Atmos. Meas. Tech.

, 16, 4975–5007, 2023

https://doi.org/10.5194/amt-16-4975-2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Irradiance and cloud optical properties from

solar photovoltaic systems
James Barry1,2 , Stefanie Meilinger1 , Klaus Pfeilsticker2 , Anna Herman-Czezuch1 , Nicola Kimiaie1 ,
Christopher Schirrmeister1 , Rone Yousif1 , Tina Buchmann2 , Johannes Grabenstein2 , Hartwig Deneke3 ,
Jonas Witthuhn3 , Claudia Emde4 , Felix Gödde4 , Bernhard Mayer4 , Leonhard Scheck4,5 ,
Marion Schroedter-Homscheidt6 , Philipp Hofbauer7 , and Matthias Struck7
1 International Centre for Sustainable Development, Hochschule Bonn-Rhein-Sieg, Sankt Augustin, Germany
2 Instituteof Environmental Physics, Heidelberg University, Heidelberg, Germany
3 Leibniz Institute for Tropospheric Research, Leipzig, Germany
4 Meteorological Institute, Ludwig-Maximilians-Universität München, Munich, Germany
5 Hans Ertel Centre for Weather Research, Munich, Germany
6 German Aerospace Center (DLR), Institute of Networked Energy Systems, Oldenburg, Germany
7 egrid applications & consulting GmbH, Kempten, Germany

Correspondence: James Barry (james.barry@iup.uni-heidelberg.de)

Received: 31 December 2022 – Discussion started: 4 April 2023

Revised: 10 August 2023 – Accepted: 29 August 2023 – Published: 27 October 2023

Abstract. Solar photovoltaic power output is modulated by the mean bias error is 5.88 and 41.87 W m−2 under clear and
atmospheric aerosols and clouds and thus contains valuable cloudy skies, respectively.
information on the optical properties of the atmosphere. As During completely overcast periods the cloud optical
a ground-based data source with high spatiotemporal resolu- depth is extracted from photovoltaic power using a lookup
tion it has great potential to complement other ground-based table method based on a 1D radiative transfer simulation, and
solar irradiance measurements as well as those of weather the results are compared to both satellite retrievals and data
models and satellites, thus leading to an improved charac- from the Consortium for Small-scale Modelling (COSMO)
terisation of global horizontal irradiance. In this work sev- weather model. Potential applications of this approach for
eral algorithms are presented that can retrieve global tilted extracting cloud optical properties are discussed, as well as
and horizontal irradiance and atmospheric optical proper- certain limitations, such as the representation of 3D radiative
ties from solar photovoltaic data and/or pyranometer mea- effects that occur under broken-cloud conditions. In princi-
surements. The method is tested on data from two mea- ple this method could provide an unprecedented amount of
surement campaigns that took place in the Allgäu region ground-based data on both irradiance and optical properties
in Germany in autumn 2018 and summer 2019, and the re- of the atmosphere, as long as the required photovoltaic power
sults are compared with local pyranometer measurements as data are available and properly pre-screened to remove un-
well as satellite and weather model data. Using power data wanted artefacts in the signal. Possible solutions to this prob-
measured at 1 Hz and averaged to 1 min resolution along lem are discussed in the context of future work.
with a non-linear photovoltaic module temperature model,
global horizontal irradiance is extracted with a mean bias er-
ror compared to concurrent pyranometer measurements of
5.79 W m−2 (7.35 W m−2 ) under clear (cloudy) skies, aver- 1 Introduction
aged over the two campaigns, whereas for the retrieval using
coarser 15 min power data with a linear temperature model An accurate determination of incoming solar radiation at the
Earth’s surface is important not only for both climate and
weather research, but also for the stable operation of the

Published by Copernicus Publications on behalf of the European Geosciences Union.

4976 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

electricity grid in the future. In Germany alone there are to 7.93 W m−2 for the Copernicus Atmospheric Monitoring
2.6 million photovoltaic (PV) systems installed, with a nomi- Service (CAMS) radiation data (see Sect. 3.3).
nal power of 71 GWp (Holm, 2023) so that accurate forecasts The idea of using PV systems as radiation sensors has been
of solar PV power generation are indeed becoming indis- explored by several authors. In Engerer and Mills (2014),
pensable for cost-effective grid operation. In this context the Killinger et al. (2016), and Marion and Smith (2017), meth-
proliferation of PV systems provides a unique opportunity to ods are developed in order to use the output of one PV system
characterise global irradiance with unprecedented spatiotem- to predict that of another, which is in essence done by infer-
poral resolution, which would lead to improvements in both ring GHI from PV power measurements. In all three cases
weather and climate models. Solar panels can in this way be empirical models for the decomposition of irradiance into di-
seen as a dense network of sensors for atmospheric optical rect and diffuse components are used, and system parameters
properties. This new information could facilitate the devel- such as orientation and PV module efficiency are required
opment of highly resolved PV power forecasts and can play inputs. Engerer and Mills (2014) achieve a mean bias error
a role in improving climate models, in particular since the of 1.09 % for the PV power output under clear-sky condi-
highly variable nature of cloud cover as well as uncertainties tions, but the accuracy diminishes for partly cloudy skies, as
in cloud microphysics result in the greatest uncertainty in our expected; Killinger et al. (2016) achieve a mean bias error
understanding of the radiative forcing of the climate. between −3.9 % and −9.8 % for the GHI, depending on the
It has been shown by several authors (see for instance Ur- empirical model used for irradiance transposition, and Mar-
raca et al., 2018; Ohmura et al., 1998; Frank et al., 2018; ion and Smith (2017) achieve a mean bias error for the GHI
Zubler et al., 2011) that the estimates of global horizon- of within ±1.5 % using south-facing PV modules at 10, 25,
tal irradiance (GHI) from both the global ECMWF (ERA5) and 40◦ tilt angles. A similar approach is taken in Elsinga et
and the regional (COSMO-REA6) numerical weather pre- al. (2017), in this case using a single-diode PV model and
diction (NWP) model reanalyses deviate from ground-based a different decomposition model. In Halilovic et al. (2019)
measurements. In Urraca et al. (2018), comparisons are the authors replaced the original iterative approach used in
made with pyranometer measurements from the Baseline Killinger et al. (2016) by an analytical method, to minimise
Surface Radiation Network (BSRN) (Ohmura et al., 1998) computational cost, and achieved a mean bias error of be-
and from a dense network of pyranometers operated by Eu- tween 0.1 % and 2.1 % for the resulting GHI, using data from
ropean meteorological services. In general the model re- silicon reference cell measurements at different tilt and az-
analyses overestimate the irradiance under cloudy skies, imuth angles.
which is largely due to an underestimation of cloud opti- In Nespoli and Medici (2017) a different method is intro-
cal depth (COD). The mean positive bias of ERA5 daily duced, in this case without the need for system-specific in-
mean irradiance is +4.05 W m−2 (3.47 %) over Europe and formation such as orientation or nominal power. A similar
+4.54 W m−2 (2.92 %) worldwide. On the other hand, the re- approach is taken in Saint-Drenan (2015) and Saint-Drenan
gional COSMO-REA6 dataset underestimates GHI on clear- et al. (2015), where system parameters are estimated by sta-
sky days, with a mean bias of −5.29 W m−2 (−3.22 %), tistical methods. In addition, Scolari et al. (2018), Laudani
which can be attributed to the use of an aerosol climatology et al. (2016), Carrasco et al. (2014), and Abe et al. (2020)
with a too large aerosol optical depth (AOD), as discussed have also described the inference of solar irradiance from PV
in Frank et al. (2018). Although the COSMO-D2 data use current and voltage measurements using an equivalent-circuit
a different aerosol scheme, these negative biases in the GHI model. In this case greater accuracy is achievable, provided
are still present, especially in summer (Zubler et al., 2011). the module temperature is also measured.
Satellite datasets perform a lot better, with data from the Sur- This work builds upon the proof-of-concept study pre-
face solAr RAdiation Heliosat (SARAH) showing a mean sented in Buchmann (2018) (for clear-sky days only); how-
bias of only +0.86 W m−2 in the daily mean GHI (compared ever it is unique in that empirical models for the separation of
to +4.22 W m−2 from ERA5) over Europe (Urraca et al., radiation components are avoided – rather an explicit simu-
2018). Interestingly SARAH overestimates in most cases, lation of the diffuse radiance distribution is performed using
with only some stations showing a negative bias related to libRadtran (Mayer and Kylling, 2005; Emde et al., 2016). Al-
snow detection. Overall the satellite measurements display though this is computationally more intensive, it has several
a smaller absolute error than reanalysis products. The posi- advantages over the usual approach (see for instance Perez et
tive bias of the GHI from satellite retrievals is confirmed by al., 1992): by using a state-of-the-art radiative transfer code
Yang and Bright (2020): their comprehensive global evalu- one can more accurately model the clear-sky irradiance, es-
ation of hourly satellite irradiance data reveals a mean bias pecially for larger solar zenith angles, and one can explic-
error1 of 4.67 W m−2 for hourly SARAH-2 irradiance com- itly take into account information on aerosol load or ground
pared to the nine BSRN stations over Europe (excluding albedo from freely available datasets. In addition it is pos-
the Austrian station Sonnblick at 3100 m altitude), compared sible to include information on the state of the atmosphere
from weather models, which is particularly relevant for in-
1 Calculated using Table 3 in Yang and Bright (2020). cluding the effects of precipitable water on incoming irradi-

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4977

ance. The radiative transfer solvers DISORT (DIScrete Ordi- plane-parallel atmosphere with 1D radiative transfer, which
nate Radiative Transfer) (Stamnes et al., 1988; Buras et al., leads to a bias in the extraction of cloud optical properties, in
2011) and MYSTIC (Monte Carlo code for the physically particular under broken-cloud conditions. By neglecting 3D
correct tracing of photons in cloudy atmospheres) (Mayer, effects, the horizontal transport of photons is not considered,
2009) are used for forward model calibration as well as which however plays an important role in real-life situations.
for inferring atmospheric optical properties and GHI from These 3D effects can for example lead to an enhancement of
ground-based irradiance measurements and PV power data. solar irradiance (Schade et al., 2007) so that the GHI exceeds
In order for a PV-based determination of solar irradiance the clear-sky irradiance due to reflected light from the edges
to viably complement the global coverage of state-of-the- of clouds. The inherent 4D variability in clouds also compli-
art satellite measurements, a mean bias error of the order of cates the comparison of ground-based and satellite retrievals
5 W m−2 would be desirable (see the discussion on CAMS of cloud properties, since one compares the time average of
and other satellite-based products above). This level of accu- a point measurement with a spatially averaged quantity.
racy also corresponds to the target accuracy for global radia- The goal of this work is to demonstrate that PV systems
tion measurements from the BSRN (McArthur, 2005). How- can indeed be used as ground-based sensors for GHI as well
ever, even if this is not achieved, ground-based irradiance as to infer the optical properties of the atmosphere, in partic-
measurements and retrievals can be seen as complementary ular the COD. The first results are presented from two mea-
since they have the added advantage of superior spatiotem- surement campaigns carried out in autumn 2018 and sum-
poral resolution. The first step to achieve this is to accurately mer 2019 in the Allgäu region in southern Germany, as part
model the generated power as a function of system-specific of the research project entitled “Development of innovative
parameters, such as the array’s elevation and azimuth angle, satellite-based methods for improved PV yield prediction on
conversion efficiency, and temperature dependence, and then different time scales for distribution grid level applications”
extract those parameters from measured power data using a (MetPVNet) (Meilinger et al., 2021b, a). In Sect. 2 the for-
fitting procedure. In order to remove any biases related to ward model and its calibration are discussed, and the inver-
atmospheric conditions, it makes sense to first calibrate the sion methods are outlined in detail. Section 3 provides a de-
systems under clear skies. Once this has been done to suffi- tailed description of the data from the measurement cam-
cient accuracy one can use measured PV power to infer at- paigns. The results are presented in Sect. 4, with a focus on
mospheric optical parameters such as aerosol or cloud optical both tilted and horizontal irradiance as well as cloud optical
depth under different sky conditions, enabling the inference depth under overcast skies, and a summary and conclusions
of GHI as well as in some cases of direct and diffuse irradi- are given in Sect. 5. Further details of the PV modelling as-
ance components. pects and radiative transfer simulation can be found in Ap-
The more parameters used to model the PV power, the pendix A.
greater the uncertainty in the retrieved irradiance. For this
reason it is of course desirable to obtain as much a priori
metadata about the PV systems as possible, such as datasheet 2 Photovoltaic power model: calibration and inversion
parameters and array orientation. However, this information
is not always readily available, especially when considering In order to infer local atmospheric optical properties from
a large number of PV systems over a wide area. In that sense, measured PV data, accurate modelling of both atmospheric
there will always be a trade-off between quantity and qual- radiative transfer and PV power generation is required. In
ity of the data, which then plays itself out in the accuracy this section both the PV model and the libRadtran radiative
of the retrieved irradiance. The advantage of PV systems or transfer model is described, along with the calibration and
any ground-based devices is that one can achieve a much inversion procedure.
higher spatiotemporal resolution compared to satellite data
or weather models, which thus allows one to study high- 2.1 Forward model: from atmospheric properties to
frequency fluctuations in global irradiance. photovoltaic power
In the European context, irradiance variability is domi-
nated by the optical properties of clouds and less by those The power generated by a solar PV module depends primar-
of aerosols. Ground-based COD retrievals using broadband ily on incoming short-wave solar irradiance and module tem-
measurements from pyranometers have been carried out in perature, both of which depend on atmospheric conditions.
several studies (see for example Leontyeva and Stamnes, Once this dependence is properly described in a model, PV
1994; Boers, 1997; Boers et al., 1999; Deneke, 2002). In- power and/or current measurements can be used to infer the
deed, the transmission of irradiance through a cloud is the irradiance in the plane of the array, taking into account the
most sensitive to its optical depth and less sensitive to droplet geometry of the system, i.e. the elevation and azimuth angle
radius, single-scattering albedo, or asymmetry factor (Leon- of the solar panels. After extracting this global tilted irradi-
tyeva and Stamnes, 1994). In most previous studies the ance (GTI) from PV data, one can go on to infer atmospheric
clouds are assumed to be horizontally homogeneous in a optical properties such as cloud optical depth and global hor-

