Other Model Simulations
National Ensemble
National WRF
Model Details and Status
Runtime Status
Ensemble Formulation
Ensemble Goals
Ensemble Examples
Model Domain
Acknowledgments
Cluster Machine Status
Newest Output Fields
Added: 10 November 2009
Multi-Panel Output Fields
Standard Deviation Fields
FSU Southeast US WRF Ensemble Simulation Case Studies
This page presents an ever-expanding list of examples highlighting the utility of a mesoscale ensemble forecast system across the southeast United States. Where possible, background from the schemes is used to provide reasoning for why the model forecasts show what they do. Verification is presented to give an assessment for the best performing member(s) of the model suite. If you happen to notice something particularly interesting, please feel free to contact me at acevans-at-met.fsu.edu and I'll look into adding it to the list.
(9/24/09): Ensemble Spread & Multi-Member Output Fields
This summer, new ensemble spread and multi-member output fields were added to the regular suite of products produced from the FSU Southeast US WRF-ARW Ensemble. Lacking until now, however, was a brief tutorial on how to interpret these plots. In this installment, we will examine the ensemble spread and multi-member output of surface-based CAPE from Tallahassee, FL from the 1800 UTC 23 September 2009 ensemble suite. In the ensemble spread image, the magnitude of the surface-based CAPE field at Tallahassee is depicted on the y-axis as a function of time on the x-axis. The ensemble mean, or simply the average surface-based CAPE value of all 16 ensemble members, is displayed in red. The maximum and minimum value of surface-based CAPE from any of the 16 ensemble members is displayed on the outside in blue. This shows you the entire range of forecast surface-based CAPE values at a particular time at Tallahassee. One standard deviation away from the ensemble mean value, both to the high side and the low side, is depicted in yellow. This is done to show you the most likely range of ensemble members; e.g. one would expect that most of the ensemble members fall within the yellow range while all fall within the blue range. The smaller (larger) the area beneath the yellow and blue curves, the less (more) variation there is amongst members. For instance, there is very little variation among members between about 08-13Z 24 September, whereas there is very large variation among members around 00Z 24 September.
In the multi-member output image, each ensemble member is depicted on the y-axis as a function of time on the x-axis. Here, the magnitude of the surface-based CAPE field is shaded within the plotting region according to the color scale shown at the bottom of the image. For instance, the very top row of the image contains the hourly forecast surface-based CAPE values for Tallahassee from the NAM-QNSE ensemble member. These values start at greater than 1500 J/kg in the first few forecast hours, then continue to fall throughout the evening and nighttime hours. The variation between the ensemble members referenced above with the ensemble spread image comes to light here. At 00Z 24 September, we see two members -- NAM-CUBL and NAM-ALL -- with surface-based CAPE values in excess of 1500 J/kg, while we also see three members -- NAMCtrl, NAM-CU, and NAM-WSM6 -- with surface-based CAPE values between 250-500 J/kg. Meanwhile, at 12Z 24 September, we see all but one member -- NAM-GD -- with 250 J/kg of surface-based CAPE or less. Thus, through just two images, we have been able to gain an understanding of the spread amongst ensemble members as well as the actual values from each ensemble member from this forecast period at Tallahassee.
Further uses of these images are possible. The aforementioned example could be extended to seeing how these differences and spreads appear within the hourly precipitation fields; for instance, one might expect a member with precipitation to have a different surface-based CAPE value than a member without precipitation. This might impact the temperature fields as well. Similarly, the aforementioned example shows a case where two similar members -- NAM-CUBL and NAM-ALL (equivalent to NAM-CUBL except for the microphysics scheme) exhibit large variation compared to the remaining ensemble members. Knowing how the ensembles are formulated numerically and physically -- what varies between members, what it impacts, and so on -- in conjunction with forecast experience (such as from the examples below) can allow you to decude what members are most likely to be correct from the entire suite. Even without this knowledge, however, these images are a powerful tool that can aid in the short-term forecast process.
(5/22/09): Heavy Precipitation Events
The recent heavy precipitation event in eastern Florida associated with a non-tropical area of low pressure brought to mind memories of Tropical Storm Fay from August 2008. As you may recall, Fay dumped over 12" of rain in many areas of the Florida panhandle and southwest Georgia with maximum observed values of 24"+ between Tallahassee, FL and Moultrie, GA. In real-time, the FSU Southeast US WRF Ensemble System had a stellar performance, effectively highlighting the areas that would see the heaviest rainfalls, the times at which those rainfalls would occur, as well as the peak rainfall values that would be observed. These forecasts were recently replicated to provide yet another example of the utility of this modeling system to real-time forecasts across the region. The forecast ensemble mean accumulated precipitation field as well as the maximum accumulated precipitation field (derived from all ensemble members) highlight excellent placing and magnitude of the observed precipitation maximum with Fay, as seen by comparison to the post-storm analysis from NWS WFO Tallahassee. Fields such as those maximum accumulated precipitation fields will soon be added to the real-time ensemble forecast output and may have utility for extreme precipitation and temperature events, particularly when used in conjunction with the raw member output as well as knowing of the ensemble formulation biases.
