[TRNSYS-users] Parameter identification.
Enriquez Miranda, Ricardo
ricardo.enriquez at ciemat.es
Thu Jun 27 03:10:02 PDT 2013
Hello TRNSYS World, hello Didier.
I have one more cent to add to this huge topic.
The technique you should use for Parameter identification will depend on your final objective (model validation, model calibration, model order reduction, assessment, forecasting ...). I agree with David and Angel. Keep things as simple as possible, but no more. In a general basis I would recommend you the following steps:
1. Sensitivity analysis. This will show you in a quantitative manner the minimal set of active parameters of the model (Those which explain the variance of the model output).
2. Solve correlations. Take care of those active parameters which are correlated. That is a limitation of the method unless you take action. If you are at the experimental design phase move to a multioutput framework, it will solve if you are dealing with a building.
3. Choose a parameter estimation algorithm. For dealing with a single objective problem I use Differential Evolution, which is pretty fast (http://www1.icsi.berkeley.edu/~storn/code.html). If you solve correlations in the multioutput way, you should move to a multiobjective framework, most probably dealing with a multidimensional Pareto Front characterisation (ask me personally).
4. Uncertainty assessment. You should also assess your model and experimental uncertainty. There is a ISO guide from 1995 to the experimental. You can assess your model uncertainty from a Montecarlo Analysis, for example. This is important in two ways: if your model uncertainty is bigger than the experimental one you should redesign your experiment. On the other hand, any model whose outputs are within the experimental uncertainty are physically indistinguible with that experiment.
5. Residual analysis. If your error is greater than the uncertainty you cannot consider the model valid. To improve it you can perform residual analysis together with a parameter space analysis technique, for example.
6. Choose your software. Input files are plain text written in some descriptive language with a predetermined format. You can manage simulations in many ways. I started managing with GenOpt, then move to Matlab and right now in Python, since I'm more comfortable with my own scripts. The important issue here is: think carefully what will you need for your study and then move to the tool. Python is quite flexible for doing this tasks. Don't be afraid to make your own code, you will have a lot of fun.
You can download a paper for free with some of this studies applied to a Spanish Building:
Enríquez et al., Energy Procedia, Volume 30, 2012, Pages 580-589 (http://www.sciencedirect.com/science/article/pii/S1876610212015846)
As a general issue there is right now an ongoing IEA-EBC (former IEA-ECBCS) Annex, number 58, which deals with parameter identification in buildings and related problems. I'm a participant in that Annex, so don't hesitate to contact me for further information.
Hope it helps,
Sunny Regards from Spain,
De: trnsys-users-bounces at cae.wisc.edu [mailto:trnsys-users-bounces at cae.wisc.edu] En nombre de angel.carrera at aiguasol.coop
Enviado el: miércoles, 26 de junio de 2013 22:15
Para: TRNSYS users mailing list at the Solar Energy Lab,UW-Madison
CC: Gossard, Didier
Asunto: Re: [TRNSYS-users] Parameter identification.
Just to add up our 2 cents from our experience I will throw a few comments to David words.Two points, one on type 56, and other on model calibration.
For model calibration,we believe that in a complex model usually a couple of around 10 parameters explain well over 90 % of the variance of the model output. So, we opt for a computationally cheap technique for identifying those parameters (usually Morris method, modified by Campologno and Saltelli) to detect those.
Once there, fitting thisd limited set is more or less simple; we feel comfortable with Wilmott sensitivity index instead of integrated values (we never worked with them, so we could be deeply wrong) as a cost function due to it accounts for temporary tendencies on model output. I have to say that we get modest statistical indices with this approach, but very good results when comparing integrated values.
About managing the parametrics and type 56,for v16, there is a document by Michael Kummert that explains how to edit the bui file with Trnopt. However, we decided last year to develop an application to manage big parametric sets, to ease dck and bui generation together with output management, wich is the real issue to make feasible the process. At the end, as a friend showed us, the bui file is a plain text file. Detailled radiation mode can be a limitation to this task.
To sum up, we now routinely run tens of thousands of simulations editing the building, the thing is what you do with the results. I think my colleague Damien Tavan, in charge oh this work, already uploaded our tool in some freeware web (at your own risk), I will ask him to send the link to the list if I am right. I also recommend the site by Marcus Jones as a source for running TRNSYS in parallel processors, that is an additional issue when diving in this field.
Sorry if I've been unable to focus the answer, there are a lot of little things in this issue. Hope it helps! Best
BlackBerry de movistar, allí donde estés está tu oficin@
From: David BRADLEY <d.bradley at tess-inc.com>
Sender: trnsys-users-bounces at cae.wisc.edu
Date: Wed, 26 Jun 2013 13:38:01
To: TRNSYS users mailing list at the Solar Energy Lab, UW-Madison<trnsys-users at cae.wisc.edu>
Reply-To: "TRNSYS users mailing list at the Solar Energy Lab,
UW-Madison" <trnsys-users at cae.wisc.edu>
Cc: Gossard, Didier<Didier.Gossard at saint-gobain.com>
Subject: Re: [TRNSYS-users] Parameter identification.
TRNSYS-users mailing list
TRNSYS-users at cae.wisc.edu
TRNSYS-users mailing list
TRNSYS-users at cae.wisc.edu
More information about the TRNSYS-users