Buildings Models with QSS Derivative Sensitivities

[DeadParrot]

Some Buildings library models have state variable derivatives, particularly those of self-observer variables, that have poor QSS derivative accuracy with noisy derivatives (distinct from any noise caused by numerical differentiation) that cause excessive requantizations and, sometimes, solution mis-tracking. This needs to be understood and addressed with QSS extensions and/or controls.

Analysis of this behavior revealed a few things:

• When run with a --dtMax small enough to essentially change QSS into a traditional uniform time stepping solver the solutions track correctly. This gives the clue that the problem is due to the mismatch between the variable states at step times and the between-step trajectory values propagated to derivative computations.
• The derivative displayed "yo-yo-ing" trajectories between requantizations that made it appear to be related to standard QSS's use of "stale" quantized trajectories in computing derivatives. An experimental QSS variant was built that propagates the higher order and current "continuous" trajectories to derivative computations to test this idea. The QSS variant did greatly improve the derivative tracking but after some simulation time some of these derivatives would become noisy and go off track, presumably due to the injection of higher order representations into the apparently sensitive self-observer feedback loop. The ideas in this QSS variant may be useful but it creates additional QSS order dependency issues to be dealt with and is not really "QSS" anymore.

Models showing this behavior include ASHRAE2006, TwoFloor_TwoZone, and SimpleHouse. The larger models are hard to diagnose with because the derivatives are built from deeply layered expressions that depend on some "opaque" internal FMU variables that are not exposed. The smaller SimpleHouse model is probably a better basis for diagnosing the behavior and will be used in the next phase of work on this issue.

[mwetter]

In SimpleHouse, what state shows the noisy derivatives?

[DeadParrot]

I have not yet tracked back the SimpleHouse behavior to the states that first mis-track like I did for TwoFloor_TwoZone (as in my Mar 25, 2021 email) but I can see the same effect of the QSS variant that does continuous trajectory propagation fixing mis-tracking for the rad.vol[*].dynBal.U variables.