How do I determine or better specify the horizon for present and past predictions? - pvlib

Question

I am having a conceptual problem with pvlib's predictions: The problem is that if I ask for "past predictions" then I do not know what the temporal horizon is for the prediction. For actual future predictions it is a little more obvious, naively I would just subtract the present time from the timestamp of the prediction returned, although if data requests (model dependent) are made only at hourly or 6-hourly intervals, then it seems like I would have to add that uncertainty to the horizon, so I am still unsure.
For past predictions, I just have no idea what the horizon is. How can this be determined?
This question applies to both pvlib-python's standard way to get data/ forecasts and I think it also applies to the special script https://github.com/wholmgren/get_nomads to get data for predictions further into the past.
Any help would be appreciated in understanding this situation.
To try to make this question more concrete I am including this bit of code taken from forecast_to_power.ipynb with the start and end times modified to be in the past:
# built-in python modules
import datetime
import inspect
import os
# scientific python add-ons
import numpy as np
import pandas as pd
# plotting stuff
# first line makes the plots appear in the notebook
%matplotlib inline
import matplotlib.pyplot as plt
import matplotlib as mpl
# finally, we import the pvlib library
from pvlib import solarposition,irradiance,atmosphere,pvsystem
from pvlib.forecast import GFS, NAM, NDFD, RAP, HRRR
# Choose a location.
# Tucson, AZ
latitude = 32.2
longitude = -110.9
tz = 'US/Mountain'
surface_tilt = 30
surface_azimuth = 180 # pvlib uses 0=North, 90=East, 180=South, 270=West convention
albedo = 0.2
# for this example, let's predict into the past:
start = pd.Timestamp(datetime.date.today(), tz=tz) - pd.Timedelta(days=14) # 14 days ago
end = start + pd.Timedelta(days=7) # 7 days from start
fm = GFS()
forecast_data = fm.get_processed_data(latitude, longitude, start, end)
forecast_data.head()
temp_air wind_speed ghi dni dhi total_clouds low_clouds mid_clouds high_clouds
2019-02-25 06:00:00-07:00 6.581512 1.791610 0.000000 0.000000 0.000000 33.0 0.0 0.0 33.0
2019-02-25 09:00:00-07:00 4.832214 0.567790 392.833659 668.164855 121.831040 0.0 0.0 0.0 0.0
2019-02-25 12:00:00-07:00 3.409973 0.860611 794.120954 910.658669 118.492918 0.0 0.0 0.0 0.0
2019-02-25 15:00:00-07:00 6.841797 0.942555 529.425232 515.727013 222.689391 22.0 0.0 0.0 22.0
2019-02-25 18:00:00-07:00 24.458038 0.466084 11.339769 0.000000 11.339769 52.0 0.0 0.0 52.0
What is the temporal horizon for this back prediction? Can I adjust it? If so how?

Answer 1

For actual future predictions it is a little more obvious, naively I would just subtract the present time from the timestamp of the prediction returned
Yes, this is correct for true predictions. For past predictions, you should define the horizon in a way that is consistent with your ability to make true predictions.
NCEP maintains a model status page that details the typical times at which the weather model data is available on its servers. Each model has a different delay between its initialization time and its forecast availability.
The Solar Forecast Arbiter definitions document might also help.