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4978 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

izontal irradiance by further inverting the radiative transfer an order of magnitude greater (roughly −0.4 % K−1 ) so that
model. this simple linear relationship breaks down when consider-
The most physically correct method of modelling the ing the PV power. In this work a parameterised power model
power output of a PV plant is with an equivalent-circuit is used (see Buchmann, 2018; Skoplaki and Palyvos, 2009;
model that captures the properties of semiconductors, such Dows and Gough, 1995), with AC PV power described as2
as the two-diode model (see for instance Mertens, 2014). In 
6 6
this way the temperature dependence of the current and volt- P AC,mod ' Gtot,PV,τ b1 + b2 Gtot,SW,τ
age is explicitly defined according to the Shockley equation 
(Shockley, 1949). A drawback of such models is their com- + b3 T ambient + b4 v wind + b5 T sky , (1)
putational complexity and reliance on parameters found on
module datasheets, which are in the most general case not in the case of the linear temperature model defined in
always available. There are however several parameterised Eq. (A3) (TamizhMani et al., 2003) or as
models in the literature that attempt to reduce the power gen-
6
eration equation to a simple relation among incoming plane- 6
 Gtot,SW,τ
of-array irradiance, module area, and temperature-dependent P AC,mod ' Gtot,PV,τ b10 +
b20 + b40 v wind
efficiency, with the latter described as a function of am- 
bient conditions. Several such models exist (see Skoplaki + b30 T ambient + b50 T sky , (2)
and Palyvos, 2009, for a review), with some of the most
popular being that of the PV Performance Modeling Col- in the case of the non-linear temperature model defined in
laborative (https://pvpmc.sandia.gov/, last access: 17 Octo- Eq. (A4) (Faiman, 2008; Barry et al., 2020). This means that
ber 2023) from Sandia National Laboratories (King et al., the modelled AC power P AC,mod is a non-linear function of
2004, 2007) or the Huld model used in the online PVGIS tool 6
plane-of-array irradiance Gtot,PV,τ , together with the effects
(Huld et al., 2011, https://re.jrc.ec.europa.eu/pvgis.html, last of ambient temperature T ambient , wind speed v wind , and sky
access: 17 October 2023). Since the goal here is an inversion, temperature T sky that influence module temperature and thus
the choice of model depends on the availability of measured efficiency. Note that the subscript PV for the tilted irradiance
data: in this work and in the context of the AC power data 6
Gtot,PV,τ refers to the fact that only the relevant wavelength
from the MetPVNet campaign, a simplified parametric power (in this case 300 to about 1200 nm for silicon PV modules)
model is employed. The model is described here briefly, and is considered, and the subscript τ indicates that transmission
more details are given in Appendix A. through the glass surface of the PV panels has been taken into
In order to correctly capture the effects of the variable so- account with an optical model. Further details of the model
lar spectrum one also needs to take into account the spectral employed here are given in Sects. A1 and A2 in Appendix A,
response of the PV technology in question (Alonso-Abella et and the refractive index n of the glass covering is one of the
al., 2014), which in the case of an equivalent-circuit model parameters varied in the optimisation procedure. The sub-
can then be included in the calculation of the photocurrent script SW refers to all incoming short-wave photons – the
(see for instance the libRadtran-based spectral PV model dependence of the spectral mismatch between the PV and
described in Herman-Czezuch et al., 2022). In the case of SW irradiance bands on atmospheric water vapour and other
parametrised PV power models, this so-called “spectral mis- factors is discussed in Sect.
match”, i.e. the difference between the entire spectrum of
2.3.
The parameters bi bi0 and (i = 1. . .5) in Eqs. (1) and (2)
incoming radiation and the range utilised by a certain PV depend on nominal power, efficiency, the temperature coef-
module, is usually simply absorbed into the PV model pa- ficient for power, and the temperature model parameters and
rameters, leading to a site-specific bias that may not take into are discussed in more detail in Appendix A, which includes
account variations in the spectrum from local atmospheric a definition of all parameters listed in Table 2. In practice
conditions. By using libRadtran for calibration and inversion the module temperature can be either measured or modelled,
along with information on the state of the atmosphere from depending on the availability of measurements and/or meteo-
weather models, one can take these variations into account in rological data. Within the PV power models described above,
the radiative transfer (RT) simulation and subsequent inver- the PV module temperature is a static quantity; i.e. the heat
sion, as discussed in Sect. 2.2 below. In particular the water capacity (C) of the PV system is not taken into account.
vapour column and aerosol optical depth at each site need to However, when dealing with high-frequency measurements
be taken into account (see Sect. A3 in Appendix A). of PV power, it is necessary to employ a dynamic tempera-
It can be shown using the diode model (see for instance ture model, as discussed in Barry et al. (2020). The charac-
Sauer, 1994; Abe et al., 2020) that the maximum power point teristic time constant (= C/J , with J being the net thermal
(MPP) current generated by a PV module is linearly depen- energy flux of the PV modules) of typically 10 min means
dent on the incident irradiance and only very weakly depen-
dent on temperature. However, the dependence of MPP volt- 2 The inverter efficiency is included in the parameter s; see
age on temperature (which itself is a function of irradiance) is Sect. A in Appendix A.

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4979

that the large fluctuations in irradiance do not translate di- (OPAC) species library with the option “continental average”
rectly to module temperature variations; i.e. the temperature (see Table 3 in Hess et al., 1998).
response is smoothed out. For simplicity the dynamic tem- In order to speed up the simulation the code is parallelised
perature model is not included in the present study, since to run on multiple processors: the simulation times are di-
most of the systems had power data collected in 15 min res- vided into batches, and libRadtran is then called multiple
olution. times as a subprocess from Python. In this way the clear-sky
simulation takes approximately 1 s per time step (8 s on an 8-
2.2 Model calibration under clear-sky conditions core machine), using a diffuse radiance field of 5◦ resolution
in both elevation and azimuth angle, atmosphere files mod-
In order to infer the irradiance in the plane of array (GTI) ified from COSMO data, and modified aerosol inputs from
from measured PV power or current, the PV model parame- AERONET and OPAC.
ters need to be determined, either from datasheets or with a
forward model calibration. This is accomplished using data 2.2.2 Non-linear optimisation for PV system
under clear-sky conditions, together with an accurate simu- parameters
lation of the irradiance, followed by a multi-parameter opti-
misation to find the parameter values. This section describes The simplified parametric model described above can be
the technical details of the clear-sky simulation and the rel- written as
evant atmospheric input parameters used. Figure 1 displays
this procedure graphically, and further explanations are given P AC,mod ≡ y = F(x, h) (3)
in the following sections.
for the forward model F described by Eq. (A1) and state
2.2.1 Radiative transfer simulation with libRadtran space defined by (see Table 2)

6
The clear-sky simulation of tilted irradiance Gtot,PV,τ is per- x ≡ (θ, φ, n, s, ζ, ui ) (4)
formed with the freely available libRadtran software package
so that the calibration procedure is effectively a non-linear,
(Mayer and Kylling, 2005; Emde et al., 2016), with the input
multi-parameter optimisation problem with eight (for the
parameters shown in Table 1 and the wavelength range from
non-linear temperature model)3 or nine (for the linear tem-
300 to 1200 nm for silicon PV applications. The correspond-
6 perature model) unknowns. As shown in Table 3, the param-
ing broadband simulation (Gtot,SW,τ ) is also performed, as eter space h in Eq. (3) contains the irradiance proxy from the
an input to the temperature model. Spectral integration is car- libRadtran simulation as well as temperature and wind speed
ried out using the Kato parameterisation in order to simplify data from either the COSMO model or the measurements,
the effects of water vapour absorption by using the so-called which are interpolated to 15 min resolution. In addition the
correlated-k approximation (Kato et al., 1999). The DISORT measured sky temperature (see Sect. 3.1) is used. This in-
solver allows for an explicit calculation of the diffuse radi- version problem can be solved with the methods detailed in
ance distribution on a predefined lattice of elevation and az- Rodgers (2000). In this case the Levenberg–Marquardt algo-
imuth angles, and the pseudospherical approximation is em- rithm is used, with the Jacobian matrix calculated explicitly
ployed so that only radiative transfer calculations at solar at each iteration.
zenith angles (SZAs) of up to 80◦ can be reliably performed. If one varies all parameters in x it quickly becomes appar-
The Python package PyEphem (Rhodes, 2022) is used to ent that there are not enough degrees of freedom in the sig-
accurately determine the sun position for the correspond- nal to uniquely determine a solution with the Bayesian for-
ing latitude, longitude, and time coordinates. Consortium for malism, since several parameters are highly correlated with
Small-scale Modelling (COSMO) model data (see Sect. 3.2) each other, for instance the inclination angle θ with the scal-
are interpolated by the package cosmomystic (see Barry ing factor s or the orientation angle φ with the tempera-
et al., 2023a) in both time and space in order to create atmo- ture model parameters or the coefficient ζ (see the discus-
sphere profile files suitable as input for libRadtran, in 15 min sion in Sect. 4.1). It is thus expedient to extract the tempera-
resolution. In this way variations in water vapour and other ture model parameters separately using the measured module
atmospheric trace gases are taken into account, and the atmo- temperature for different PV system configurations (if avail-
spheric layers are cut off at the appropriate altitude of each able) and then fix those parameters in the optimisation proce-
site. Concurrent measurements by an AErosol RObotic NET- dure. In Barry et al. (2020) a dynamic model was developed
work (AERONET) sun photometer (Holben et al., 1998) are by fitting the measured and modelled module temperatures
used to extract the Ångström exponent α and turbidity co- using three different PV systems with different mountings.
efficient τ a,1 with the aeronetmystic (see software sup-
plement) software package. Other aerosol optical properties 3 In the non-linear Faiman model there are less temperature pa-
such as the single-scattering albedo and asymmetry factor rameters as the ambient and sky temperatures are not independent,
are taken from the Optical Properties of Aerosols and Clouds as is the case in the linear model; cf. Eqs. (A3) and (A4).

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4980 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 1. Schematic diagram showing the different steps of the calibration procedure, with input data sources in green, model steps and
algorithms in blue, simulated parameters in orange, and system parameters (see Table 2) in dark red. Note that in this case only clear-sky
days or time periods are considered.

Table 1. Model parameters for the libRadtran simulation of clear-sky days, including information on their source.

