Model Potpourri

 Spent time discussing the forecast issues of the day (and it lasted all day) for 5/8:
1. lack of synoptic forcing on the strong side
2. boundary layer (dry air HCRs initiating storms on the dryline, warm sector HCRs initiating storms east of the dryline in a narrow moist tongue)
3. marginal to sufficient shear for supercells
4. boundary layer oddities (mesoscale drying in the post MCV environment that advected north in OK, moist plume that got sandwiched between the dryline and the dry air pocket)

Models: signal or noise?

Todays forecast challenge had to do with storm location and initiation. From the TX panhandle northward thru KS into NE and CO. It seemed fairly evident that storms would form in this deep and relatively dry boundary layer (at least for May). With the PBL depth approaching 3km, models were not shy in breaking out storms to enjoy the roughly 40-50 kts of vertical shear and lovely quarter circle hodographs. It was relatively obvious that the PBL would control where and when storms broke out.

2013 Spring Experiment Week 1 Day 1

The Spring Experiment within the Hazardous Weather Testbed kicked off today. The Experimental Forecast Program has a running list of objectives (there just is too much to explore) so lets hit the high points of producing convective outlooks:
1.  Can we merge human and ensemble forecasts to produce shorter time scale forecasts (in this case create a long period (20 hour) forecast and turn that into multiple 3 hour periods)?

What we are testing involves making the best use of the models and incorporating what the forecasters are good at. The models are terrible in getting severe weather at the right times and in the right places. Challenges in predictability (i.e. small scale errors in observations, errors in how we represent model processes, etc) add up to prevent being truly spot on in the forecasts. But forecasters, after years of experience and training, can be great at pattern recognition and referencing climatology about severe storms and severe storm processes. Merging of the two can thus help make better forecasts as we demonstrated last year. We call this technique Temporal Disaggregation (TD). It just means we take the ensemble forecast and apply it to the area outlined by forecasters. It worked well and compared favorably to what forecasters could draw independently ... in other words: what we drew in fine grained forecasts matched well what the model probabilities would be if we corrected the location!

As far as time, we learned that if the periods were long enough we could account for poor time forecasts. So the next big question is:

2. Can we make good short period forecasts? What resources can we take advantage of to get around the predictability challenge?

So our TD technique will produce a first guess of the forecasts for the 3 shorter periods (18-21, 21-00, 00-03 UTC).  From these first guesses, we will use updated ensemble data to update these forecasts. Here we have more models to use with the latest data (through data assimilation). One of the challenges an operational forecaster would face is: Can this be done timely and accurately while providing good risk information? Can the experimental models be used reliably?  Can we create good proxies or variables that we can extract from the model that relate either directly or indirectly to severe weather?

We have the CAPS 8 member 4-km grid spacing, radar data assimilation ensemble (SSEF), the 7 member SSEO, and 10 member AFWA. We also have the NSSL mesoscale ensemble (NME) run at 18km grid spacing with 36 members but updated hourly with 3 cycles of forecasts out to 03 UTC. This latter ensemble uses the ensemble Kalman filter to assimilate surface data, aircraft and satellite observations.

A big part of the experiment will be spent evaluating the models, the techniques, and the forecasts themselves. We hope to highlight the model capability at this higher temporal resolution. We will also make it a point to identify good metrics that reliably identify good forecasts as compared to what forecasters will also deem as good. In this way we put the metrics to the test and discuss the strengths and weaknesses of them.

Today was mostly about showing all the tools and models at our disposal while making a forecast for thunderstorms with hail in the North Carolina and Kentucky area. The SSEF, SSEO, and AFWA convection allowing models were in moderate agreement about producing a few strong storms. The uncertainty was moderately high (only half of the corresponding members of each ensemble produced these stronger storms), but confidence was high that a few of them would be capable of producing hail. Sure enough we put our forecast for 21-00 UTC for the hail and the first and biggest report occurred at 20:55 UTC, with the rest in NC occurring thereafter.

I think for the next post I can mention some of the high resolution variables that we use as proxies for severe weather. And future posts will discuss some other experimental products that we get to see including the UKMet offices' Unified Model.

