BUG: netcdf4 conversion error in forecast.py

[CameronTStark]

Describe the bug
The method get_data() from the parent class ForecastModel has a handling NETCDF data. It produces the following error on multiple tests within test_forecast.py:

TypeError: <class 'cftime._cftime.DatetimeGregorian'> is not convertible to datetime

To Reproduce
Run the command locally:

pytest pvlib/tests/test_forecast.py --remote-data --pdb

This command will ensure the tests marked as remote_data are run and you'll be dropped into a pdb debugger.

Expected behavior
The formatted data would be received without error.

Screenshots
If applicable, add screenshots to help explain your problem.

Versions:

• pvlib.__version__: 0.7.1 master
• pandas.__version__: 1.0.3
• python: 3.7.4

Additional context
I'm unsure about the underlying origin but the disconnect seems to come through get_data() -> _netcdf2pandas() -> set_time() where num2date supplies a DatetimeGregorian that Pandas can't process.

It's notable that there is a suggestion on line 279 of forecast.py about the potential to move to xarray's implementation of NetCDF rather than using the _netcdf2pandas() method. This avenue likely has the potential to be easier than continuing a pvlib implementation.

[wholmgren]

Duplicate of #920?

[CameronTStark]

I don't think so but I can't say they're not related.

The error described in #920 is not the same as here nor does the solution proposed in #921 fix this issue.

[kandersolar]

FWIW I cannot reproduce the errors on my machine.

Maybe not relevant, but just in case: #900 handled the cftime change that added only_use_cftime_datetimes, but there was another recent change to add only_use_python_datetimes: https://github.com/Unidata/cftime/pull/139/files#diff-9225bec7bdd479a2dfbc187bb72953d1R259

[augustecolle]

Downgrading cftime from 1.1.1 to 1.0.4 solved this issue for me.

[CameronTStark]

That did it for me. Thanks, @augustecolle!

[wholmgren]

@CameronTStark do you still see this problem when using pvlib/pvlib-python/master? I think @kanderso-nrel's suggestion might have fixed it and #947 might be unnecessary. Sorry I got confused about which issue to reference.