Parameter Symbol Source

Latitude, longitude, altitude, time ϕ, ϑ, z0 , t Set by PV data
Solar zenith angle, solar azimuth angle θ 0, φ0 Calculated with PyEphem(Rhodes, 2022)
Temperature profile T (z) COSMO (Baldauf et al., 2011)
Pressure profile p(z) COSMO
Water vapour [ H 2 O ](z) COSMO
Ozone [O3 ](z) US standard atmosphere
Albedo ρ Constant (0.2)
Ångström turbidity coefficient τ a,1 AERONET (Holben et al., 1998)
Ångström exponent α AERONET
Other aerosol optical properties – OPAC continental average (Hess et al., 1998)

These results (for the static model case) are used in the over- method employed depends on the prevailing weather condi-
all optimisation, where appropriate. tions, specifically on the amount (and type) of cloud cover.
The calibration algorithm is designed to allow certain pa- In a nutshell, using a 1D DISORT-based method, one can
rameters to be fixed if they are known, whereas unknown pa- use GTI to extract AOD or COD during clear or completely
rameters are varied with a given a priori error, which in turn overcast periods, respectively, which thus allows the determi-
affects the parameter retrieval error and thus propagates into nation of the direct and diffuse irradiance components. Un-
the uncertainty in the inferred irradiance. der broken-cloud conditions, a 3D MYSTIC-based method
allows one to determine the GHI from GTI directly. The DIS-
2.3 Model inversion under all sky conditions ORT and MYSTIC lookup tables (LUTs) are provided in an
open data repository (Barry et al., 2023b).
The calibrated PV systems can now be used as sensors to
extract information about the state of the atmosphere. This 2.3.1 Global tilted irradiance from PV model inversion
section describes the different methods used to infer both ir-
radiance and atmospheric optical properties from PV power Once the PV system has been calibrated under clear-sky con-
data, as summarised in the schematic diagram in Fig. 2. The ditions, the system parameters can be fixed, and the mea-

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4981

Table 2. List of PV model parameters in x. In the calibration procedure, the parameters in x known to a certain degree (from datasheets or
other sources) of accuracy are fixed, whereas all others are varied.

Parameter (x) Symbol Source (if available)

Tilt angle θ laser scanning or theodolite
Azimuth angle φ laser scanning or theodolite
Glass refractive index n optimisation
Scaling factor s optimisation
Temperature coefficient ζ datasheet or optimisation
Temperature model parameters ui (i = 0, 1, 2, 3) optimisation and/or model (Barry et al., 2020)

Table 3. List of additional inputs in h used for calibration on clear-sky days. The subscripts PV and SW refer to the different wavelength
bands used for integration; see the discussion in Sect. 2.2.1.

Parameter (h) Symbol Source

6
Direct tilted irradiance Gdir,PV(SW) libRadtran simulation (see Table 1)
6
Diffuse tilted irradiance Gdiff,PV(SW) libRadtran simulation (see Table 1)
2 m ambient temperature T ambient COSMO or measured
Wind speed at 10 m v wind COSMO or measured
↓
Long-wave downwelling irradiance GLW Measured

sured PV power can be used to infer the global tilted irra- tral mismatch, it is a good first approximation, which will
6 be improved upon in future work (see also Rivera Aguilar
diance (GTI; also denoted as Gtot,SW ) under all sky condi-
tions. The temperature model makes use of broadband irra- and Reise, 2020, for an alternative method). In addition one
diance (see Eqs. 1 and 2), whereas the PV power model uses could modify this algorithm to include operational satellite
only the relevant spectral range of silicon PV modules (300– retrievals of atmospheric parameters such as ozone concen-
1200 nm) so that the spectral mismatch between the light tration, if required.
6
converted to electricity (Gtot,PV,τ ) and the entire short-wave Once the spectral mismatch factor ξ spec,GTI has been cal-
spectrum needs to be taken into account when inverting the culated, the next step is to extract the plane-of-array irra-
model chain. diance from the PV power, which in the case of the mod-
6 6
The ratio of PV-relevant (Gtot,PV ) to broadband (Gtot,SW ) els given in Eqs. (1) and (2) is simply the solution to the
6
tilted irradiance is a function of the system geometry, time quadratic equations in Gtot,SW,τ ; i.e.
of day, and local atmospheric conditions, with the largest  6 2
contributing factor being the precipitable water in the at- ξ spec,GTI b2 Gtot,SW,τ
mosphere. In order to take this into account, the libRadtran

clear-sky irradiance simulations (see Sect. 2.2.1) are used to + b1 + b3 T ambient + b4 v wind + b5 T sky
characterise the ratio 6
ξ spec,GTI Gtot,SW,τ − P AC,meas = 0 (6)
6
Gtot,PV,τ for the linear temperature model, and
ξ spec,GTI ≡ 6
= f (2, [H 2 O] , τ a ) (5)
Gtot,SW,τ ξ spec,GTI  6 2
Gtot,SW,τ
as a function of incident angle 2, precipitable water H 2 O, (b2 + b4 v wind )
and aerosol optical depth τ a , for each station and measure- + b1 + b3 T ambient + b5 T sky


ment campaign. In this way the available information on
the water vapour column and aerosol extinction from the 6
ξ spec,GTI Gtot,SW,τ − P AC,meas = 0 (7)
COSMO model and AERONET can be taken into account in
the PV model inversion. Details are given in Sect. A3 in Ap- for the non-linear temperature model. These equations can
pendix A. The fitting function could in principle be extended be solved with the quadratic formula, using the calibrated pa-
to include ozone column abundance, which is however not 0
rameters b1,2,3,4,5 (b1,2,3,4,5 ) as defined in Eq. (A5) (Eq. A6)
included here, since this information is not available from for the linear (non-linear) temperature model; the available
the COSMO model data. Note that although this method data for T ambient , v wind , and T sky ; and the spectral mismatch
does not take into account the effect of clouds on the spec- factor for GTI defined in Eq. (5). Note that the inverted

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4982 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 2. Schematic diagram showing the different steps of the inversion procedure, with input data sources in green, model steps and
algorithms in blue, simulated and retrieved parameters in orange, and system parameters in dark red. Note that in this case all atmospheric
conditions (all sky) are considered.

6
Gtot,SW,τ is the irradiance impinging upon the PV module for each time step, allowing the data to be separated into
under the glass covering so that the optical model has not clear, overcast, and broken-cloud time periods. The clearness
been inverted yet. In order to compare this quantity with index is then used to estimate the cloud fraction, which is dis-
pyranometers, the transmission of light through the glass cussed in more detail in Sect. 2.3.3.
τ PV,rel (2) (also known as the “incidence angle modifier”; On clear days (or during clear time periods) the aerosol
see Eq. A9) must be taken into account so that the final GTI optical depth (AOD) can be inferred, whereas under cloudy
is given by (see also Eq. A14) conditions the cloud optical depth (COD) can be found, de-
pending on the degree of cloud cover. In this work the extrac-
6 tion of COD using a DISORT-based LUT under completely
6 Gtot,SW,τ
Gtot,SW = . (8) overcast skies is examined in more detail in Sect. 2.3.3.
τ PV,rel,eff
An in-depth analysis of aerosol optical properties will be
For the direct extraction of GTI an empirical formulation is carried out in future work. For broken-cloud conditions, a
used to find the effective incidence angle for the diffuse com- MYSTIC-based LUT is used to infer the global horizontal
ponent, whereas for the inversion onto optical properties the irradiance from tilted irradiance measurements, as discussed
refractive index is explicitly taken into account within the ra- in Sect. 2.3.5 (see Chap. 9 of Meilinger et al., 2021b).
diative transfer simulation. More details are given in Sect. A2
2.3.3 Cloud optical depth with DISORT lookup table
in Appendix A.
Cloud optical properties are functions of microphysical prop-
2.3.2 Clearness index and irradiance variability erties such as the droplet size distribution, droplet number
classification concentration, and thermodynamic phase. For water clouds,
the absorption and scattering of solar irradiance can be effi-
Using the global tilted irradiance extracted from the mea-
ciently characterised (Hu and Stamnes, 1993) by the effective
sured PV power data, different methods are used in order to
radius reff and cloud liquid water content (LWC), which can
extract atmospheric optical properties and global horizontal
be related to cloud optical depth (COD, τc ) through
irradiance, depending on the prevailing weather conditions.
By combining the inverted tilted irradiance with the corre- 3 LWP
τc = , (10)
sponding clear-sky curve one can calculate a clearness index 2 reff ρH2 O
6 where the liquid water path (LWP) is the integral of the LWC
Gtot,SW,τ,inv
ki = (9) across the height of the cloud. The derivation of this equa-
6
Gtot,SW,τ,clear tion (see for instance Petty, 2006) assumes large Mie extinc-

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4983

tion, which is justified since clouds appear to be (mostly) Table 4. Cloud parameters for the DISORT simulation of a conti-
white in the solar spectrum. Although both τc and reff can nental stratus cloud (Hess et al., 1998).
be accurately retrieved from spectral measurements of re-
flected radiation, the transmission of light through clouds is Cloud parameter Value
mostly sensitive to the cloud optical depth. This is due to the Liquid water content (LWC) 0.28 g m−3
fact that changes in transmission due to variations in single- Effective radius (reff ) 7.33 µm
scattering albedo and the asymmetry factor (which depend Cloud height (h) 1–2 km
on reff ) are small compared to those due to changes in opti-
cal depth (Leontyeva and Stamnes, 1994). For illustration, in
the two-stream approximation for conservative scattering (no olution of 10◦ in both azimuth and elevation angles.4 The
absorption), the transmittance T can be shown to be (see for LUT is then used to find the COD by first calculating the
instance Petty, 2006) plane-of-array irradiance for the corresponding PV system
or pyranometer orientation (see Sect. 2.3.1) and then inter-
1 polating the COD in time to match the resolution of the mea-
T = , (11) sured data. For this purpose the original 1 Hz pyranometer
1 + (1 − g) τc
and PV data are smoothed to 1 min resolution, whereas the
low-frequency PV data are kept at 15 min resolution (see
where g is the asymmetry factor. For liquid water clouds, Sect. 3.1).
scattering is mostly in the forward direction, with g ' 0.85, In order to determine the exact time points at which a cloud
whereas for ice clouds g ' 0.7. In both cases the variations is above the sensor, the cloud fraction is determined by cre-
in g are small so that τc is the primary factor influencing ating a mask based on the clearness index in Eq. (9) using a
T . The hyperbolic dependency of T on τc means that the threshold of 0.8 and overshoot limit of 1.1; i.e.,
transmission curve is rather steep at small optical depths but

flattens out for COD & 15. This has implications for the ac-  1 if k i ≤ 0.8
curacy of ground-based retrievals, as discussed in detail in cf = 0 if 0.8 < k i ≤ 1.1 (12)
Sect. 4.4. It must also be noted that in the algorithm described 
nan if k i > 1.1.
in Sect. 2.3.1, spectral variations in cloud optical properties
are not taken into account. In practice this means that vari- This binary cloud mask (clear = 0, cloudy = 1) is then also
ations in the single-scattering albedo at higher wavelengths smoothed with a moving average function over 60 min in or-
around 1 µm (silicon PV modules are still sensitive to wave- der to create an estimate of the cloud fraction (hcf i60 ). Vary-
lengths up to 1.2 µm) may be unaccounted for. ing the thresholds in Eq. (12) shows that the cloud fraction
In this work a lookup table for the optical depth of a typical computed in this way depends less on the exact threshold
stratus cloud is constructed using DISORT in 15 min time in- used but more on the window size chosen for the moving av-
tervals, under the assumption of a pseudospherical or plane- erage. Indeed, comparison with concurrent cloud camera re-
parallel atmosphere with horizontally homogeneous liquid trievals shows that 60 min is a reasonable averaging time to
water, clouds, and a completely cloudy sky. This means that use, when averaging a cloud mask created with data at 1 min
3D effects are not taken into account, and the results need resolution. However, the algorithm is limited by the viewing
to be interpreted with care, especially in situations with bro- angle of the respective PV system or pyranometer, so it can
ken clouds. In addition, different cloud types such as thicker be inaccurate when there are many clouds on the horizon, for
cirrus clouds, mixed-phase clouds, or multi-layer clouds are instance.
not properly represented by the LUT. Due to the pseudo- The COD is then only extracted for data points for which
spherical approximation only SZAs up to 75◦ are consid- cf = 1, i.e. by finding the values of τ c,550 nm for which
ered (for SZAs above 75◦ with cloud cover the pseudospher-
6 6

ical DISORT solver is unstable). The cloud parameters in Gtot,SW,meas/inv = Gdir,SW,cloudy τ c,550 nm
Table 4 are input into libRadtran, and the COD at 550 nm 6

(τc,550 nm ) is varied on a 16-step logarithmic scale between + Gdiff,SW,cloudy τ c,550 nm (13)
COD = 0.5 and COD = 150, using the default Hu parame-
for all points under a cloud, where meas or inv refers to mea-
terisation (Hu and Stamnes, 1993) and 16 streams. Note that
sured or inverted GTI from pyranometers or PV systems, re-
the COD LUT also implicitly contains aerosol information
spectively. The corresponding direct and diffuse components
as an input, since here the aerosol properties are fixed using
can then also be extracted from the LUT, although in this
the OPAC database (Hess et al., 1998), and the Ångström pa-
case the direct irradiance is basically zero (beneath a cloud).
rameters from AERONET are used.
As described for the clear-sky simulation in Sect. 2.2, the 4 Note that for more accurate radiance calculations one could use
direct irradiance and diffuse radiance field are calculated the Mie option in libRadtran, which uses pre-calculated tables for
with libRadtran, the latter in this case with a coarser res- Mie scattering but is however computationally more expensive.

