Buildings.Examples.DualFanDualDuct.ClosedLoop (Dymola)

[DagBruck]

We (meaning the Dymola R&D team) looked at this problem.

Dassl stalls around time = 20000; very small steps are taken with order 1. At this point, there is no change of step size at all. There is really fast dynamics in (some of?) the junction volumes and it seems that Dassl keeps the step small to resolve these dynamics.

When simulating with Dassl we get fast oscillations in the junction volume medium energies. That the dynamics are fast is probably physical, but the oscillations seem not to be; the corresponding eigenvalues are real, negative and large. As the oscillations aren't physical we think that Dassl should be able to damp them, especially, in this case as it is running order 1. But Dassl fails in this, so we are stuck. Frankly speaking, we don't know why Dassl cannot damp the oscillations.

But note that some other solvers work well for this model, especially Sdirk34hw. I guess the unsatisfactory conclusion is that we have a selection of solvers because some integrators will not be great for certain problems.

[DagBruck]

Analysis from my colleague Erik Henningsson.

Dassl stalls around time=20000. Very small steps are taken with order 1. At this point, no change of step size at all. There are really fast dynamics in (some of?) the junction volumes. It seems like Dassl keeps the step small to resolve these dynamics.

When simulating with Dassl we get fast oscillations in the junction volume medium energies. That the dynamics are fast is probably physical, but the oscillations seem not to be; the corresponding eigenvalues are real, negative and large.

We don't know why Dassl cannot damp the oscillations. There may be

1. Some nonlinear dynamics in the model not covered by the linearization
2. Some unfortunate interaction between Dassl, its corrector iteration, and our special logic for accepting good enough solutions when the order is 1.

We have experimented with our "accepting-good-enough" logic, but it hasn't resulted in any significant improvements.

Note that our Godess solvers work well for this model. Especially Sdirk34hw. In that light, the situation is ok. Maybe a bit unfortunate is that we cannot give a priori motivations to why implicit Runge--Kutta should be better. Other, than: "they have better damping". But, as there aren't any highly-oscillatory modes in the linearization, Dassl's damping should be enough.