Added rounding of aa and b in _schumaker_qspline.py, see Issue #1311

[Antoine-0]

• Closes Spline fitting failure because of rounding #1311
• I have read the contributing guidelines
  I'm not sure how to do the following items :
• Tests added
• [ ] Updates entries to docs/sphinx/source/api.rst for API changes.
• Adds description and name entries in the appropriate "what's new" file in docs/sphinx/source/whatsnew for all changes. Includes link to the GitHub Issue with :issue:`num` or this Pull Request with :pull:`num`. Includes contributor name and/or GitHub username (link with :ghuser:`user`).
• Pull request is nearly complete and ready for detailed review.
• Maintainer: Appropriate GitHub Labels and Milestone are assigned to the Pull Request and linked Issue.

[cwhanse]

I'll take a look at the failing test. I can add the text for the whatsnew document, and take care of assigning labels. Don't worry about the API item.

[cwhanse]

I removed the data that was causing the test failure. Those data were made up to check consistency between Matlab and python implementations of pvlib.ivtools.sde._fit_sandia_cocontent. The test still has 2 of the original 3 made-up IV curves, which pass the test criteria. So I am not concerned about dropping part of the test data, the test still meets its intent. The removed data that are removed would still produce values for the 5 IV curve parameters, but they would not be close to what is produced by the Matlab code (e.g. 16% relative difference in one parameter).

I also split the EPS parameter into EPS_slope to determine when a slope is 0, and EPS_val to determine when a number is zero. The EPS_slope value is around 10^-6, the EPS_val value is single digit precision.

Why did the removed data fail the test? The single-curve fitting function is rather sensitive to its input data. Rounding small values to 0 changes the spline fit, which affects the cocontent integral, which affects a regression, etc. This sensitivity is why the diode model fitting in pvlib.ivtools.sdm.fit_pvsyst_sandia (and similar) retains only the shunt resistance from the single IV curve fits; the other four parameters aren't reliably estimated by the cocontent method.

[cwhanse]

any objections to merging this?

[Antoine-0]

no, this is all fine for me, thanks again for running the tests !

[wholmgren]

Without carefully engaging with the issue and only looking at the diff, I have two quick comments:

1. We need a 0.9.1 entry in the whatsnew.rst file here: https://github.com/pvlib/pvlib-python/blame/master/docs/sphinx/source/whatsnew.rst#L8 Edit - fine to do this in a follow up PR.
2. Would it be better to add a new test that asserts the correct behavior rather just remove a test? Edit - I read Cliff's comment above and it seems fine just drop the test. Sorry for writing too quickly.

[cwhanse]

will merge when CI is done

[cwhanse]

Thanks @Antoine-0

[Antoine-0]

I merely noticed an issue, thanks to you for updating the tests and the great work you put into pvlib !