(12/17/08): Simulated Maximum and Minimum Temperature Issues
If you analyze output from today's 1800 UTC ensemble suite, particularly the outputted temperatures from the various members of the ensemble, you'll note something peculiar: with some of the members, maximum and minimum temperatures are significantly below reality in specific locations. This isn't limited to today's simulation, either -- it's something you'll find in a number of runs of the ensemble suite. So, you may ask, what is contributing to this evolution? There appears to be something in the combination of the YSU PBL scheme and the Noah canopy model that results in significant low biases for temperatures in various locations, oftentimes urban areas (though not always the same locations). Compare output from today's 1800 UTC simulations from the YSU PBL scheme to that from the MYJ PBL scheme and note the unrealistic bullseys over Memphis, Little Rock, and near Muscle Shoals, AL. As three of the five ensemble members use this PBL scheme, effects are seen in the ensemble mean fields as well. Attribution as to why this occurs is still to be assigned, but my guess is that something in the closure for the lowest model level temperature fields for the YSU scheme conflicts with output from the Noah canopy model under certain conditions, leading to these extreme low biases. Caution should be used when viewing the ensemble mean fields due to effects such as this; the output is not a 'black box' from which a forecast can be made but instead a tool that requires proper interrogation (as this and the below examples highlight) for best forecast use.
(7/30/08): Convective Adjustment Processes
In today's ensemble suite, the first run to 36 hours, significant differences in 18 hr forecasts of 700 hPa temperature (among other fields) were noted between the two convective schemes employed within the ensemble suite, the Kain-Fritsch and Betts-Miller-Janjic schemes. As noted on the goals page, the KF scheme uses a mass flux and downdraft adjustment scheme when representing convection, taking the observed thermodynamic profile and using mass flux characteristics from there to compensate for the presence of convection. The BMJ scheme uses a set of observed or easily definable profiles to perform convective adjustment. The differences between how and when these convective schemes turn on can be extreme, as highlighted by the 18 hr ensemble mean and individual member accumulated precipitation fields. The ensemble mean field shows widespread convection and is an amalgam of the spotty, deep convection shown by the KF+YSU control run and the the BMY+MYJ ensemble member. (Note that there is little sensitivity to boundary layer parameterization noted.) How do these convective adjustment issues manifest themselves in the real atmosphere? The 700 hPa temperature fields tell the tale: note that across the regions where convection had occurred or was occurring, 700 hPa temperatures were approximately 2 C warmer with the KF run than with the BMJ run. Despite this mid-level cooling, surface-based CAPE values are significantly lower in the BMJ runs than in the KF runs, likely due to further convective adjustment processes in the temperature and moisture fields. The verifying 700 hPa temperature field ended up closely resembling the ensemble mean, with a relatively uniform temperature distribution across the region. In all, significant impacts to the short- and medium-range forecasts are realized, highlighted upon an interrogation of the entire suite of model products and help further stress the importance of not just ensemble mean forecasts but also knowing the strengths and weaknesses of each ensemble member.
(7/28/08): Boundary Layer Temperatures
The selected boundary layer schemes, Yonsei University and Mellor-Yamada-Janjic, differ in how they close the turbulence parameterizations within the PBL. The Yonsei Univ. scheme uses non-local eddy mixing triggered by surface heating, while the Mellor-Yamada-Janjic scheme uses local mixing. Each has its advantages and disadvantages, leading to the Mellor-Yamada-Janjic scheme often resulting in too shallow of a boundary layer and too moist in the PBL -- e.g. most of its negative influences lie in the PBL itself -- and the Yonsei Univ. scheme often having too shallow of a convective inhibition layer and thus being too moist above the PBL -- e.g. most of its negative influences lie at the top of the PBL. The impacts that these differing mixing schemes can have are highlighted by two selected meteograms, one using the Yonsei Univ. scheme and the other using the Mellor-Yamada-Janjic scheme. Significantly improved boundary layer mixing is noted in the Yonsei Univ. run for the selected location, resulting in a slightly drier/warmer boundary layer and a maximum surface temperature 4 F higher than with the Mellor-Yamada-Janjic scheme. The verifying high temperature for that day was 91 F, resulting in 3 F and 7 F errors for the Yonsei Univ. and Mellor-Yamada-Janjic scheme respectively. Such differences impact available atmospheric instability, potentially leading to mesoscale differences in convective evolution independent of the more significant changes noted between convective schemes.
© 2006-2009, Clark Evans. Disclaimer: These forecasts are experimental and NOT official forecasts. As with any model, these forecasts are prone to large forecast error. Please refer to official NWS forecasts for the latest weather information.
