Environmental Engineering Reference
In-Depth Information
Steve Hanna:
Would it help to use second order closure in the models and therefore
directly predict the uncertainty (pdf) in addition to the mean?
Answer:
This is a technique that we have used to transform emission uncertainty
into concentration uncertainty (see Galmarini et al., ACP, 2008). Unfortunately
that could work only for uncertainties relating to surface-atmosphere interaction
where the boundary layer and its turbulence act as system for the transfer of
surface information to the upper atmospheric levels.
Peter Builtjies:
Could we use the fact that in most cases the ensemble/average has
a higher correlation than the single model as information about 'stable' and
more uncertain model aspects?
Answer:
To answer your question you would need to gather precise evidence and
in formation. However you would agree with me that we are lacking fundamental
research on ensemble behaviours. It is clear by now that the ensemble produces
a results that compensate for opposite behaviours of individual models, however
being a practice, based on heuristic approaches, it does not guarantee of maximum
exploitability neither gives hints on which pre-requisite a model should have to
produce yet a better result. A theoretical base would help greatly in that setting the
boundaries of applicability of the practice and ingredients for its improvements.
Iversen Trond:
I tend to disagree with you that combining mediocre model results
into ensemble is not a valid strategy compared to improving the models. I think
it well established theory that as long as the ensemble does not sample the
entire uncertainty space, any added, bias-free, ensemble member will improve
the ensemble beyond the single ensemble member.
Answer:
Well it depends on what we need with mediocre, and how bad is bad
enough. Your argument would be correct in principle if the members would be
independent. We know that our models are not independent at all and they may
all be biased toward a specific solution that would be closer to reality but still
further than it could be if we could get more clear idea on what it means being
independent, estimate the level of dependency, get to know the impact of the
dependency on the final result. I find paradoxically that from one side efforts
and resources are invested in model development, extremely detailed process
modelling, getting our models as deterministic as possible but then we do not
care on the quality of their results and their impact of an ensemble because
finally averaging will sort all out.