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4984 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

As mentioned above, this approach is limited by the fact Table 5. Limits on the input parameters for the MYSTIC LUT.
that a 1D radiative transfer solver such as DISORT cannot
take into account horizontal transport of photons so that 3D Input parameter Limits
effects such as radiative enhancement under broken-cloud SZA (θ 0 ) [20◦ , 60◦ ]
conditions (see for instance Schade et al., 2007) are not taken Tilt angle (θ) [0◦ , 50◦ ]
into account. For this reason only situations with overcast Relative azimuth (|φ − φ 0 |) [0◦ , 90◦ ]
conditions will be considered when applying this method. In Cloud fraction [0.13, 0.82]
situations with low overall cloud cover, the COD is not the
main determinant of the total irradiance received by the sen-
sor or PV system but rather the cloud fraction and the AOD.
To this end a complementary approach using a MYSTIC- 2.3.5 From tilted to horizontal irradiance with
based LUT (see Sect. 2.3.5) is used in order to translate mea- MYSTIC lookup table
sured tilted irradiance to horizontal irradiance under broken-
cloud conditions. In order to extract the global horizontal irradiance from
the tilted irradiance (from pyranometers or PV systems), a
2.3.4 Aerosol optical depth with DISORT lookup table MYSTIC-based LUT for the GHI was developed using large-
eddy simulation (LES) cloud fields (Črnivec and Mayer,
As mentioned above, in this work the extraction of the AOD 2019), taking into account various factors such as albedo, wa-
is not discussed in detail. However the procedure will briefly ter vapour, sensor geometry, and cloud fraction. Detailed 3D
be described here, since this is used as an alternative method radiative transfer simulations were carried out, and the most
for determining the GHI from tilted irradiance measure- important factors simply turned out to be the sensor and sun
ments. An AOD-GTI lookup table can be created in a sim- geometry as well as the cloud fraction. A detailed descrip-
ilar way to the COD LUT described in Sect. 2.3.3, where tion of the MYSTIC LUT is given in Chap. 9, Sect. 9.1.5, of
in this case the AOD at 500 nm is varied on a logarithmic Meilinger et al. (2021b).
scale in 16 steps between AOD = 0.01 and AOD = 1. In ad- Table 5 shows the limits of applicability of the MYSTIC
dition, the aerosol properties are fixed to the so-called conti- LUT, for which there are three major reasons. Firstly, despite
nental average scheme from the OPAC database (Hess et al., several optimisations like the reduction in the number of pho-
1998), and for the inversion procedure the AERONET-based tons used for the Monte Carlo simulations, the computational
Ångström parameters are not used as input. In this context demand for calculating the LUT is high. For this reason, 20◦
it must be noted that the typical dust event reaching Europe is chosen as the lower limit for the SZA, since in the lat-
does not have such a high AOD but is rather characterised by itudes under investigation no smaller values occur. Similar
small values of the Ångström exponent of less than 1, indicat- considerations apply to the tilt angle of PV panels – in All-
ing the presence of coarser dust particles. For example, one gäu, Germany, one rarely encounters title angles larger than
study of the climatology of dust events found a mean AOD of 50◦ . The second limiting factor relates to the derivation of a
0.155, 0.32, and 0.122 for dust plumes in southern, central, cloud mask and cloud fraction from the radiation measure-
and northern Europe, respectively (Mandija et al., 2018). ments (see Eq. 12). Firstly, this is only possible when there is
Using the AOD-GTI lookup table, the AOD can be ex- a direct line of sight between the sun and the module or sen-
tracted on clear-sky days, and from this also the direct and sor, which limits the relative azimuth angle between the sun
diffuse irradiance as well as the global horizontal irradiance. and the PV panel. Secondly, the derivation of cloud fraction
In this way the AOD is used as an intermediate step for the from temporally resolved radiation measurements becomes
reverse transposition of GTI to GHI. In Germany and espe- imprecise at large SZAs for geometrical reasons so that the
cially in the Allgäu region the AOD is usually very small upper limit of the SZA is set to 60◦ . Finally, the cloud frac-
(during the measurement campaigns it did not exceed 0.5 at tion limits are determined by two factors: firstly, the LUT
550 nm) so that any errors in the calibration procedure lead model was developed and tested for partly cloudy situations.
to large relative biases in the AOD. This then leads to bi- The special cases of 0 % and 100 % cloud fraction are con-
ases in the direct and diffuse components, but since there are sidered separately with the DISORT-based LUTs, as other
very few absorbing aerosols, these errors have opposite signs parameters like AOD and COD become relevant here. Sec-
and largely cancel out in the determination of the GHI. In ondly, the exact cloud fraction limits (0.13 and 0.82) given
Sect. 4, the GHI extracted via both COD and AOD under in Table 5 are constrained by the available cloud scenes from
different conditions is compared to that measured by pyra- LES simulations.
nometers and satellites, as well as the GHI predicted by the The measured or inverted tilted irradiance, together with
COSMO weather model. However the inferred AOD itself is the average cloud fraction over the last hour (as described
not examined in detail. above; see Eq. 12), is fed into the MYSTIC LUT, along with
the sensor and sun geometry. In this way the GHI can be
extracted from the GTI under broken-cloud conditions. This

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4985

method can however not be used to determine the optical compensate for errors in mounting the plane-of-array pyra-
depth, nor can the direct and diffuse irradiance components nometers, the calibration algorithm described in Sect. 2.2 is
be separated from each other, since the fit was created using also applied to the pyranometer data, in this case without an
the GTI and GHI. optical model and only using data up to an SZA of 60◦ . Due
to the substantial cosine bias, a correction factor is empiri-
cally determined:
3 Measurement and validation data
C(µ) = −3.01µ3 + 5.59µ2 − 3.34µ + 1.45, (14)
3.1 Ground-based measurements where µ = cos θ 0 for horizontal sensors and µ = cos 2 for
tilted sensors (θ 0 is the SZA and 2 the angle of incidence;
Model calibration and inversion are performed with PV see Eq. A8). The pyranometer data are used for comparison
power data recorded over two measurement campaigns in au- with the inverted irradiance (both tilted and horizontal) and
tumn 2018 and summer 2019 as part of the MetPVNet mea- for finding atmospheric optical properties using the lookup
surement campaign (see Chap. 3 of Meilinger et al., 2021b). table method.
There were a total of 24 stations spread out in the region In order to validate the PV- and pyranometer-based COD
around Kempten (47.715924◦ N, 10.314006◦ E), as shown in retrievals, it would be appropriate to use another ground-
Fig. 3, with 22 of them equipped with silicon-based pyra- based source of cloud optical properties; however unfortu-
nometers measuring both GHI and GTI in the plane-of-array nately there are no appropriate meteorological stations in the
of the PV system. Two master stations (MS01 and MS02) immediate area that could have been used for this purpose.
were also equipped with secondary standard pyranometers Although there are several DWD stations in the Allgäu re-
and pyrheliometers in order to measure both components of gion (in Kempten, Oberstdorf, and Hohenpeissenberg), these
the incoming short-wave radiation, cloud cameras, and spec- provide information on irradiance (direct and diffuse) but not
trometers to record spectral information. The MS01 station on cloud optical properties (see Becker and Behrens, 2012).
also contained a sun photometer to determine aerosol prop- Thus, a true validation would have to be done for another
erties, as part of AERONET, and a pyrgeometer to measure dataset with PV systems closer to a measurement station
long-wave downwelling irradiance. that has ground-based retrievals of COD. For this reason, the
The PV power data were for the most part provided by COD retrievals are simply compared to the corresponding
the local distribution network operator Allgäuer Überlandw- cloud properties from weather models and satellite data.
erk GmbH (AÜW), recorded in 15 min intervals. These data
represent the amount of energy generated in the last 15 min 3.2 Weather model data
so that care needs to be taken to translate them into a mea-
sured power corresponding to a specific time stamp. For this The Consortium for Small-scale Modelling (COSMO) nu-
purpose the data are simply shifted by half a period and re- merical weather model is a nonhydrostatic regional model
sampled, since by integration of power over 15 min one ef- developed by the DWD (Baldauf et al., 2011). Note that this
fectively smooths the power curve. In addition there were model was recently replaced by the so-called ICOsahedral
five stations equipped by egrid GmbH with high-frequency Nonhydrostatic (ICON) model, which has been fully opera-
power measurement devices: for these stations the power was tional since January 2021. Since the measurement campaigns
recorded in 1 Hz resolution. took place in 2018 and 2019, in this work the COSMO-EU
Analysis of the measured data revealed a total of 12 clear- model with a spatial resolution of 2.2 km is used, as input to
sky days that occurred between 12 September and 14 Oc- the clear-sky irradiance simulation (see Sect. 2.2.1), for PV
tober 2018, as well as 9 clear-sky days between 25 June model calibration (see Sect. 2.2.2), and for validation and
and 13 August 2019, as shown in Table 6. COSMO model comparison of the inverted irradiance with weather model
data for the corresponding days were procured from Ger- predictions. For the clear-sky simulation, temperature and
many’s national meteorological service, the Deutscher Wet- pressure profiles as well as the water vapour column are ex-
terdienst (DWD), in order to accurately recreate atmospheric tracted from COSMO data, whereas for both the calibration
conditions using cosmomystic. These days are used for and the inversion procedure the surface temperature and wind
calibration of each PV system. speed are used. For comparison and validation both direct
The network of PV systems was equipped with low-cost and diffuse downward irradiance data are used.
silicon-based pyranometers, with two devices per station: In order to compare COSMO COD data with the cloud op-
one mounted in the plane of the PV array and one horizon- tical depths extracted from the PV systems, a 2D COD field
tal, with 1 Hz resolution and an overall accuracy of 5 %. An must be computed from the 3D cloud variables generated by
absolute calibration of these sensors had been carried out at the COSMO model. For each grid cell, a cloud fraction vari-
the Leibniz Institute for Tropospheric Research (TROPOS) able in COSMO indicates which fraction of the cell is cov-
prior to the campaign by comparing their output to that of a ered by clouds. To derive a vertically integrated COD, an as-
secondary standard pyranometer (2 % accuracy). In order to sumption needs to be made as to how these clouds overlap

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4986 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 3. Map showing the locations of PV systems used in the MetPVNet measurement campaign (taken from © Google Earth). The top-left
corner is at 47.85◦ N, 10.09◦ E and the bottom-right corner at 47.38◦ N, 10.52◦ E. The yellow line marks the border between Germany and
Austria; the grey line is the border between the states of Bavaria and Baden-Württemberg.

Table 6. Dates of clear-sky days during the measurement campaigns in autumn 2018 and summer 2019.

