Thursday, February 3, 2011

Verification of the composite snow forecast parameters

In the days leading up to the blizzard that struck the southern plains and midwest early this week, all the buzz was on different model forecasting sites that were spitting out impressive snow totals. The pretty colors and high totals had these images going viral minutes after they were released. The most common of which was likely WGN's RPM model run on Sunday afternoon which showed a staggering swath of 30"+ totals over northern Illinois and surrounding areas. It's a fairly well know fact that buying completely into these composite maps is simply naive, but just how accurate are they? They're generally posted with a "totals will be slightly less than this" disclaimer. I decided to save the model outputs from Earl Barker's WXcaster site in particular as it's one of the more commonly used models and does not seem to put out the outrageous totals that some in house model suites do for the days leading up to the blizzard in order to see how well the model rendered version of the NAM and GFS models did in the totals and placement of the heaviest totals.

On each map, the yellow shaded figures depict the actual measured storm total snow fall, overlaying the model forecast totals.

We'll start with the 12z GFS for Monday, the day before the storm:

It's hard to argue with this one. On a whole, the idea of taking the extreme totals and shaving a couple inches off seems to jive well with what the map shows. In addition, it handled the rain/snow cutoff line in the south fairly well. As expected, the northern extent of the heavier forecast swath didn't do quite as well. A northward shift in the system as it developed and moved through the area is partially to blame. From Iowa City, IA to Rockford, IL we see a lot of 9-12" totals in the red shaded areas that would have suggested totals surpassing 15". Then, take a look at the Kankakee area in northeast Illinois. I mentioned in my forecasts prior to the event that I expected both dry-slotting and a higher amount of precipitation falling in the form of sleet to cut down on totals along Interstate 57 in the Kankakee area which is exactly what transpired (pat on the shoulder for the Pritchard model!). This run of the GFS would have suggested in surplus of 20" of snow, where they really only experience about 30% of this forecast total. Dry-slotting is a hard to forecast, but crucial element to these powerful storms. A combination of dry-slotting during the evening, and a mix of precipitation types is also to blame for almost 50% reductions in forecast totals from Springfield to Bloomington, IL. The model at 36 hours in advance of the storm handled the overall southern transition line between liquid and frozen precipitation well, but did not handle the blend of sleet and snow well at all. This is probably the single most important reason to not buy into extremely high snow totals automatically. Ask yourself if they truly make sense, and what reasons other than a truly historic snowfall could be responsible for the model output.

Now, the 12z NAM forecast:

This model was on a whole, way too far south with it's heaviest snow totals swath. Again, it included the Interstate 72 corridor in 20" snow totals where in reality a blend of snow and sleet kept totals at only about 30% of this level. In fact, the entire eastern half of the heavy band in the forecast output fell victim to mixed bag precipitation types and dry-slotting as the storm moved into the area. The northern part of Illinois was significantly under forecast on this model run, which owes to the northerly shift in the system that neither the GFS or the NAM handled.

18z GFS Monday, the day before the storm:
The 18z GFS, oh those lovely off-hour forecasts that always seem to throw in some extreme variable to excite weather enthusiasts. Upper air sounding are only done twice a day under typical circumstances, at 12z and 0z. The 6z and 18z model suites have no new upper air data ingested, so falling for their every move isn't advisable. This particular off hour run of the GFS did not handle the southern line of the heaviest banding well at all. Areas along that St. Louis to Kankakee line are forecast to be along that southern line of totals in excess of 1 foot, but experienced prolonged periods of freezing rain and sleet. The northward shift and precip type difficulty again plaguing the model.

Now, we've got the 00z NAM for the night before the storm:
The most noticeable trend on the 00z output is the widespread reduction in snow totals. The path is on a whole unchanged, but totals have taken a significant hit. The biggest reason for this was the forecast increase in convection in the deep south near the Gulf Coast. Widespread convective storms were forecast to break out across the deep south (which did in a sense occur, but not to the extensive level forecast) which will commonly rob the conveyor belt of moisture from feeding into the low pressure system to the north and reduce snow totals. This is a legitimate concern that needs to be addressed even if the model run does not suggest it will be an issue. This was also the model run that spit out incredible 24-30" totals in the Oklahoma City area which went viral seconds after the model was released. I was immediately skeptical of the over-forecasting of snow totals due to mixed bag precipitation, which indeed turned out to be the case in most areas. The one thing this model did begin to handle well, was the enhanced snowfall accumulations near the Lake Michigan lake shore in Chicago.

12z Tuesday NAM from the day of the storm:
The morning of the storm, we see the models catch on to the northerly shift in the system track. Areas along Interstate 72 that I suggested would see a high amount of sleet and a high reduction in their snow totals are now beginning to be backed up by the model output. Without being too critical of the model's transition line forecast, 4-8" totals from St. Louis, MO to Danville, IL are on a whole fairly well handled. The yellow shading of the snow totals along this line clash with the colors a bit, but I assure you they're there and do jive with what the NAM suggested. Also notice that snow totals have increased dramatically again. The overall track of the heaviest band is well handled, but again, shaving 10% off the forecast total seems fair. The tan band from NE Missouri into the Chicago area suggesting 18-20" is spotted with totals that do fit in to that range. A few outliers are easily noticeable inside the band where totals on the order of only 13" or so occurred, and then 20"+ totals were found near Lake Michigan where lake effect snow occurred before and after the storm. The variations between actual totals in the heaviest areas were due to the model's inability to forecast where the heaviest mesoscale banding and embedded thunderstorms were to occur. Among precipitation type and dry-slotting, depicting areas that will experience the heaviest banding inside the deformation zone is something that a model of any resolution simply cannot forecast in advance. It can suggest a region where it is likely, as we see here, but getting down to fine resolution isn't going to happen.

And just for kicks, the 18z NAM from Tuesday, as the storm began impacting the area:
Most notable is the dramatic over-forecasting of totals in northeast Illinois from Kankakee southwest to just north of Champaign. The NAM, at even mere hours out still can not handle the fact that the dry-slot was already pushing in to southern Missouri taking aim on this area. I mentioned in my forecast discussion, as did many others forecasting the storm that portions of eastern and northeastern Illinois could see a lull in precipitation because of this, but it just is not something that forecast computer guidance has a good handle on. For what it's worth, the 4km WRF precipitation forecast model did try to handle this factor with relative success. Some over forecasting of totals is seen in eastern Iowa around Iowa City, which is likely due to the model suggesting inaccurately that mesoscale banding and heavy convective snows would set up in this area, which did not occur.

This isn't ground breaking evidence, as the difficulties that higher resolution models experience in forecasting snow accumulation were already commonly known but it's an interesting overlapping look at where the models seem to struggle the most. Exaggerated totals were common along the southern track of the heaviest forecast due to a tandem of inability to establish the crucial correct track of the center of low pressure, and the inability to distinguish precipitation type in high reflectivity zones where heavy freezing rain or sleet may be the dominant mode rather than obscene snow totals. Dry-slotting along the southern edge of the heavy forecast totals swatch is also another factor that the models commonly struggle with.

Ultimately these pretty colors are nothing more than that; pretty colors that can give a broad suggestion of the area that will see the heaviest totals, and how heavy these totals will be. Taking these specific totals as gospel is falling victim to meteorological cancer. Take this model suggestion, and pick it apart asking yourself if it makes sense given the above mentioned impacts.

No comments: