From: Phil Jones <p.jones@uea.ac.uk>
To: "Humphrey, Kathryn (CEOSA)" <kathryn.humphrey@DEFRA.GSI.GOV.UK>, "Stephens, A (Ag)" <A.Stephens@rl.ac.uk>
Subject: RE: Questions on the weather generator
Date: Fri Jan  4 12:07:45 2008
Cc: "David Sexton" <david.sexton@metoffice.gov.uk>, <C.G.Kilsby@newcastle.ac.uk>, "Jenkins, Geoff" <geoff.jenkins@metoffice.gov.uk>

    Kathryn,
          I did talk to the Metro yesterday - no idea what they used. Maybe a few will
    have read it - before copies are tossed around on the tube!
       Added Geoff on this email.

         Ag has answered the second question. I may come back to that after
    trying to answer the first part.
         There are two aspects to the WG work we're doing. The first, which I've mentioned
    on a number of occasions, is to prove that the perturbation process used with the WG
    works. Colin Harpham sent around a load of plots to Chris/Ag/David/Geoff just before
    Christmas. I have a rough draft of a paper on this which I sent to Chris yesterday. This
    involves the UKCIP08 WG, but is totally independent of the change factors David is
    developing for UKCIP08. This uses some earlier HadRM3 model runs. The WG is fit to
    10 grid box series across the UK and then perturbed according to the differences between
    the future model integrations and the control runs. We then generate future weather and
    show that its characteristics are similar to what HadRM3 got directly. This has used
    the same change factors (same variables) but from a different set of RCM runs.
      The whole purpose of this exercise is to show that the perturbation process works.
    The only way we can test this is to use RCM model runs - because they have future
    runs with a big climate change. We can't use past weather data as it doesn't have
    enough of a climate change. This is validation of the perturbation process.
       We can additionally validate the WG using observational data - which we've done
    earlier.
      Return to Q2. Ag has said how the model variants get chosen. The model variants
    used have a variety of ways of being chosen. Let's say we start with the 50th percentile
    for rainfall. We select all model variants between 45 and 55%. Then we want temperature
    at the 90th percentile. We then do a second selection of the variants already selected
    that have temperature changes between 85 and 95%. As we had initially 10,000
    variants, the first selection reduced this to a 1000 (as we chose 10% of them). The
    second selection reduced this to 100 (as we've again chosen only 10% of them).
      Now with these 100 variants, most users will average the change factors (from David)
    across these 100. These average change factors (which will approximately be
    at the 50% and 90% value for precipitation and temperature respectively) get passed
    to the WG. The WG then simulates 100 runs of 30 years - for the already
    pre-selected location (small area) and future period.
       There are obviously loads of permutations as we will be allowing users to select all
    percentile levels (singly for temperature or precipitation) or jointly for both from
    5 to 95 % in steps of 5.
    The percentile levels can be chosen based on seasons (4) and years (1). If you
    select summer say, users will also get the rest of the year - using the change factors
   that
    go along with those for the selected model variants.
       Another possibility is to select one model variant within the chosen percentile bands
    and pass these change factors to the WG.
      There are other possibilities, but I think we've limited the choices to these two.
    The other possibility was a variant (can't think of a better word here - but not
    related to the model variants) to the first. As you have 100 chosen model variants
    in this example, you could chose one at random or allow each of the 100 WG
    integrations to be based on a different one of the model variants. These generated
    sequences will likely have greater variability than that based on the average of the
    100 or that based on the single model variant.
     I think this may open up a can of worms with Ag when he reads it !


     Whichever of these are chosen, the use should still run the WG for
    100 30-year sequences.
      I think I've made the last bit on model variant selection complicated
    and haven't gone back to look at what Ag has written in the User Guidance.
    It ought to tell you how the change factors that the WG needs will get selected.
    Cheers
    Phil

   At 10:07 04/01/2008, Humphrey, Kathryn (CEOSA) wrote:

     Hi Ag,

     Yes that makes perfect sense in terms of selecting one/several model variant/s, thanks.
     I'm still a bit confused about the utility of random sampling though as this won't give
     you results for a particular probability level (will it?).  I think Phil was going to
     get back to me on this as well as the change factors question.

     Phil, I liked your quote in the Metro this morning!

