I have a mathematical model for the transformer that needs parameters values(which are MANY) which is evaluation outcome curve fits with my experimental measurement test.
in order to find the optimum values I am doing MANY iterations(permutation for the parameter values) that would take not less than 5 years.
I guess I am in need to know about simulation optimization. but there are many optimization methods?
what should I study?
I suggest you to use the simplex algorithm to implement a least-square parameter identification. If you use matlab, you can use the fminsearch function.
Hi Youssef,
Does transformer stands for the electrical device or for the model structure (transformer = advanced RNN) ?
Hi Laurent Dewasme ,
Does transformer stands for the electrical device ?
Yes, https://en.wikipedia.org/wiki/Transformer
Hello Michael Short ,
It might help if you could provide some further details: is your underlying model linear in the parameters, or non-linear?
It is non-linear system.
What criteria are you using for quality of fit (e.g. sum squared errors, sum absolute errors, max absolute error etc)?
Well, I can tell you that the measurement curve has peaks and troughs and what I really care about is the peaks and troughs that come out of simulation to be identical or correlated as much as it could with the measurement ones. but I do not know which method of you mentioned above would be suitable for my case.
How many is 'many' parameters in the model: 10? 100? 100,000?
40*40*40*40*40*40=4.096e9
but still it depends on my measurement case
is it okay if I contacted you further?
Thank you in advance
Anne Richelle Thank you for you reply. I will read about it.
Ok. It seems like the number of parameters is huge...As Anne said, fminsearch in MATLAB is probably a good solution. If you have some a priori knowledge on the order of these parameters, you can also use fmincon to bound the parameter values and accelerate the optimization.
Best regards,
Laurent
Michael Short However the number of parameters seems awfully large for an electrical transformer (but not out of the question depending upon what you are doing)... have you tried simplifying the model at all, e.g. using a lumped-parameter simplification?
Well, it seems I am having a misconception regarding to parameters. and I will talk deeper and in details about my problem.
the model I have for transformer is represented in state space form.
dx/dt =[A]x(t) + [B] u(t) (A matrix dimention is 12 row *12 column and B matrix is 12*1)
y(t)= [C]x(t)+[D]u(t)(C matrix is 1*12(matrix C does not change) and D is zero)
so replacing/changing the matrix elements of A and B would change the phase and gain margin(Bode Diagram). on my simulation software I did, it would take a long time to find all phase and gain margin of all values of A and B (that has constrains) and afterward, it picks the most correlated to the measurement.
Is there any shortcut way to make a software that finds the most correlated matrix to my measurement with acceptable error based on few iterations(not 4e9 iterations that takes 4 years) to some values ?
Key words:
Risk assessment, PCA and multivariate analysis.
Go to tje library and have fun.
Michael Short EXACTLY, that is what I have figured out the day before yesterday. I should have learn least squares method(Guess-Newton), Jacobin Matrix...etc. and yes I have constraints regarding to some theories and another measurements I do.
Is your measurement data time-domain or frequency domain?
Frequency domain.
I would really like to thank you or reply sir.
Can you recommend me a book that has exercises, problems and solutions?
Please consider this review in advance:
- DOI: 10.1093/bioinformatics/bty736