12 September 2018 17 September 2018 20 September 2018 27 September 2018 30 September 2018 4 October 2018
First campaign
5 October 2018 8 October 2018 10 October 2018 12 October 2018 13 October 2018 14 October 2018
26 June 2019 27 June 2019 28 June 2019 29 June 2019 30 June 2019 4 July 2019
Second campaign
23 July 2019 24 July 2019 25 July 2019

in a model column. Following Scheck et al. (2018), the com- 3.3 Satellite data
monly used random-maximum cloud overlap assumption is
adopted, along with the method of Matricardi (2005), in or- The Copernicus Atmospheric Monitoring Service (CAMS)
der to compute the vertically integrated COD for a number of radiation service (Qu et al., 2017; Schroedter-Homscheidt et
subcolumns within each model column. From these subcol- al., 2022) is an online satellite-based and numerical-model-
umn values a mean COD for the cloudy part of the column is based service with radiation as well as cloud and aerosol
derived. A total COD is then computed as the average of the data available to download for free, covering the period from
column mean COD over 5 × 5 columns centred around the February 2004 to the present. The spatial coverage is Europe,
column containing the relevant ground station. Africa, the Middle East, the eastern part of South America,
and the Atlantic Ocean, interpolated to the point of interest,
and with a time resolution of up to 1 min. In this work the

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4987

global, direct, and diffuse components of irradiance are im- ing algorithm distorts the clear-sky days, whereas for the sys-
ported from CAMS (version 4.0), for each station and for all tems at PV11 the different PV arrays at the site are not well
days in the two measurement campaigns. These data are used characterised (see the point labelled 11,3 in Fig. 4). In other
as a comparison for the irradiance inverted from PV systems. cases shading effects played a role: in most cases the calibra-
In addition to irradiance, CAMS provides data on cloud tion performed better when using both summer and autumn
and aerosol properties. In this work, the cloud parame- data, since in summer the sun is much higher, and shading
ters from the Advanced Very High Resolution Radiome- effects play a smaller role. In general the model calibration
ter (AVHRR) Processing scheme Over cLoud, Land, and works best when using as much data as possible, since there
Ocean Next Generation (APOLLO_NG) analysis (Kriebel is for instance more variation in temperature in order to find
et al., 2003; Klüser et al., 2015) are used, using data from more reliable temperature model parameters.
the Spinning Enhanced Visible and InfraRed Imager (SE- As discussed in Sect. 2.2.2, several parameters are corre-
VIRI) instrument on board the Meteosat Second Generation lated with each other: the size of the PV system (captured by
(MSG) satellite. In this case the so-called Stephens method the factor s) correlates with the tilt angle θ , whereas the az-
(Stephens et al., 1984) is used to determine the COD, using a imuth angle φ shows a large correlation with the parameters
two-stream solution of the radiative transfer equation, along of the temperature model, since the warming up and cooling
with an updated algorithm using a probabilistic approach for down of the PV system are delayed with respect to the diur-
cloud detection (Klüser et al., 2015). For comparison with nal variation in solar irradiance. In general the use of mea-
the COD inferred from the PV systems, APOLLO_NG data sured module temperature leads to better calibration results.
are extracted for the closest pixel to each station. It turns out that the calibration algorithm presented here does
not perform well when using the non-linear Faiman temper-
ature model and 15 min power data, even though this model
4 Results couples irradiance and wind speed in a more physically cor-
rect way (see for instance Faiman, 2008; Barry et al., 2020).
The calibration and inversion procedure described in Sect. 2 The benefit of this model is lost for coarsely resolved 15 min
is applied to the data from the measurement campaign de- data so that in the end the algorithm proposed here does not
scribed in Sect. 3 in order to extract irradiance and optical always find an optimal solution, specifically if the tempera-
properties from the PV systems in the Allgäu region. After ture model parameters are unknown. The bias that then oc-
a brief summary of the calibration results in Sect. 4.1, the curs in the final tilted irradiance inversion results can be seen
retrievals of tilted and horizontal irradiance are presented in in the plots in Sect. 4.2 as well as in the results in Tables 8
Sect. 4.2 and 4.3, and the inferred COD results are shown in and 9. However, this bias in the tilted irradiance does not al-
Sect. 4.4. ways translate into a bias in GHI, as is seen below. Table 7
lists the PV systems used in this work, along with the corre-
4.1 Model calibration and uncertainty sponding time resolution of their data.

The PV models in Eqs. (1) and (2) are used together with the 4.2 Global tilted irradiance from PV power data
clear-sky simulation described in Sect. 2.2.1 and the clear-
sky days (in Table 6) in order to calibrate each system. In In this section the plane-of-array irradiance from PV power
each individual case the days that turned out not to be clear retrievals is compared to the tilted pyranometer measure-
are discarded from the calibration dataset, and the data are re- ments at selected stations during the two measurement cam-
stricted to the time periods in which the required inputs such paigns. The results are obtained using two different ap-
as ambient temperature, atmospheric long-wave irradiance, proaches for module temperature: (i) the linear temperature
and wind speed are available. In order to validate the calibra- model (Eq. A3) and (ii) the non-linear Faiman temperature
tion results, the retrieved elevation and azimuth angles are model (Eq. A4). The results are compared in Tables 8 and 9,
compared to ground truth data from the Bavarian Agency for and scatterplots for both models are shown. Both stations
Digitisation, High-Speed Internet and Surveying (LDBV). with 15 min power data as well as those with high-frequency
The so-called Level of Detail 2 (LoD2) database (https:// data (1 Hz data, smoothed to 1 min resolution) are included
ldbv.bayern.de/produkte/3dprodukte/3d.html, last access: 24 in the analysis, and in each case a comparison is made with
October 2023) contains a 3D building model constructed us- the measurements from TROPOS silicon-based pyranome-
ing airborne laser scanning so that the roof pitch of individual ters, except for the MS02 master station, where the tilted
buildings can be extracted. Figure 4 shows a comparison of Kipp & Zonen CMP11 pyranometer is used for validation.
the retrieved orientation angles with the ground truth values In all cases a limit of 80◦ is imposed on both the solar zenith
for each system and using the linear temperature model. angle and the incident angle in order to avoid possible errors
In most cases the algorithm finds reasonable values for the from both the radiative transfer simulation and the optical
angles: the larger deviations can usually be explained for in- model.
dividual cases; for instance for PV04 the inverter MPP track-

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4988 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 4. Comparison of the retrieved elevation angles (θopt , left plot) and azimuth angles (φopt , right plot) with the corresponding ground
truth angles (θactual and φactual ) from airborne laser scanning data, for PV systems with 15 min AÜW (red points) and 1 Hz egrid (blue points)
data (cf. Table 7 and the description in Sect. 3.1).

Table 7. List of the PV systems used for this work (see also the map in Fig. 3). The data resolution column indicates whether a particular
system is used for this analysis or not, with an explanation given in the cases where the system is omitted. Note that stations MS01 and PV02
had no PV systems, only pyranometers and other measurement equipment. PV11 has four separate PV systems.

Station Data resolution Mounting Comments

2018 2019
MS02 15 min 15 min/1 s Ground
PV01 15 min 15 min Rooftop
PV03 – – Rooftop Calibration errors
PV04 – – Rooftop Calibration errors
PV05 – – Rooftop No data
PV06 15 min 15 min Ground
PV07 – – Rooftop Calibration errors
PV08 15 min – Rooftop No data in 2019
PV09 – – Rooftop open Calibration errors
PV10 15 min 15 min Ground
PV11,1 15 min 1s Rooftop
PV11,2 15 min 15 min Rooftop
PV11,3 15 min 15 min Rooftop
PV11,4 1s – Ground No data in 2019
PV12 1s 1s Rooftop
PV13 – – Rooftop No 2018 data, calibration problems
PV14 15 min 15 min Rooftop open
PV15 1s 1s Rooftop
PV16 15 min 15 min Rooftop
PV17 15 min 15 min Rooftop
PV18 15 min 15 min Rooftop
PV19 1s 1s Rooftop
PV20 – – Rooftop No data
PV21 15 min 15 min Rooftop
PV22 – – Rooftop Calibration errors

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4989

Table 8. Mean bias error (in W m−2 ) and relative mean bias error (in brackets in %) of GTI retrievals compared to tilted pyranometers. The
values marked with ∗ are high due to calibration errors using the non-linear temperature model with 15 min data.

2018 2019
Linear Non-linear Linear Non-linear
1 min 16.23 (3.2) 5.29 (1.0) 21.12 (4.1) 12.25 (2.4)
15 min 28.74 (5.9) 87.55∗ (18.0)∗ 40.20 (7.8) 96.40∗ (18.6)∗

Table 9. Root mean squared error (in W m−2 ) and relative RMSE (in brackets in %) of GTI retrievals compared to tilted pyranometers. The
values marked with ∗ are high due to calibration errors using the non-linear temperature model with 15 min data.

2018 2019
Linear Non-linear Linear Non-linear
1 min 72.34 (14.3) 82.68 (16.4) 73.97 (14.2) 79.71 (15.3)
15 min 108.27 (22.2) 172.96∗ (38.1)∗ 83.94 (16.3) 197.83∗ (38.2)∗

Figures 5 and 6 show a comparison between the retrieved angles. Another possible reason for the positive bias could
GTI and that measured by pyranometers for the 1 and 15 min be a systematic bias in the tilted pyranometer measurements,
data, respectively, for each measurement campaign and us- even after the bias correction described in Sect. 3.1.
ing both the linear and the non-linear temperature models. The two different temperature models achieve similar re-
The corresponding statistical measures of the mean bias er- sults for the 1 min data, with the non-linear model show-
ror (MBE), defined by ing an MBE of 5.29 W m−2 (12.25 W m−2 ) in autumn 2018
(summer 2019) compared to an rMBE of 16.23 W m−2
n
1X
(21.12 W m−2 ) for the linear model. In general the algo-
MBE = Xinv,i − Xref,i , (15)
n i=1 rithm performs worse with 15 min data, which has to do
with errors from the calibration procedure and uncertainties
and the root mean squared error (RMSE), defined by in the PV power measurements – the systems with high-
v frequency measurements are thus in general better charac-
u n
u1 X
2 terised and deliver more accurate irradiance retrievals, as
RMSE = t Xinv,i − Xref,i , (16) shown in Sect. 4.3 below. This effect is quite extreme for the
n i=1
non-linear Faiman temperature model (see the values marked
with ∗ in Tables 8 and 9 as well as the plots in the lower
for the inverted quantities Xinv,i and the reference quantities panels of Fig. 6), since in some cases the calibration algo-
Xref,i are shown in Tables 8 and 9, along with the relative rithm cannot find an optimal solution, and the a priori val-
error metrics rMBE and rRMSE calculated by normalising ues have to be relied upon, leading to an average MBE of
the MBE and RMSE with the mean of the reference quantity, 91.98 W m−2 .
hXref i. The scatterplots throughout this work are coloured ac-
cording to a probability density function calculated using the 4.3 Global horizontal irradiance from PV power data
multi-variate Gaussian kernel density function gaussian_kde
in the Python toolbox SciPy, with yellow (light grey) for In the following the global tilted irradiance (GTI) retrievals
high- and blue (dark grey) for low-frequency points in the are converted to global horizontal irradiance (GHI) and com-
colour (black and white) version. In Figs. 5 and 6 one can see pared to the measurements from pyranometers as well as to
that most points lie close to the 1 : 1 line, for both campaigns, the satellite and weather model data. This conversion is per-
albeit with a positive bias in all cases. The 1 min data show a formed in three different ways, depending on the prevailing
slightly larger spread of points than the 15 min data, since in weather conditions, as described in Sect. 2.3. In the case of
the former case there are more outliers caused by (i) temper- clear skies, the DISORT-based LUT is used to find the value
ature effects, (ii) 3D radiative transfer effects, and (iii) spa- of aerosol optical depth, and the inferred AOD is then used
tial effects due to the differences in cloud cover and the sen- to calculate the direct and diffuse horizontal irradiance com-
sor position between the PV and pyranometer. In addition ponents, which result in the GHI. The same method is used
the slightly different geometry of flat PV arrays compared under cloudy skies using the DISORT-based COD LUT but
to glass-dome-shaped pyranometers could play a role, espe- only for times at which the sensor is under a cloud, using the
cially when it comes to their sensitivity to different viewing inferred cloud fraction method as described in Sect. 2.3.3.

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4990 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 5. Combined comparison of GTI retrieved from PV power measurements with that measured by tilted pyranometers, using data in
1 min resolution from MS02, PV11, PV12, PV15, and PV19 (see Table 7), for 2018 (a, c) and 2019 (b, d), together with the linear (a, b) and
non-linear (c, d) temperature models. The relative mean bias error (rMBE) and relative root mean squared error (rRMSE) are shown in the
inset, along with the mean of the reference GTI and the number of data points used, denoted as hGref i and n, respectively.

For this reason, this method tends to underestimate the GHI 4.3.1 GHI retrieval validation with pyranometer
under broken-cloud conditions, which is discussed in detail measurements
below.
The third approach for finding the GHI is using the As discussed in Sect. 4.2 above, the PV systems with 1 min
MYSTIC-based LUT described in Sect. 2.3.5, where the data show the best calibration results and the most accurate
LUT input parameters are simply the array geometry, sun tilted irradiance retrievals. The scatterplots in Fig. 7 com-
position, and cloud fraction. In this case there are certain re- pare the GHI retrieved from these systems to that measured
strictions on these parameters, as shown in detail in Table 5. by horizontal pyranometers, using all three methods and for
This means that not all of the retrieved GTI data points can both temperature models. The statistical measures of the dif-
be transformed into GHI using this method – in particular the ferent retrievals are shown in Table 10, where it can be seen
SZA is limited to between 20 and 60◦ and the cloud fraction that the mean bias error reaches the goal of 5 W m−2 only in
to between 0.13 and 0.82 so that neither completely overcast certain cases.
nor clear-sky conditions are taken into account. This method Under clear conditions (top row of Fig. 7), the linear model
thus deals with the case of mixed-/broken-cloud conditions, applied to 1 min PV data achieves an rMBE for the GHI
in which it is more likely that there will be errors due to 3D of 18.15 W m−2 (3.9 %) in autumn 2018 and 9.44 W m−2
effects and the sensor position. (1.4 %) in summer 2019. Note that the mean irradiance is
higher in summer, but there are less points that can be classi-
fied as clear (n ' 9400 compared to n ' 13 400 in autumn).