     Kathryn
       ___________________________________________________________________________________

     From: Stephens, A (Ag) [[1]mailto:A.Stephens@rl.ac.uk]
     Sent: 04 January 2008 08:56
     To: Humphrey, Kathryn (CEOSA)
     Cc: Phil Jones; David Sexton; C.G.Kilsby@newcastle.ac.uk
     Subject: RE: Questions on the weather generator
     Hi Kathryn,

     I can comment on your second question. Here is my understanding:

     Firstly, users must run a minimum of 100 WG runs regardless of which ones they run. This
     is to enforce the use of a "probabilistic" approach.

     Selection by model variant will only make sense once a user has produced some runs.
     After any run they will have access to the model variant IDs that were used. The use
     case that gave rise to us including "selection by model variant ID" was as follows:

     1. Person X does some WG runs (sampling by whatever method she chooses).
     2. She uses/analyses a set of runs to produce some interesting results.
     3. She is keen to do more/different analyses using the model variants that represented
     that part of parameter space.
     4. She has the list of model variant IDs so she can publish these so that others can use
     them or she can re-use them herself in other experiments.
     5. Person Y can read about what Person X did and re-produce exactly her results, or use
     the same set of interesting model variants for some other experiments.

     Does that make sense?

     Cheers,

     Ag
       ___________________________________________________________________________________

     From: Humphrey, Kathryn (CEOSA) [[2]mailto:kathryn.humphrey@DEFRA.GSI.GOV.UK]
     Sent: 03 January 2008 16:58
     To: Stephens, A (Ag)
     Subject: FW: Questions on the weather generator
     ______________________________________________
     From:  Humphrey, Kathryn (CEOSA)
     Sent:  03 January 2008 16:55
     To:    'Phil Jones'; 'Chris Kilsby'; 'Stephens, Ag'
     Subject:       Questions on the weather generator
     Phil/Chris/Ag,
     I'm putting together a "quick and easy" presentation on the UKCIP08 methodology for
     Defra officials to give them some idea of how it's all done so they can better
     appreciate what's it's potential uses may, and may not, be.
     However I'm getting stuck still on some of the WG methodology!  Can you help?  (I'm not
     planning on telling them this level of detail about the WG but am just bothered by the
     issues below).
     I'm firstly confused about the RCM change factors; are you using these to validate the
     WG runs (which I do understand) or to generate them (which I don't as I thought they
     were being generated using the data in final PDFs themselves)?
     And I'm still confused about the reasons for allowing users to select runs by model
     variant.  I think by model variant you mean each perturbed version of HadCM3 or other
     single model run or emulator result that creates a point in parameter space.  Is this
     right?  If so then I understand why you can't run your WG on all model variants (too
     many) so selecting a random sample is a representation of parameter space.  But my
     initial understand of how the WG works is that you pick a point on the PDF (say 50th
     percentile) with a given probability and run the WG for that point.  But this doesn't
     make sense if you are allowing users to select random/ single model variants seasons
     etc. because these won't reflect a particular percentile.   Maybe it's the case that you
     don't need a particular percentile for whatever use the WG data is for, but if you don't
     know, how do you know how likely your WG output is and therefore what to do with the
     result in terms of planning?
     Apologies for my ignorance and assistance would be gratefully received!
     Kind Regards,
     Kathryn
     Kathryn Humphrey
     Climate Change Impacts and Adaptation Team, Defra
     Zone 3F Ergon House, Horseferry Road, London, SW1P 3JR
     tel 0207 238 3362 fax 0207 238 3341
Department for Environment, Food and Rural Affairs (Defra)

This email and any attachments is intended for the named recipient only.
If you have received it in error you have no authority to use, disclose,
store or copy any of its contents and you should destroy it and inform
the sender.
Whilst this email and associated attachments will have been checked
for known viruses whilst within Defra systems we can accept no
responsibility once it has left our systems.
Communications on Defra's computer systems may be monitored and/or
recorded to secure the effective operation of the system and for other
lawful purposes.

   Prof. Phil Jones
   Climatic Research Unit        Telephone +44 (0) 1603 592090
   School of Environmental Sciences    Fax +44 (0) 1603 507784
   University of East Anglia
   Norwich                          Email    p.jones@uea.ac.uk
   NR4 7TJ
   UK
   ----------------------------------------------------------------------------

References

   1. mailto:A.Stephens@rl.ac.uk
   2. mailto:kathryn.humphrey@DEFRA.GSI.GOV.UK

