Let solver integrate flows

[SouthEndMusic]

Fixes #1431.

To do:

• Let the solver integrate edge flows
• Let the solver integrate basin forcings
• Let the solver integrate flows for allocation input
• Let the solver integrate realized user demand flows
• Update Jacobian sparsity and related test
• Docstrings etc. (maybe some parts of the documentation that now only mention the storages and the PID integral term as states)

[SouthEndMusic]

There seems to be something broken in adaptive timestepping. It looks like timesteps are allowed to be arbitrarily small (no matter the solver settings), which would explain why running the backwater model takes a ridiculous amount of time and this doesn't throw an error:

Edit: this has now been resolved. I think the state-specific tolerances in function Model and the internalnorm which only looks at the storage now aligns the solver with what we care about. The point below about the comment stil stands.

Ribasim/core/test/bmi_test.jl

Lines 14 to 15 in 79aa2df

	# cannot go back in time
	@test_throws ErrorException BMI.update_until(model, dt0 / 2.0)

It also looks like that comment is outdated, because you don't give a negative dt but I suspect a too small one.

[visr]

It also looks like that comment is outdated

It doesn't look outdated to me. First you update to t=dt0, then to t=dt2/0, which is before the current t. The SciML methods expect a dt argument, whereas BMI expects a t argument.

[SouthEndMusic]

If desired, we could apply different tolerances to different components of the states this way

[visr]

Looks good, though the "mean_flow" test perhaps needs some double checking to see if it is ok. A flow of 0.84 where it should be 1 seems quite a large difference, though if it is only for tiny dt and evens out properly over time it should be fine.

[evetion]

Can't comment on the solver changes/setup, but overall looks ok to me, some small comments.

FWIW I would be useful for others to have a nice PR description (which turns into a commit message after (squash) merging) at least detailing what changed in the internal API for other developers who didn't see these changes.

[evetion]

Also, this seems a net change in terms of lines added/removed, contrary to the original idea in #1431. Do we think this lowered our maintenance? How is the performance affected?

[SouthEndMusic]

So it turns out that this is pretty bad for performance, it basically negates all performance gain from #1457 as measured with the HWS model. So here's a wild idea: What if we set up a parallel ODE problem for the flow integration with an explicit solver?

[visr]

Too bad this didn't work out performance wise, good effort though.