Nominal values of iteration variables in non-linear system solvers

[Lennart Ochel]

The solvers for linear, non-linear and mixed systems are working on wrong variable attributes, e.g. nominal values. This can be investigated using the following model:

   model problem10
     parameter Real xn = 1e5;
     parameter Real yn = 1e-5;
     Real x(start=1.5*xn, nominal=xn);
     Real y(start=1.5*yn, nominal=yn);
   equation
     0 = -(x/xn)*exp(-(x/xn)*(y/yn))-1;
     0 = (y/yn)*exp(-(x/xn)+(y/yn))-1;
   end problem10;

The issue is, that the attributes get copied for the solvers before the parameters are calculated. I already worked out a patch that fixes this problem. Unfortunately, there are models that need to solve non-linear systems (by defect) to evaluate parameters, see #1651. Because cyclic dependencies in parameter bindings is not allowed, that ticket should get fixed first.

[Martin Sjölund]

Note that cyclic dependencies in parameter equations is allowed in initial equations. Since NLS are allowed, it should be possible to create the nominal attributes on the fly however.

[Lennart Ochel]

Replying to sjoelund.se:

Note that cyclic dependencies in parameter equations is allowed in initial equations. Since NLS are allowed, it should be possible to create the nominal attributes on the fly however.

As far as I know it is not allowed to use parameters with fixed=false as, e.g., nominal attribute. Is that what you mean? Can you please given an example?

[Lennart Ochel]

BTW: this is only about non-linear systems.

[Martin Sjölund]

Replying to lochel:

As far as I know it is not allowed to use parameters with fixed=false as, e.g., nominal attribute. Is that what you mean? Can you please given an example?

No, it was just a general comment that parameters need to be able to solve NLS.

[Lennart Ochel]

Yes, but that is only true for parameters with fixed=false, right? This is about parameter with binding expressions and those have to be acyclic.

[Lennart Ochel]

fixed in r22519