period of this forecast is an 18–24 hour forecast, it is fairly impressive, with the model doing a very good job of forecasting the organization into an accelerating, bowing line. This line forms in about the right place even though the MM5 essentially misses all of the storms that around 0900 UTC began to form in a north-northwest to south-southeast line from central Kansas to west-central Nebraska. These storms continued to expand in about the same place, and appear to have been, at least initially, somewhat elevated type storms that developed just ahead of a warm frontal boundary. The earlier times of this MM5 forecast never included anything but some mid-level reflectivity, and even then it was west of where the line actually occurred. The difficulty in handling convection that may not have been surface-based or forced by a distinct low-level boundary is similar to the problems that all of the models had with the elevated convection in the Texas Panhandle discussed earlier.

The MM5 runs during IHOP extended out only to 12 hours, compared to 24 hours for the reruns, so for comparison, runs beginning at 1500 UTC for the 12-km MM5 and at 1800 UTC for the 4-km MM5 will be shown. The 1500 UTC MM5 12-km IHOP run is shown in Figure 17. Note how the initialization from LAPS nicely captures the ongoing convection at 1500 UTC (Figure 17a), although the storms are quickly lost, mostly in the first hour. This occurred at times with MM5 during IHOP, as noted earlier, and for this case may have been exaggerated somewhat because much of this convection near an apparent warm front may not have been surface-based. Unfortunately there is such a loss of echo that by the 3-hour forecast (valid at 1800 UTC), a composite echo is forecast but none is predicted to reach the surface. Right after 1800 UTC, however, the mid-level echo shown entering northwest Kansas in the 1800 UTC forecast strengthens rapidly, then expands to form the line segment shown in the forecast valid at 2100 UTC. This line segment then moves southward and strengthens and expands, bowing somewhat at 0000 UTC but then becoming more of a straight line by 0200 UTC. The line in this forecast does not accelerate as fast as in the 0600 UTC forecasts from the WRF and MM5 reruns shown earlier.

a)

It is apparent that all the models were able to predict the upscale growth and organization of the convection into a line with good location and timing for the most part. There was good agreement between the different models and usually between the different initialization times. A consensus forecast from an ensemble viewpoint of the various runs would have been a good one. The dprog/dt method did not necessarily verify as well, however, especially for the MM5 4-km runs, with the 2100 UTC run not as good a forecast as earlier runs.

Summary and Future Work

The model forecasts examined here from two IHOP days encompass a variety of forecast problems typically encountered especially east of the Rocky Mountains, including convective initiation along a dryline, prediction of supercells, upscale growth and organization of storms into a squall line, and the very tricky forecast of overnight elevated convection. The forecasts presented well represent the behavior of the models used during the IHOP period, and indicate that there is potential for such models to offer forecast guidance that can be valuable to forecasters trying to predict convection. The model was most successful when the convective initiation was forced by a well-defined surface boundary, as in the 2 June 2002 case, and had the most difficult time with storms forced by more subtle boundaries (like the warm front on 15 June) or by no apparent surface boundary, like the elevated storms in the Texas Panhandle on 15 June. Some of the forecasts of supercell formation and movement, as well as upscale growth that occurred on 15 June were impressive, and there was even skill shown for such developments beyond the typical 6–12-hour limit that one might suspect for convective forecasts.

During IHOP, the RUC and MM5 special model runs by FSL were extensively used to help make short-range forecasts, with the models displayed on the FSL FX-Net workstation. Partial examination of an extensive real-time questionnaire completed by the forecasters for as many model runs as possible during IHOP has yielded constructive insight into various model issues that occurred, as well as how much the forecasters trusted some of the predictions.

Often these predictions carried far more detail as well as forecast precipitation (convection) than would be indicated by the operational models (Eta or GFS), and in some cases, some forecasters needed a spin-up time to understand whether the forecasts could be believed and how best to use them. For this study, we will continue completion of the analysis and questionnaires. We also want to examine model performance over a broader spectrum of IHOP days, not in the detail as was done for 15 June but more by phenomenon, such as the different convective types discussed for this same case.

Editor's Note: A complete list of references and more information on this and related topics are available at the main FSL Website www.fsl.noaa.gov, by clicking on "Publications" and "Research Articles."

(Edward Szoke is a meteorologist in the Forecast Research Division headed by Dr. John McGinley. He can be reached at Edward.J. Szoke@noaa.gov, or at 303-497-7395.)