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4991

Figure 6. The same as Fig. 5 using data in 15 min resolution from MS02, PV01, PV06, PV08, PV10, PV11, PV14, PV16, PV17, PV18, and
PV21 (see Table 7) for 2018 (a, c) and 2019 (b, d). Note that the large errors in the non-linear model in the bottom row result from errors in
the calibration procedure (see the values marked with ∗ in Tables 8 and 9).

Table 10. The mean bias error and root mean squared error (in W m−2 ) along with the rMBE and rRMSE (in brackets in %) of GHI retrievals
using the 1D DISORT (AOD and COD) and the 3D MYSTIC LUTs compared to horizontal pyranometers, for 1 min and 15 min data.

Data Measure LUT 2018 2019

Linear Non-linear Linear Non-linear
1D AOD 18.15 (3.9) 1.83 (0.4) 9.44 (1.4) 9.75 (1.5)
MBE 1D COD 18.32 (8.2) 16.68 (7.4) 13.85 (4.7) 8.73 (2.9)
3D GHI 29.90 (5.6) 0.60 (0.1) 20.69 (3.1) 3.39 (0.5)
1 min
1D AOD 30.17 (6.5) 33.55 (7.3) 33.81 (5.1) 35.24 (5.3)
RMSE 1D COD 63.10 (28.2) 64.70 (28.6) 66.72 (22.7) 70.38 (23.3)
3D GHI 89.76 (16.8) 87.89 (16.3) 102.69 (15.4) 111.28 (16.5)
1D AOD 10.01 (2.3) 1.76 (0.4) 1.75 (0.3) 6.44 (1.0)
MBE
1D COD 37.34 (15.5) 40.65 (17.5) 46.39 (15.6) 44.55 (15.4)
15 min
1D AOD 38.96 (9.0) 39.59 (9.5) 35.93 (5.2) 36.89 (5.5)
RMSE
1D COD 78.32 (32.5) 82.74 (35.6) 91.96 (30.8) 100.71 (34.8)

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4992 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 7. Combined comparison of GHI retrieved from PV power measurements with that measured by horizontal pyranometers, using data
in 1 min resolution from MS02, PV11, PV12, PV15, and PV19, for 2018 (left two columns) and 2019 (right two columns), and both the
linear and the non-linear temperature models. The top row shows GHI retrieved via AOD under clear skies using the DISORT AOD LUT,
the middle row is for cloudy periods via the COD using the DISORT COD LUT, and the bottom row is for broken-cloud periods using the
MYSTIC 3D LUT.

The non-linear model performs significantly better in au- 17.50 W m−2 for the linear model. In autumn the lower aver-
tumn, with an rMBE for the GHI of 1.83 W m−2 (0.4 %), but age irradiance of 225 W m−2 leads to a higher relative MBE
in summer the bias is similar to the linear model [9.75 W m−2 than in summer, where the average irradiance is 298 W m−2 .
(1.5 %)]. Interestingly the linear temperature model performs At this point it is worth mentioning that the algorithm only
better in summer than in autumn, whereas the non-linear finds the COD and thus the GHI when the sensor is un-
model performs better in autumn. This could be attributed to der a cloud, hence the lower average irradiance in compar-
differences or uncertainties in the calibration of the tempera- ison with that under clear skies, and the RMSE is higher
ture models. In general these results show that the DISORT (66.23 W m−2 on average) than in the case of clear skies,
LUT method performs comparably well for extracting GHI where it is 33.19 W m−2 on average, as expected. This also
from PV power measurements under clear conditions. In all means that averaging these results over 60 min can lead to
cases the rRMSE is of the order of 5 % to 7 %, on average erroneous values for the irradiance, especially under broken
33.19 W m−2 . clouds, since the periods of cloud enhancement within each
It is evident from the middle row of Fig. 7 that the bias 1 h window will not be taken into account.
is greater under cloudy skies, with an average over both The GHI retrieved from the 3D MYSTIC LUT shows sig-
campaigns of 11.29 W m−2 for the non-linear model and nificantly lower bias in the case of the non-linear temper-

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4993

ature model (2.00 W m−2 compared to 12.71 W m−2 using and cloud optical depths and that from the CAMS retrieval,
the DISORT COD method, averaged over both campaigns), for 1 and 15 min power data, respectively. Under clear-sky
but in the case of the linear temperature model the bias is conditions the retrieved GHI shows an average MBE of
higher (25.30 W m−2 compared to 16.09 W m−2 ). The good 15.95 W m−2 for the linear model and 7.54 W m−2 for the
performance of the non-linear model could indeed be a re- non-linear model. These values are similar to those found
sult of the improved treatment of the PV module tempera- by comparing with ground-based pyranometers, confirming
ture during broken-cloud conditions: although neither model the accuracy of the CAMS data. On the other hand, the GHI
contains a dynamic term, the non-linear model couples ir- retrieved using the DISORT COD LUT under cloudy skies
radiance and wind speed in a more physically correct way. shows a significant negative bias compared to the CAMS
However, it must be noted that due to the restrictions on the retrieval (−37.77 W m−2 in autumn and −86.87 W m−2 in
MYSTIC LUT (see Table 5), the number of inferred irradi- summer). There are two possible reasons for this: firstly the
ances included in the statistics is far lower than for the COD simplification to one cloud type means that thinner clouds or
method. Another confounding factor could be that the PV multi-layer cloud situations are not properly represented in
panels show better efficiency at higher irradiance, and since the model (see the discussion on COD retrievals in Sect. 4.4
the MYSTIC method also takes overshoots into account, the below). However, the main reason is related to the retrieval
irradiance is on average higher: 603.78 W m−2 compared to algorithm: by only considering measurements where the PV
261.45 W m−2 for the COD method. The larger fluctuations system is under a cloud, only the periods with lower irra-
in irradiance under broken-cloud conditions also lead to a diance values are retrieved, and overshoots are ignored so
larger RMSE than for the case of cloudy skies, and in sum- that at 1 min resolution a large negative bias in irradiance
mer the RMSE is the highest, as expected. is found. Since the CAMS irradiance retrieval is based on
Figure 8 shows the GHI retrievals from the AÜW systems the Heliosat-4 method, in which cloud properties from the
with 15 min PV power measurements, under clear (top row) APOLLO_NG method are updated every 15 min (Qu et al.,
and cloudy (bottom row) skies. In this case the MYSTIC 2017), one should expect this bias to reduce at coarser reso-
3D LUT is not used, since the determination of the cloud lution. Indeed, the 15 min data in the bottom row of Fig. 10
fraction with coarsely resolved data leads to erroneous re- confirm this: the average bias is reduced to −3.73 W m−2 in
sults, and the rapid fluctuations in irradiance under broken autumn and −15.07 W m−2 in summer. Note that an averag-
clouds are not properly captured at 15 min resolution. The ing to 60 min is not performed due to the limitations of the
DISORT 1D LUT performs well under clear skies, as to be DISORT COD algorithm, as discussed in Sect. 4.3.1.
expected, with the linear model showing an average MBE The comparison with COSMO model data is shown in
of 5.88 W m−2 and the non-linear model 4.10 W m−2 . Once Fig. 11 and Table 12, where in this case the data are av-
again the non-linear model outperforms the linear one in au- eraged over a 60 min period. It is evident that the COSMO
tumn 2018, but this trend is reversed in summer 2019. This model shows a bias under clear-sky conditions: here the as-
systematic effect is most probably due to uncertainties in sumed AOD is too high so that the irradiance turns out to
the temperature model calibration. Under cloudy skies the be too small, with an average bias of 60.92 W m−2 . On the
GHI retrievals show a significant positive bias of on aver- other hand, under cloudy conditions and especially under
age 41.87 W m−2 (42.60 W m−2 ) for the linear (non-linear) low-light conditions in summer, the irradiance from COSMO
model, which means that the retrieved COD is too small. This is larger than that retrieved from PV plants, which means that
is discussed further in Sect. 4.4. Interestingly the large bias the COD in COSMO is too small. Here the average bias is
errors in tilted irradiance for the non-linear model are not −38.36 W m−2 . These results confirm the findings of Frank
evident in the horizontal irradiance results, which is proba- et al. (2018) and Zubler et al. (2011) and are discussed fur-
bly due to the fact that far less points are taken into account ther in connection with the cloud optical depth in Sect. 4.4
in the statistics for GHI (compare the values of n in Figs. 8 below.
and 6). In other words, the outliers have been removed by se-
lecting either clear-sky days or periods for which the cloud 4.4 Cloud optical depth retrievals
fraction is 100 %.
As discussed in Sect. 2.3.3, the cloud optical depth is re-
4.3.2 Comparison to satellite and weather model trieved from both PV systems and pyranometers, using a
irradiance data DISORT-based LUT. In order to avoid errors due to 3D ef-
fects, in this work only data with a cloud fraction of 1 are
One of the main aims of this work is to show that PV systems considered, in other words only completely overcast con-
can provide a reliable source of information on global hori- ditions. The results for the linear temperature model are
zontal irradiance that is complementary to that from satel- shown in Figs. 12 and 13, compared to the APOLLO_NG
lite and weather models. Figures 9 and 10 (with correspond- and COSMO data, respectively. As can also be seen in Ta-
ing statistical measures in Table 11) show the comparison bles 13 and 14, in most cases a smaller COD is extracted
between GHI retrieved from PV power using the aerosol from PV systems, except for the comparison between the

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4994 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 8. Combined comparison of GHI retrieved from PV power measurements with that measured by horizontal pyranometers, using data
in 15 min resolution from MS02, PV01, PV06, PV08, PV10, PV11, PV14, PV16, PV17, PV18, and PV21, for 2018 (left two columns) and
2019 (right two columns), and both the linear and the non-linear temperature models. The top row shows GHI retrieved via AOD under clear
skies using the DISORT AOD LUT, and the bottom row is for cloudy periods via the COD using the DISORT COD LUT.

Figure 9. Combined comparison of GHI retrieved from PV power measurements with that from CAMS, using data in 1 min resolution from
MS02, PV11, PV12, PV15, and PV19, for 2018 (left two columns) and 2019 (right two columns), and both the linear and the non-linear
temperature models. The top row shows GHI retrieved via AOD under clear skies using the DISORT AOD LUT, and the bottom row is for
cloudy periods via the COD using the DISORT COD LUT.

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4995

Figure 10. Combined comparison of GHI retrieved from PV power measurements with CAMS, using data in 15 min resolution from MS02,
PV01, PV06, PV08, PV10, PV11, PV14, PV16, PV17, PV18, and PV21, for 2018 (left two columns) and 2019 (right two columns), and
both the linear and the non-linear temperature models. The top row shows GHI retrieved via AOD under clear skies using the DISORT AOD
LUT, and the bottom row is for cloudy periods via the COD using the DISORT COD LUT.

Table 11. The mean bias error and root mean squared error (in W m−2 ) along with the rMBE and rRMSE (in brackets in %) of GHI retrievals
using 1D DISORT (AOD and COD) and the 3D MYSTIC LUT compared to CAMS, for 1 and 15 min data.

Data Measure LUT 2018 2019

Linear Non-linear Linear Non-linear
1D AOD 18.07 (3.9) 1.16 (0.3) 13.83 (2.1) 13.91 (2.1)
MBE
1D COD −37.68 (−13.4) −37.86 (−13.5) −85.85 (−22.1) −87.88 (−22.3)
1 min
1D AOD 23.51 (5.1) 32.19 (7.0) 31.99 (4.8) 32.48 (4.9)
RMSE
1D COD 112.68 (40.2) 112.78 (40.2) 163.15 (42.0) 163.61 (41.6)
1D AOD 5.74 (1.3) −1.25 (−0.3) 1.31 (0.2) 5.78 (0.9)
MBE
1D COD −6.09 (−2.1) −1.37 (−0.5) −13.70 (−3.8) −16.44 (−4.7)
15 min
1D AOD 28.74 (6.5) 30.95 (7.3) 29.26 (4.3) 25.43 (3.8)
RMSE
1D COD 75.78 (26.6) 79.69 (29.0) 114.36 (32.1) 117.01 (33.6)

summer campaign and COSMO data, in which case a pos- trend similar to the PV-based ones: once again it is evident
itive mean bias of approximately COD = 8 is found. Overall, that the COD is mostly below 10, and in this range the re-
the COD is mostly in the range between 1 and 10, for both trieved data have a positive relative bias. There are several
campaigns. Taken at face value, the negative bias with re- possible reasons for this: firstly it is evident from Eq. (11)
spect to APOLLO_NG would imply a positive bias in GHI that the retrieval is more sensitive to errors in inverted irra-
with respect to CAMS, which is not seen in the 1 and 15 min diance (transmission) for smaller COD. On the other hand,
retrievals. However these results cannot be directly compared it must also be noted that the efficiency of both PV mod-
due to the effect of both spatial and temporal averaging as ules and silicon-based pyranometers shows a logarithmic de-
well as the limitation of the DISORT COD LUT algorithm, pendence on irradiance so that any inaccuracies in the PV
which ignores 3D effects. model parameters or the pyranometer calibration will have a
Figures 14 and 15 show the same results using measured larger effect on the inverted irradiance under low-light condi-
pyranometer data to infer the COD. These retrievals show a tions (higher COD). In addition, since the LUT is constructed

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4996 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Figure 11. Combined comparison of GHI retrieved from 60 min averaged PV power measurements under clear (top row) or completely
cloudy (bottom row) conditions using the optical depth via DISORT LUT with that from the COSMO model, for all stations and for 2018
(left two columns) and 2019 (right two columns), using both the linear and the non-linear temperature models.

