Jump to content
Snow?
Local
Radar
Cold?
IGNORED

Numeric Modelling and Ensemble Discussion


Dawlish

Recommended Posts

Posted
  • Location: Steeton, W Yorks, 270m ASL
  • Location: Steeton, W Yorks, 270m ASL

The analogy I would tend to use for ensembles is the simple model used by Richard Dawkins twenty odd years ago to demonstrated genetic mutation. He programmed a simple routine - this in the days when a C64 was a powerful home-use machine and the zenith of home computing - that started with a given shape, and then mutated this shape in one of a number of ways using random numbers for each new generation. If you can't get your head around this, think of a simple left-right decision. At layer 2 there are two positions: left or right: at layer three there are three: twice left; twice right; or one left + one right (which can be arrived at two ways). A pyramid build, based if I remember rightly on Fibonaci/s sequence, of possibilities.

Met modelling will follow the same principles; the further out you go the more options there are for outcomes, hence why the tendency to pick the correct one becomes less and less, and why forecast accuracy reduces. The question is whether or not the ensembles actually add any value a priori (they might, a posteriori[?] - to which I shall return). The added confusion is that the choice of branches is not equally probablistic; nearer t=0 there is more certainty in specific outcomes, further out the certainty reduces.

The models work, as I understand, by using the control run starting from "now" and loading best known probabilities; ensemble runs then change the odds on certain factors. This begs a few questions:

- If an ensemble turns out better than a control run, despite being based on "flawed" assumptions, then the assumptions underpinning the control run clearly are not correct;

- If the control run is based on best knowledge then ensembles are largely meaningless, unless the confidence in the control is low, in which case the ensemble runs lust themselves be flawed as they branch out from the control run.

A posteriori ensembles might be helpful, because they tell us which factors, asumed beforehand to be wrong, actually were not so wrong. I suspect, from a modellers perspective, the ensembles are most use after the event, just like animals that die in medical experiments can be more useful than the ones that live.

The bottom line is this, whatever the cluster of the ensembles out beyond the point where they diverge, the divergence itself is the point at which all downstream confidence is lost. There are a few on here who will occasionally look at ensembles which then reconverge (be it mild or cold), and suggest that there is agreement. For sure, there is apparent agreement, but don't overlook the fact that all that is happened is that different runs have arrived at the same WRONG point by making the same set of wrong decisions, albeit in different order. A model cannot randomly recover from a critical wrong turn, other than by chance; for all runs to do the same, so far out, is mathematically so unlikely as to not be worth considering.

Link to comment
Share on other sites

Guest
This topic is now closed to further replies.
  • UK Storm and Severe Convective Forecast

    UK Severe Convective & Storm Forecast - Issued 2024-03-29 07:13:16 Valid: 29/03/2024 0600 - 30/03/2024 0600 THUNDERSTORM WATCH - FRI 29 MARCH 2024 Click here for the full forecast

    Nick F
    Nick F
    Latest weather updates from Netweather

    Difficult travel conditions as the Easter break begins

    Low Nelson is throwing wind and rain at the UK before it impacts mainland Spain at Easter. Wild condtions in the English Channel, and more rain and lightning here on Thursday. Read the full update here

    Netweather forecasts
    Netweather forecasts
    Latest weather updates from Netweather

    UK Storm and Severe Convective Forecast

    UK Severe Convective & Storm Forecast - Issued 2024-03-28 09:16:06 Valid: 28/03/2024 0800 - 29/03/2024 0600 SEVERE THUNDERSTORM WATCH - THURS 28 MARCH 2024 Click here for the full forecast

    Nick F
    Nick F
    Latest weather updates from Netweather
×
×
  • Create New...