Difference in Clearsky DNI Results Between PVLib and Solcast

[HrvojeR]

Dear friends,

I am using the PVLib library to calculate clearsky DNI values on an hourly basis for a specific geographic location. However, I have noticed significant differences between the results obtained using the PVLib method and the values provided by the Solcast API. To test the accuracy of PVLib methods, I used data from Solcast as input for PVLib. The codei s shown below:

import pandas as pd
from pvlib.location import Location

# Loading Solcast data
data = pd.read_excel('data.xlsx', parse_dates=['period_start'], index_col='period_start')

# Defining geographic location
latitude = _my location data_
longitude = _my location data_

location = Location(latitude, longitude, 'Europe/Zagreb')

# Solar position - data from Solcast
solar_position = pd.DataFrame(
    {
        'apparent_zenith': data['zenith'],
        'zenith': data['zenith'],
        'apparent_elevation': data['apparent_elevation']
    },
    index=data.index
)

# Calculating clearsky values using PVLib
clearsky = location.get_clearsky(
    times=data.index, 
    model='ineichen', 
    solar_position=solar_position, 
    pressure=data['pressure']
)

# Obtained Clearsky DNI values
print(clearsky['dni'])

For comparison, here are the clearsky DNI results obtained using PVLib and the values returned by the Solcast API:

PVLib Clearsky DNI:

pvlib clearsky_dni:
0
0
0
0
0
0
0
190.459769
504.166109
657.054597
727.709442
750.209492
727.709442
673.140539
556.541914
285.565017
0
0
0
0
0
0
0
0
solcast clearsky_dni:

0
0
0
0
0
0
2
382
683
795
839
849
837
785
666
397
13
0
0
0
0
0
0
0

The differences are evident, particularly during periods of high solar irradiance. My questions are:

1. What could be the reasons for these differences?
2. How can I adjust PVLib settings (e.g., model selection, additional parameters) to align the results more closely with those provided by Solcast?
3. Is it possible that Solcast applies a specific model or correction that PVLib does not implement?

Thank you in advance for your assistance.

Best regards,

[cwhanse]

It's difficult to comment when location, time, and SolCast data are omitted. There may be differences due to the time resolution and alignment: values from pvlib are instantaneous at the timestamp; SolCast may be an average or other statistic for the interval (I don't know). Also, the comparison is between two models one of which isn't public: https://www.solcast.com/irradiance-data-methodology

[mikofski]

In addition to input differences, please note there are several clear sky models. The default in pvlib is “Ineichen” but we don’t know what clear sky model Solcast uses. Please see pvlib clear sky documentation for a list of models implemented in pvlib, & note there may be other models.

[wholmgren]

Solcast uses REST2 + modifications. https://www.solcast.com/irradiance-data-methodology

Text below is from DNV review of Solcast historical data

[adriesse]

Are the extensive validations of REST2 valid after "proprietary configurations and improvements"??

[AdamRJensen]

Solcast is based on actual current aerosol data, whereas the pvlib function you use is based on climatological average aerosol data (high uncertainty).

I would recommend comparing against the free McClear clear sky data (this can be retrieved using the pvlib.iotools.get_cams function).

[HrvojeR]

Thank you all for your responses; they have been extremely helpful. Guided by the suggestion @AdamRJensen which we tested, we achieved the most accurate results so far, and we hope they will prove useful in practice. Once again, thank you all!
Results:
https://prnt.sc/qCKBmrQuVSMa