Forums (Forecast Builder)

Back

Superblend observation

MF
Michael Fuhs, modified 8 Years ago.

Superblend observation

Youngling Posts: 8 Join Date: 4/15/16 Recent Posts
This note really has nothing to do with forecast builder, but addresses superblend.  It would be really neat if we could somehow incorporate a lot of ECMWF weighting in superblend.  This last go 'round with recent warm weather this past weekend in the forecast area was totally missed by the GFS, and to some degree the GEM.  I mean we are talking about 15 degrees too cold for highs this past weekend on the superblend.  The Weather Channel forecast, whatever they use, was much better then ours.  A 12 hour old ECMWF is still "better" than a new GFS and a new GEM.  The ECMWF picked up on the warming this past weekend, and I remember making a point of that early last week in my AFD in that if the ECMWF panned out for this past weekend, temperatures would be much warmer.  My guess it would take a heck of a lot of work to incorporate the ECMWF.  In addition, in the long run this is all just a moot point with the NBM coming down the road.  But trust me, if things need changed, you are not going to get anybody to change temperatures more then just 2 or 3 degrees from the third period on, and days 4 through 7, forget it, even though many times the ECMWF is a superior model.  Conversely in mid May when we had a cold and rainy week with an anomalous closed low in the plains, the ECMWF and even the GEM had a lot of highs in the 40s (pressing record cold max's) with the constant rainfall while superblend had highs in the 50s to near 60 in days 3 through 7.  Even fully mixed from 850mb on the GFS, you would have trouble hitting the forecasted superblend highs, with superblend and consall consistently the warmer guidances, even much warmer than consraw or bcconsraw.  It was interesting to watch the superblend highs drop a few degrees each day upon getting closer to the rainy days, playing catchup with the ECMWF.  Most concerning, with our region basing DSS graphics off the superblend forecast from the third period on, it really can be a tricky siutation.  With the NBM incorporating virtually no ECMWF that I am aware of, this problem will likely only get worse. 
JW
Jerry Wiedenfeld, modified 8 Years ago.

RE: Superblend observation

Youngling Posts: 4 Join Date: 11/21/12 Recent Posts

Hello,

Thanks for the post.  Currently in the extended the ECMWF is weighted equally with GFS, Canadian, MOSGuide, ADJMEX and ADJECE in SuperBlend.  For a while we did double weight the ECMWF but when we switch to the ForecastBuilder test we dropped the double weight.  We have not noticed a significant degradation by reducing the weight in the bulk verification statistics. It is true that the NBM will not contain the ECMWF, but again they did some early comparisons by including it and not including it and while there was some degradation it mostly feel in the noise level. At this point the use of the ECMWF in the NBM is out of our hands.  All we can do is use it as an input model for the messaging.  If that message has an impact on our gridded forecast like your example we would have to collaborate grid modifications with neighbors.

 

Thanks again,

Jerry

CN
Christopher Noles, modified 7 Years ago.

RE: Superblend observation

Youngling Posts: 1 Join Date: 10/23/15 Recent Posts

Something to consider, regardless of ECMWF incorporation. This fall, the GFS has been a rather poor performer for our area. The SuperBlend 6 hour PoP grids we collab over (at least in Central Region), in the longer term, incorporates the GFS's "raw" numerical number. So, it is quite common for us to have to collaborate PoPs every 12 hours (initialization twice a day from the Forecast Builder CRON) in a more active pattern because of multiple things going on. One, the GFS has been a poor performer. Two, the 06z/18z initialization in effect writes over a previous forecaster's collab effort because of minimal previous forecast weighting. Three, this work load aspect is increased even more by the fact that temporally, we are having to look at and adjust for 6 hour periods beyond the near term. Farther out in time, timing swings can be anywhere from 6 to 18 hours for significant systems. It would be nice to see a greater uncertainty element built in (option to use 12 hour PoPs?). Because, while there is nothing wrong with high PoPs in the long term if confidence is high... there are too many instances of going from 50-80 PoPs down to 10-20 PoPs for a 6 hour forecast primarily day 3 or 4 and beyond. WPC still creates 12 hour PoP grids Days 4 and beyond. There may be a reason for that. One other issue that will come up. During the collab process, sometimes forecasters will just punt and say, "well I don't have a good feel for things, so I'll just go with the blend." That mind set has to disappear. The writing is on the wall, that this agency doesn't want us spending a whole lot of time in the long term portion of the forecast. Well, if so, some adjustments may still need to be made. Users and partners use our long term forecasts for planning purposes. Maybe there is a way to make it easier on us, while still putting out good products.

One a more positive note, the near term hi res grids have proven invaluable at times. Thanks for the work that has been done on all of it so far. These are just some thoughts, observations, not necessarily complaints.

JM
Joseph Moore, modified 7 Years ago.

RE: Superblend observation

Youngling Posts: 72 Join Date: 11/1/13 Recent Posts

Christopher, I definitely agree with you on the concern that the extended grids are driven too much by basically just three deterministic models. For an agency that has preached the utility of ensemble modeling from top to bottom, it's extremely disappointing to me that in 2017 we have no real way to incorporate any ensemble guidance in our medium/long term forecast, when it is of great value. I would love to see NIC include at least GEFS guidance, if not NAEFS.

That said, there is some good news: NBM incorporates a heck of a lot more guidance than we currently have in GFE/D2D (see below), and in my experience because it has such a diverse set of input data it often does a much better job on spread. It has higher and more widespread POPs in general while eliminating the silly splotches we get ALL THE TIME from some random CMC/GFS run that we often have to clean up. I know CRGMAT would love to go with NBM over superblend as soon as they feel it is a reliable replacement. (ForecastBuilder Developers: While I think there are some issues with NBM in the near-term, I think it would be an excellent candidate to phase in starting with the day 4-7 period. (I realize the current code can't handle this at the moment, but in my opinion this would be a great way to start using the guidance in an operational manner.)

NBM Extended models: The NBM Vlab page has more details (see the table graphic below the cadence/timing graphic), but the short version is that NBM has (1) GEFS, (2) GEPS (aka the CMC ensemble), and (3) NAVGEME.