Off-by-one error in clear sky mid-month calculation for Linke turbidity

[josephmckinsey]

Description

On line 260 at

pvlib-python/pvlib/clearsky.py

Line 260 in 0781787

dayofyear = time_utc.dayofyear

, we see that _interpolate_turbidity uses time_utc.dayofyear which is 1-indexed. When the middle of the month is calculated, this uses 0-indexing.

Usually in other parts of the code, this is corrected by an offset of 1, but in this case, it is not.

Reproduction

For instance, this year, February 15th 12:00 PM was the midpoint of February. This is the 46th day of the year, so the index is 45.5 which is correctly calculated by _calendar_month_middles.

Expected Behavior

This can be seen for some midpoints which you would expect to be exactly what's in the HDF5 file, but are not. For example, we would expect $\frac{57}{20} = 2.85$ for these coordinates at this time:

>>> pvlib.clearsky.lookup_linke_turbidity(pd.DatetimeIndex([datetime.datetime(2024,2,15,12)]), 38, -93)
2024-02-15 12:00:00    2.861667

I imagine this is of very low priority, since these calculations only have the resolution of a single day, but I noticed it and thought I should make an issue.

• pvlib.__version__ == '0.11.1'
• pandas.__version__ == '2.2.3'
• Python 3.13.0

[cwhanse]

Maybe the issue is a lack of clarity in what to expect. What value should be used to represent the entire month in these cases?

February 2024: the middle of the month day is noon of Feb 15 (14.5 days before, 14.5 days remaining). The returned index is 45.5, which is what I would expect (31 + 14.5). But starting from day of year (46 for Feb 15, 2024) I can understand why seeing 45.5 appears to be off.

February 2023: the middle of the month is midnight on Feb 14. The returned index is 45. The day of year for that datetime is 46, but 1 second earlier would be 45.

import datetime
import pandas as pd
from pvlib import clearsky


dt = pd.DatetimeIndex([datetime.datetime(2024,2,15,12)])
print(dt.dayofyear)

middles = clearsky._calendar_month_middles(2024)
print(middles[2])

dt = pd.DatetimeIndex([datetime.datetime(2023,2,15,0)])
print(dt.dayofyear)
dt = pd.DatetimeIndex([datetime.datetime(2023,2,14,23,59,59)])
print(dt.dayofyear)

middles = clearsky._calendar_month_middles(2023)
print(middles[2])

The lookup and interpolation methods originated with pvlib code, so we aren't bound to follow a reference here and can choose how to do this.

[josephmckinsey]

I don't really have too much stake in this; I just thought it was weird. My own suggestion would be to use actual times instead of dayofyear calculations, so the interpolated value is a continuous function of datetime. You still need to get the middle of the month in epoch time or something like that.

I haven't tested the sensitivity of a final PV power output with respect to turbidity, so I'm not sure how much this matters, especially given how the input data is in uint8.