Table 12. The mean bias error and root mean squared error (in W m−2 ) along with the rMBE and rRMSE (in brackets in %) of 60 min
averaged GHI retrievals using 1D DISORT (AOD and COD) and the 3D MYSTIC LUT compared to COSMO model data.

Data Measure LUT 2018 2019

Linear Non-linear Linear Non-linear
1D AOD 65.90 (16.2) 57.72 (14.7) 58.74 (9.9) 61.33 (10.5)
MBE
1D COD 14.34 (5.2) 16.77 (6.1) −37.50 (−9.4) −39.21 (−9.8)
60 min average
1D AOD 72.71 (17.9) 67.70 (17.2) 65.65 (11.1) 67.32 (11.5)
RMSE
1D COD 124.16 (44.8) 125.12 (45.4) 143.86 (36.1) 144.96 (36.4)

with water clouds, the effect of optically thin ice clouds is have enough information to accurately determine cloud opti-
not properly taken into account. Since ice particles scatter cal properties in all situations.
slightly less in the forward direction, 1 − g ' 0.3 is larger Notwithstanding the bias in COD retrievals, the ground-
than for water clouds (1 − g ' 0.15), and thus a smaller op- based method presented here can still complement satel-
tical depth could lead to similar irradiance at the ground (see lite retrievals, in particular due to the potentially higher
Eq. 11). Thirdly, since clouds become more absorbing in the spatiotemporal resolution achievable with large amounts of
near infrared and considering that silicon PV is sensitive to high-frequency data spread over a large area. Once again, for
wavelengths up to 1200 nm, spectral effects could also lead the summer months the COSMO data show a large mean bias
to a bias in the results. In general it must be said that even error of COD = 7.69, even for large values of COD, confirm-
with measurements at two different wavelengths there ex- ing the findings of Frank et al. (2018).
ists an ambiguity in the determination of effective radius and
COD (Nakajima and King, 1990) so that in the case of spec-
trally integrated PV-inverted irradiance one cannot expect to

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4997

Figure 12. Combined comparison of COD retrieved from PV power measurements under completely cloudy conditions using the DISORT
LUT with that from APOLLO_NG, for all stations and for 2018 (a) and 2019 (b), using the linear temperature model and averaged over
60 min.

Figure 13. The same as Fig. 12 but compared to the total COD from the COSMO model.

Table 13. The mean bias error and root mean squared error (in W m−2 ) along with the rMBE and rRMSE (in brackets in %) of COD
retrievals from PV systems compared to the APOLLO_NG data.

Data Measure 2018 2019

Linear Non-linear Linear Non-linear
MBE −3.22 (−25.0) −3.57 (−27.2) −3.58 (−15.4) −3.24 (−13.6)
60 min average
RMSE 15.20 (117.9) 15.66 (119.3) 18.99 (81.7) 19.30 (81.2)

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

4998 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

Table 14. The mean bias error and root mean squared error (in W m−2 ) along with the rMBE and rRMSE (in brackets in %) of COD retrievals
from PV systems compared to the COSMO model predictions.

Data Measure 2018 2019

Linear Non-linear Linear Non-linear
MBE −8.78 (−47.6) −9.16 (−49.0) 7.32 (59.3) 8.06 (64.6)
60 min average
RMSE 25.77 (139.7) 26.06 (139.2) 22.22 (180.1) 23.52 (188.6)

Figure 14. Combined comparison of the COD retrieved from tilted (plane-of-array – poa) pyranometer measurements under completely
cloudy conditions using the DISORT LUT with that from APOLLO_NG, for all stations and for 2018 (a) and 2019 (b), using the linear
temperature model and averaged over 60 min.

Figure 15. The same as Fig. 14 but compared to the total COD from the COSMO model.

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 4999

5 Conclusions and outlook and −8.91 W m−2 , respectively. Considering the difference
between point measurements and satellite pixels, as well as
In summary, in this work a framework for extracting both the inherent bias from considering only periods with 100 %
global tilted and horizontal irradiance from PV power data cloud fraction for the DISORT method, these results must be
has been presented, and a first test for retrieving cloud opti- interpreted with care. In the case of clear skies, small errors
cal depth is carried out. The algorithm makes use of state-of- in the temperature model can lead to large errors in extracted
the-art radiative transfer solvers in libRadtran, in conjunction irradiance, and in the case of cloudy skies, simplifying as-
with different sources of data for the state of the atmosphere, sumptions on the cloud type can lead to errors in COD and
in particular the aerosol and water vapour content. The cali- extracted irradiance.
bration procedure uses an explicit calculation of diffuse and The retrieved GHI shows a large bias when compar-
direct irradiance, taking into account the spectral response ing it with COSMO model data, thus confirming the re-
of the relevant PV technology and the optical properties of sults of Frank et al. (2018) and Zubler et al. (2011): un-
the glass surface. Where necessary the module temperature is der clear skies the 60 min averaged GHI has a mean bias
modelled using weather model data for ambient temperature error of 60.92 W m−2 , since COSMO has in general a too
and wind speed input. The PV systems are calibrated using large aerosol load, whereas under cloudy skies in summer
a libRadtran clear-sky simulation with the DISORT solver, the MBE is −38.36 W m−2 , since under cloudy conditions
with inputs from the COSMO model and AERONET, and the COSMO model generally tends to overestimate the irra-
the algorithm can be adapted to each system and situation, diance. The latter result is also confirmed by the COD re-
depending on which parameters are known and which need trievals: in summer there is a positive bias of 7.69 for the
to be determined by non-linear optimisation. retrieved COD relative to COSMO – the COD in the weather
Once calibrated, the measured PV data are used to extract model is thus on average too small.
global tilted irradiance under all sky conditions. In order to Overall, the largest source of error in the model chain
take into account the spectral mismatch between the spec- comes from the PV model itself – an accurate calibration is
tral response of PV modules and the entire broadband spec- vital in order to be able to extract irradiance reliably. In gen-
trum, a situation-dependent fit for the dependence of this mis- eral the non-linear temperature model performs the best with
match on atmospheric conditions is performed, using simu- high-frequency PV data. In this regard it is helpful to use
lated data for clear-sky conditions and the water vapour and measured module temperature rather than relying on temper-
aerosol load of each site. The GTI is then compared to that ature models. More accurate results could also be achieved
measured by tilted pyranometers: the retrieved GTI at 1 min by using PV current measurements, since in this case the
(15 min) resolution has a mean bias error of 18.68 W m−2 temperature dependence is almost negligible. This will be ex-
(34.47 W m−2 ) averaged over the two measurement cam- plored in future work.
paigns, using the linear temperature model. The non-linear The DISORT LUT is only employed during periods of per-
Faiman temperature model achieves a mean bias error of sistent cloudy or clear-sky conditions in order to infer COD
8.77 W m−2 for the systems with 1 min data, but for those or AOD, respectively, whereas the effect of 3D transport of
with 15 min data the calibration algorithm fails to find an op- photons is only taken into account with the MYSTIC LUT
timal solution in several cases so that in the end the mean bias for the GHI. This means that the algorithm in its present form
error is 91.98 W m−2 . This shows that an accurate calibration can only extract direct and diffuse components reliably under
is essential in order to accurately extract irradiance. stable conditions for which 1D radiative transfer is still a rea-
The inverted GTI is used to find the global horizontal ir- sonable approximation. A possible extension to this will be
radiance using three different methods: (i) under persisting studied in future work, in which explicit 3D simulated cloud
clear or (ii) cloudy conditions a lookup table based on a 1D fields will be used in conjunction with the 1D DISORT simu-
DISORT simulation is used in order to find either the AOD or lation in order to quantify the error that results from neglect-
the COD and thus the global horizontal irradiance; (iii) under ing 3D radiative transfer. It is only once the three dimension-
broken-cloud conditions a LUT based on 3D MYSTIC sim- ality of atmospheric radiative transfer and additional infor-
ulations is used to translate the tilted irradiance to horizontal mation on cloud type are taken into account that one can ac-
irradiance, using the geometry of the system, sun position, curately extract the direct and diffuse irradiance components
and cloud fraction as inputs. The retrieved GHI is then com- under broken-cloud conditions.
pared to pyranometer measurements: in the case of the 1D Another aspect not taken into account in this study is the
LUT method, with 1 min data under clear (cloudy) skies, the possible gain in using several different PV systems that are
mean bias error is 13.79 W m−2 (16.09 W m−2 ) for the lin- close enough to each other so as to be able to see the same
ear model and 5.79 W m−2 (12.71 W m−2 ) for the non-linear or similar portions of the sky. In this case it is conceivable
temperature model. Comparison of the 15 min GHI retrievals that one could extract more information about cloud proper-
with CAMS data reveals a positive bias under clear skies of ties and irradiance compared to that obtained from just one
3.53 W m−2 (2.27 W m−2 ) for the linear (non-linear) model, system. Indeed, the rapid proliferation of PV installations
whereas under cloudy skies there is a negative bias of −9.90

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

5000 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

could make such multi-sensor inversions an interesting fu- as

ture prospect. P AC,mod
The ultimate goal of this work is to show that PV power   6
6
data can be used to infer global horizontal irradiance and = AηDCAC ηmodule T , Gtot,SW,τ , v wind Gtot,PV,τ
optical properties of the atmosphere, and the algorithm pre-  6

sented here is the first step in this direction. Although it is = AηDCAC ηmodule T , Gtot,SW,τ , v wind
clear from the above that this is feasible, moving towards op- 
6 6

erational use would require several further steps. The largest Gdir,PV,τ + Gdiff,PV,τ , (A1)
source of bias comes from the calibration step, in particu-
lar the effect of temperature. Access to direct current (DC) where A is the surface area of the PV system; ηDCAC is
data on the inverter level would allow for a much more ac- the converter efficiency; and the direct and diffuse compo-
curate extraction of the irradiance, without the confounding nents of the irradiance in the plane of array and underneath
6
effect of temperature, since the dependence of MPP current the glass covering (see Sect. A2) are given by Gdir,PV,τ and
on temperature is an order of magnitude smaller than that of 6
Gdiff,PV,τ , respectively. The temperature-dependent module
MPP voltage and power. Since inverter data are commonly efficiency is defined by (Evans and Florschuetz, 1977)
available in industry, this should be possible provided one has  
6
the legal right to access the data. Additionally, the data pre- ηmodule T , Gtot,SW,τ , v wind
processing needs to be automated. For instance, one needs to  
exclude PV systems and/or data subsets that do not meet cer- = ηmodule,n 1 − ζ (T module − Tn ) , (A2)
tain criteria such as shading of PV modules or inverter clip-
ping, and one needs to be able to detect clear-sky periods au- where ηmodule,n and ζ are the module efficiency and temper-
tomatically, even if the system orientation is unknown. This ature coefficient under standard test conditions (STC); i.e. at
could be done by developing a hybrid approach using both 6
Tn = 25 ◦ C, Gtot,SW,n = 1000 W m−2 with an air mass of 1.5.
physical modelling and artificial intelligence (AI) algorithms Note that in principle one could include a logarithmic depen-
for pattern recognition. In addition, the calibration procedure dence of the module efficiency on irradiance (Sauer, 1994),
itself could be augmented with AI to find the unknown pa- which is however not considered here. The module tempera-
rameters more effectively. ture is modelled using both the linear model (TamizhMani et
Once these aspects are overcome and appropriate agree- al., 2003) defined by
ments with industry partners are made for access to the data,
6
there should be nothing standing in the way of operational T module = u0 T ambient + u1 Gtot,SW,τ + u2 v wind
use. If this could be achieved at a large scale it would allow + u3 T sky (A3)
for a better characterisation of solar irradiance at the ground
and open up several possibilities for improving PV power and the non-linear model (Barry et al., 2020; Faiman, 2008)
forecasts at different time horizons. At shorter timescales defined by
(sub-hourly) one could use the additional information on ir- 6
radiance variability as further input to empirical forecasts Gtot,SW,τ
T module = T ambient +
based on statistical methods, whereas at longer timescales u1 + u2 v wind


