Error in clearsky.reno in newer pandas and numpy versions

[kandersolar]

I know there are some known pandas issues already, but I think this one is new? This snippet raises an error for some combinations of pandas and numpy:

from pvanalytics.features.clearsky import reno
import pandas as pd

times = pd.date_range('2019-01-01 12:00', freq='1min', periods=3)
ghi = pd.Series([1,1,1], index=times)
reno(ghi, ghi)  # ValueError: clearsky requires regular time intervals of 15m or less

It fails with numpy==1.21.4 + pandas==1.3.0. It works with numpy==1.19.5 + pandas==1.2.5.

This is the issue:

In [7]: pd.Timedelta("1T") / np.timedelta64(1, "60s")
Out[7]: 60.0  # returns 1.0 on older versions

I did not do any digging for changes in either library that may have caused this, or whether the issue can be isolated to a change in just one of pandas or numpy.

[cwhanse]

This one is new. Probably safer to compute that integer number of minutes with either pandas or numpy but not both. Surprising to see the combination give the wrong answer.

[cwhanse]

Would that statement work in newer versions as

pd.Timedelta("1T") / np.timedelta64(1, "m")

Wondering if it's a change in how numpy handles the time unit string.

[kandersolar]

With the newer versions listed above pd.Timedelta("1T") / np.timedelta64(1, "m") gives 1.0, so perhaps that's a suitable fix. I tend to agree that sticking with all pandas is likely safer.