[StratoQ]

In recent days we have seen quite large splits in the ensembles. Many of the members going for cold solutions and just as many for a milder outcome. Using a 10 pin bowling analogy, this is like a "7-10 Split", the most difficult of all to convert to a spare. Equally these model solutions don't inspire much confidence to the average viewer.

My question is : is there a programming "learn" facility built into the models whereby previous runs are re-analised to see which member came closest and fine tune the rest? Or anything similar?

This post has been edited by StratoQ: 27 January 2012 - 14:20

[grahamread]

It would be interesting to know. I imagine the problem is "tuning" a model to better predict one type of event correctly, may then make it less good at handling a different set of starting conditions...at least that's my very non-mathematical take.

[Tony Gilbert]

I have considered this question many times for the past decade!!
Would be good to have the answer from someone who actually works with the computers that actually funnels this data into its final format. Not sure if such a question can be answered by a professional forecasters?..... Pretty much like asking the Pilot of a 747 to explain the fine technical detail of the engineering within the plane!

This aside; If the various current model forecast are not ingesting seasonal anomalies such as SST, topography, urbanisation, and climate change then we are pretty much stuck in the mud ATM. On a personal not I do not know enough about how this data is crunched into the super computers to make a reliable opinion. Though having said that, if the aforesaid detail is not being incorporated into the models then we have no hope of moving forward.

This post has been edited by Tony Gilbert: 27 January 2012 - 19:23

[smartie]

StratoQ, on 27 January 2012 - 14:19, said:

In recent days we have seen quite large splits in the ensembles. Many of the members going for cold solutions and just as many for a milder outcome. Using a 10 pin bowling analogy, this is like a "7-10 Split", the most difficult of all to convert to a spare. Equally these model solutions don't inspire much confidence to the average viewer.

My question is : is there a programming "learn" facility built into the models whereby previous runs are re-analised to see which member came closest and fine tune the rest? Or anything similar?

Modern numerical models have no 'heuristic' ability, ie. they don't learn from past 'errors' or 'successes'. However, they undergo changes, upgrades, testing and evaluation which (usually) results in improvements in their performance, as assessed by various forecast skill scores deemed useful for assessing this performance.

Unfortunately, 10 pin bowling isn't a very good analog for ensembles, ;-). But ensemble forecasts are currently considered one good way of tackling the difficullties of modelling the chaotic atmosphere/climate system and inherent forecast uncertainty. I guess as ensemble forecasting is still in it's infancy understanding is yet to filter down to 'non-specialists'.

Modern climate/global/mesoscale models include representations of all the physical quantities / variables Tony mentions. Eg. the WRF model (for real data implementations) routinely incorporates terrain (at a suitable resolution for the grid), land surface representation ( variable/ seasonal from satellite/quasi homogeneous datasets), SSTs (seasonal or satellite &/ buoy datasets) including the urban land surface. Also variables/ quantities important for radiative processes such as trace gases as Ozone/ CO2 can be assimilated/represented.

Best,
David

This post has been edited by smartie: 27 January 2012 - 20:06

[Tony Gilbert]

PS. The common statement that a good forecaster should read between the lines within the forecast models would seem a bit nonsensical if we were to consider that the whole point of the models is to be as accurate as they possibly can in the first instance?... After 12 years of amateur level forecasting I am now able to note specific variations in the models which might suggest an alternative outcome. Though why should a forecaster have to correct an under, or over estimation within the models in the first place? ...Surely if the data is entered correctly into the super computers with all the various regional and seasonal variations in the first instance then the final model output should in theory be more accurate?
Whilst I understand and agree that the broad variations in the anomalies will continue to flounder the best weather model computer, after this many years of crunching data surely we should be getting a better picture by now?...and if not is there some room for changing the way in which the data can be extrapolated to include the seasonal, topographical, regional, urbanisation etc variations??

I'm sure current UKWW forecaster and media forecaster do the best with the data that is available to them. Though I truly believe this data currently makes hard work of forecasting with alot of mis forecasting which cannot be fairly attributed to the forecaster, but moreover to the tools available to him!

To say well I've learned to decipher the models for my own use and I'm OK with that, is surely limiting the potential for a bigger picture for our future?

This post has been edited by Tony Gilbert: 27 January 2012 - 20:29

[Tony Gilbert]

smartie, on 27 January 2012 - 20:04, said:

Modern numerical models have no 'heuristic' ability, ie. they don't learn from past 'errors' or 'successes'. However, they undergo changes, upgrades, testing and evaluation which (usually) results in improvements in their performance, as assessed by various forecast skill scores deemed useful for assessing this performance.

Unfortunately, 10 pin bowling isn't a very good analog for ensembles, ;-). But ensemble forecasts are currently considered one good way of tackling the difficullties of modelling the chaotic atmosphere/climate system and inherent forecast uncertainty. I guess as ensemble forecasting is still in it's infancy understanding is yet to filter down to 'non-specialists'.

Modern climate/global/mesoscale models include representations of all the physical quantities / variables Tony mentions. Eg. the WRF model (for real data implementations) routinely incorporates terrain (at a suitable resolution for the grid), land surface representation ( variable/ seasonal from satellite/quasi homogeneous datasets), SSTs (seasonal or satellite &/ buoy datasets) including the urban land surface. Also variables/ quantities important for radiative processes such as trace gases as Ozone/ CO2 can be assimilated/represented.

Best,
David

Hi David, sorry we must have cross posted. Thank you for your input

[PJB]

Models do not "learn" as such but they do look at previous model positions vs new observations. This is known as 4D VAR assimilation. 4D-Var finds the atmospheric state (or "analysis") which "best fits" both the prior information (or "background", a short forecast from a previous analysis) and recent observations.

Models Such as the UKMO GM and ECMWF employ this type of analysis.

You can read more here

http://www.metoffice...4d-var-research

[StratoQ]

Thanks Paul/David/Tony.

I wasnt aware of 4D Var before and it is good to hear of such advancements in model development.