Leveraging the known

I left off the last post indicating that we know model biases, at least for the models that are coarser in grid spacing. But that is a lie.

Metrics: Pick your number

How exactly do you choose your favorite model?

Because that is the model you fall back to when uncertainty is large. The model you use when a big event is forecast. The model you are most familiar with. The model you use to re-calibrate yourself when a big event is forecast.

Verification

One of the many struggles with forecasting is verification, especially of "rare" events. In the severe storms world, we have storm reports. It has been shown that there are serious flaws with this database over time. They are conditional upon hitting something or someone; population density or highways. There are many areas un-accessible that may have had things like hail or high wind but yet there are no observations nearby. In the Plains this is a big challenge.

Over the course of the HWT EFP, we have debated over the so-called practically perfect methodology. At is core it is a Kernel Density Estimation (KDE) technique using a gaussian smoother (120 km) and a radius of influence technique (40 km) to map individual storm reports to a grid and produce probabilities of severe.

Given that storm reports are not the most ideal, independent, non-biased dataset out there we look to other more unbiased data. So what are the alternatives? Can we use severe storm warnings? How about data specifically from the radars like Max Estimated Hail Size or Rotation tracks? How about satellite derived data about vegetation?

Just about all of these data sets have their own problems. For the radars, we are observing rotation or hail aloft, not necessarily at the ground. This is still valuable information. But how do we switch from spotty storm reports to continuous tracks? Will the same KDE smoothing approaches be necessary?

For the warnings, it is clear that meteorology alone is not driving them. If there is a chance a storm could be severe over a highly populated area at a critical time, the edge goes to issuing warnings as opposed to not. This is not all that bad, since we would all like to err on the side of safety. Better to be safe than sorry.

Using  radar data we still have to verify that what the radar detects is actually occurring at the ground and that phenomena is as strong/large as indicated aloft. And that requires doing verification on the observations. The SHAVE folks at NSSL-OU are trying to do exactly that as are some other NWS associated folks, though at the moment their name escapes me.

Satellite data also offers some advantages on tracks of severe storms provided there is damage say from large hail stripping vegetation bare or tornadoes doing damage.  Collecting more fine resolution data is going to take a dedicated effort but in the end it helps build more complete knowledge about storms, more understanding about the successes and failures of the forecasts, and quite possibly will end up making better forecasts.

Thanks to TAMU for soundings

One of the observation components of this years EFP has been an intercomparison between the Vaisala RS92 and InterMet radiosondes to help verify the Microwave Radiometer that we (Dave Turner at NSSL) have on the roof of the National Weather Center. We were lucky to have Don Conley from Texas A & M bring his observations class on the road and visit the HWT conduct some local and mobile radiosonde intercomparisons. They did two mobile deployments: one in Concordia, KS on Wednesday and Altus, OK on Friday for helping to verify the models we are using for convection initiation.

They drove from Norman to Concordia and were able to make 3 launches (4 really), and 2 trips to Walmart (for helium and then to return said helium). Many thanks to the City of Concordia and the airport manager for allowing them to use a hangar for these balloon launches in very strong winds (which caused the failure of the very 1st balloon launch). They got to a great spot just east of 2 very long and robust Horizontal Convective Rolls both of which produced CI along the front-HCR intersection.

I haven't heard the stories from today, but I do know they got to Altus after lunch at Meers (for the Meers burger, obviously) and got off two launches again in an environment characterized by HCRs. These are great tests of the instruments, great experiences for the students, and excellent learning opportunities for the rest of us.

They (and you) should know that these soundings make their way to the SPC (something that is usually done upon request at TAMU) and prove valuable. These types of partnerships, sometimes ad hoc, but almost always mutually beneficial are what make the HWT a vibrant place for forecasting, research, research forecasting and forecasting research; and now with observations!

You can find the mobile and local soundings here.

Again, Thanks Don, Mike D., Mike C. and the whole Observations Class (Send me your names and we can make you famous* write them on here!).

*Fame not guaranteed.