New release to use SmoothedConstantInterpolation

[olivierbonte]

Hi,

Thanks to the contribution of @SouthEndMusic (many thanks!) in #367, there is now a SmoothedConstantInterpolation method.
I was wondering when there would be a new release of DataInterpolations.jl including this method, which would be greatly appreciated.
(In the meantime, I'll resort to installing the dev version directly from github for using this).

Cheers,
Olivier

[ChrisRackauckas]

Thanks for the reminder. I started the registration.

[SouthEndMusic]

Nice, what's your usecase for SmoothedConstantInterpolation?

[olivierbonte]

The same as yours I think: trying to use the ODE solvers with hydrological models!
Constant interpolation makes the adaptive timestep solvers fail, while e.g. pchip does not respect the integral under the curve (where piecewise constant is the reference), so SmoothedConstantInterpolation seems like just what I need :)

[SouthEndMusic]

Cool, yes that's our use case too. We didn't have these problems with failing solvers using adaptive time stepping and ConstantInterpolation. The smoothing is mainly for performance. Did you try adding the times from your timeseries to tstops? I created the convenience function get_transition_ts for this, which just returns itp.t, except for SmoothedConstantInterpolation, where the transition times are slightly different.

[olivierbonte]

Thanks for the tip on tstops! I knew from literature (use of SUNDIALS in land surface modelling) that it could be a good idea to let the solver stop at the discontinuities, but didn't know this feature was present already in OrdinaryDiffEq.

Out of curiosity: do you now combine both tstops and SmoothedConstantInterpolation in your application? Or is using only SmoothedConstantInterpolation suficient?

[SouthEndMusic]

We do combine these yes, although SmoothedConstantInterpolation hasn't seen much application yet because our users are used to the block interpolation so the smoothing is $0$ by default. We have quite a lot of timeseries in our simulation, so as part of our initialization we collect the tstops (with some extra processing to also take periodicity into account) from all our timeseries: https://github.com/Deltares/Ribasim/blob/49a9763fcdaf599d09fd19585e80e1913b36075e/core/src/util.jl#L986-L1072