(more than 3 h) these data could be assimilated into weather + u3 T sky − T ambient , (A4)
models. In order for this to make a difference one needs a
much larger dataset of PV systems for the analysis, which where v wind is the wind speed at 10 m above the ground,
↓
then requires further automation. The first steps in this direc- T 4sky = GLW /( σ ) defines the sky temperature, and u0,1,2,3
tion are currently being explored. are model parameters. Here an emissivity of  = 1 is as-
sumed, and σ is the Stefan–Boltzmann constant. Note that
Eqs. (A1), (A2), (A3), and (A4) can be combined into the
general non-linear expressions given in Eqs. (1) and (2) (see
for instance Skoplaki and Palyvos, 2009; Dows and Gough,
1995), where in this special case the coefficients are given by
Appendix A: Modelling details
b1 = s (1 + ζ 25), b2 = −u1 s ζ, b3 = −u0 s ζ,
A1 Parametric power model b4 = −u2 s ζ, b5 = −s ζ u3 (A5)
for the linear and
In Buchmann (2018) a simple model is proposed, based on u1
b10 = s (1 + ζ 25), b20 = − , b30 = s ζ (u3 − 1),
the combination of several different modelling steps from the sζ
literature (see for instance Skoplaki and Palyvos, 2009, or u2
Dows and Gough, 1995). Here the PV power output is written b40 = − , b50 = −s ζ u3 (A6)
sζ

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 5001

for the non-linear temperature model, where with τPV (0) means that the result is less sensitive to the
s ≡ A ηDCAC ηmodule,STC is a constant scaling factor. product κL and more to the angle of incidence. In the for-
The model equations in Eqs. (1) and (2) are used in this ward model the wavelength-integrated direct irradiance and
work, and the coefficients u0,1,2,3 , ζ , and s are allowed the diffuse radiance beams are each multiplied with the fac-
to vary freely, unless their a priori values are known from tor τ PV,rel (2) in order to take into account the attenuation
datasheets and/or from temperature modelling. In the cases due to the glass surface. The values of the extinction coef-
where measured temperature is available, the parameters ficient and thickness of the glass are fixed to κ = 4 m−1 and
u0,1,2,3 fall away. L = 0.002 m, respectively (De Soto et al., 2006), whereas the
effective refractive index n is allowed to vary (n = 1.526 with
A2 Optical model of glass covering an a priori error of 1 %). In principle one could also vary the
product κL that controls the absorption, but as mentioned
A2.1 Model formulation above the incident angle modifier approach is applicable to
a wide range of glass covers (Duffie and Beckman, 2013;
In order to model the optics of the glass surface of the PV
Klein, 1979).
modules the following equation for the transmission of pho-
tons as a function of the incident angle 2 is used (De Soto et A2.2 Optical model in forward model calibration
al., 2006; Sjerps-Koomen et al., 1996):
  In the forward model calibration, the transmission function
κL
τ PV (2) = exp − can then be used to calculate the direct and diffuse irradiance
cos 2r 6 6
Gdir,PV,τ and Gdiff,PV,τ as
" #!
1 sin2 (2r − 2) tan2 (2r − 2)
1− + , (A7) 6 cos 2 ↓
2 sin2 (2r + 2) tan2 (2r + 2) Gdir,PV,τ = G τ PV,rel (2) (A11)
cos θ 0 dir
where 2r is the angle of refraction from Snell’s law
and
(n sin 2r = sin 2), n is the index of refraction of glass, κ is
the glazing extinction coefficient, and L is the glazing thick- Z π/2−θ
2π+φ Z
ness. The incident angle 2 is the angle between the solar 6
Gdiff,PV,τ = L 0 0
diff cos 2 τ PV,rel (2 ) d ,
position vector and normal vector of the PV array, defined by
φ −θ


cos 2 = cos θ 0 cos θ + sin θ 0 sin θ cos φ 0 − φ , (A8) for cos 20 ≥ 0, (A12)
where θ is the angle of inclination of the PV array, φ is its where L diff is the diffuse radiance distribution calculated by
orientation, θ 0 is the solar zenith angle, and φ 0 is the solar DISORT. In this way there is no need for an empirical inci-
azimuth. dence angle modifier, since the direction of each diffuse pho-
In principle one should take into account the wavelength ton is explicitly described. The same formulation is used to
dependence of n and κ; however for most materials they are 6 6
calculate Gdir,SW,τ and Gdiff,SW,τ , i.e. over the whole wave-
relatively constant, with n increasing slightly at lower wave-
length range.
lengths, and since in practice the exact properties of the glass
covering for each system are unknown, it suffices to use the A2.3 Inversion of the optical model
effective values for all wavelengths.
The so-called incidence angle modifier (see also Duffie For the inversion of the PV model, two different methods
and Beckman, 2013) is defined by the ratio of the transmis- are used: for the extraction of cloud optical depth with DIS-
sion τ PV (2) and the transmission at normal incidence; i.e., ORT, the optical model can be explicitly taken into account
in the radiative transfer simulation, whereas for the direct ex-
τ PV (2)
τ PV,rel (2) ≡ , (A9) traction of GTI and its translation to GHI with the MYSTIC
τPV (0) lookup table, the empirical formulation (Duffie and Beck-
where man, 2013) for the effective angle for diffuse photons as a
  2  function of tilt angle θ ,
n−1
= e−κL 1−
τ PV (0) n+1
(A10) 2diff = 59.7 − 0.1388 θ + 0.001497 θ 2 , (A13)
4n
= e−κL (n+1)2
.
is used for all time points with a clearness index below 0.3
The use of a relative transmission coefficient is justified by so that the final inverted GTI is given by
the fact that the absolute transmittance is already taken into
6
account when characterising the solar cell under standard 6 Gtot,SW,inv,τ
conditions (Sjerps-Koomen et al., 1996). The normalisation Gtot,SW,inv = , (A14)
τ PV,rel,eff

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

5002 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

where The mismatch ratio remains at roughly 0.83 throughout

 the day, depending on water vapour column. In the case of
τ PV,rel (2dir ) if k i ≥ 0.3 PV12 in Fig. A1, one can see that in the mornings the ratio
τ PV,rel,eff = (A15)
τ PV,rel (2diff ) if k i < 0.3 is smaller, i.e. more red light, whereas in the evenings the
is the effective incidence angle modifier for clearness in- ratio is larger, since the panel is facing more to the south-
dex k i . east. This means that the panel detects more diffuse light in
the evenings. The behaviour at PV15 (Fig. A1) is the op-
A3 Spectral mismatch fitting procedure posite – in the morning the ratio is larger, since the panel
looks more to the south-west (mornings have a higher diffuse
The ratio of silicon PV irradiance to broadband irradiance as component, i.e. more blue light), whereas in the evenings it
defined in Eq. (5) is a function of the atmospheric composi- is smaller. However in the summer time when the sun goes
tion (primarily water vapour content) and angle of incidence down far to the north, the diffuse component again plays a
of the incoming solar beam. Indeed, the shape of the diurnal role so that the ratio becomes larger (right-hand plot). In gen-
variation in ξ spec,GTI is dependent on the azimuth angle of eral the variation is the greatest for larger SZAs, as seen in
the PV plant, as shown in Figs. A1 and A2, where the ratio the greyed-out areas in the plots.
is plotted for both GHI and GTI, for two different PV sys-
tems with different orientations, using libRadtran clear-sky
simulations from the different measurement campaigns. The
points in the upper (lower) plots in each figure are coloured
according to precipitable water (AOD), and it is evident that
for a given incident angle the water vapour plays the largest
role in determining the ratio.

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 5003

Figure A1. The spectral mismatch ratio at PV12 (azimuth angle φ ' 133◦ ) as a function of time (mean incident angle) as well as the water
vapour column (a, b) and AOD (c, d), for both 2018 (a, c) and 2019 (b, d) measurement campaigns. 2min is the minimum incident angle.
The solid and dashed black lines represent the mean mismatch ratio for GTI and GHI, respectively. The fit for GTI corresponding to Eq. A16
is shown as a dashed red line and is performed up to an SZA of 80◦ .

Figure A2. The same as Fig. A1 but for PV15 (azimuth angle φ ' 228◦ ).

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

5004 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

As shown in Figs. A1 and A2, the data are first grouped by resentation in this paper. While Copernicus Publications makes ev-
the time of the day in order to calculate the mean value, then ery effort to include appropriate place names, the final responsibility
split into two halves on either side of the minimum incident lies with the authors.
angle (2min ). Each branch can then be fitted with the function
 
x1 x2 Acknowledgements. The authors thank the DWD for the provision
hξ spec,GTI i = x0 exp − − , (A16)
cos 2 cos 22 of COSMO model data. James Barry would like to thank Lukas
Tirpitz for helpful comments and suggestions in the development
shown as the dashed red line in the plots. This shows the of this work. In addition, the authors thank the thorough review
general form of the diurnal variation in spectral mismatch work of the two anonymous referees, whose efforts significantly
and that for silicon PV the ratio is about 83 % for most of the contributed to improving the paper.
day.
In order to capture the effect of precipitable water and
AOD, the function Financial support. This research has been supported by the Bun-
desministerium für Wirtschaft und Energie (grant nos. 0350009A
ξ spec,GTI = p0 exp (−p1 [ H 2 O ] −p2 τ a ) (A17) and 0350009C).

is fitted to the irradiance data for each time step (aver-

aged over all days of the respective measurement campaign), Review statement. This paper was edited by Marloes Penning de
where [ H 2 O ] is the precipitable water from COSMO and τ a Vries and reviewed by two anonymous referees.
the AOD at 500 nm from AERONET. The fit coefficients are
then interpolated over the entire dataset in order to calculate
ξ spec,GTI at any time of the day. Future work will examine the
effects of clouds on the spectral mismatch factor – here the
clear-sky fit is applied to all data, which due to whitening of References
the skylight by clouds could lead to a bias in the final result.
Abe, C. F., Dias, J. B., Notton, G., and Poggi, P.: Computing So-
lar Irradiance and Average Temperature of Photovoltaic Modules
from the Maximum Power Point Coordinates, IEEE J. Photovolt.,
Code and data availability. Data are available for
10, 655–663, https://doi.org/10.1109/JPHOTOV.2020.2966362,
download from Zenodo (Barry et al., 2023b,
2020.
https://doi.org/10.5281/zenodo.8335791), and code
Alonso-Abella, M., Chenlo, F., Nofuentes, G., and Torres-
is available from Zenodo (Barry et al., 2023a,
Ramírez, M.: Analysis of spectral effects on the en-
https://doi.org/10.5281/zenodo.8336425), linked to GitHub.
ergy yield of different PV (photovoltaic) technologies:
The case of four specific sites, Energy, 67, 435–443,
https://doi.org/10.1016/j.energy.2014.01.024, 2014.
Author contributions. The two measurement campaigns were de- Baldauf, M., Seifert, A., Förstner, J., Majewski, D., Raschendor-
signed and coordinated with contributions from all authors, and fer, M., and Reinhardt, T.: Operational Convective-Scale Nu-
the installation and calibration of the various measurement devices merical Weather Prediction with the COSMO Model: Descrip-
were performed by NK, CS, RY, HD, JW, FG, PH, and MS. The tion and Sensitivities, Mon. Weather Rev., 139, 3887–3905,
PV model was conceptualised by JB, TB, KP, AHC, and SM; the https://doi.org/10.1175/MWR-D-10-05013.1, 2011.
software and simulations to implement and validate the model were Barry, J., Böttcher, D., Pfeilsticker, K., Herman-Czezuch, A., Kimi-
developed and carried out by JB, with contributions from JG, FG, aie, N., Meilinger, S., Schirrmeister, C., Deneke, H., Witthuhn,
and AHC. FG wrote several Python subroutines in order to import J., and Gödde, F.: Dynamic model of photovoltaic module tem-
weather model data and also developed the MYSTIC lookup table. perature as a function of atmospheric conditions, Adv. Sci. Res.,
CE and BM introduced several customised features into libRadtran 17, 165–173, https://doi.org/10.5194/asr-17-165-2020, 2020.
for this work; LS and MSH provided weather model and satellite Barry, J., Gödde, F., Grabenstein, J., and Herman-Czezuch, A.:
data. JB prepared the paper with contributions from all co-authors. jamesmhbarry/PVRAD: Supplement to “Irradiance and cloud
optical properties from solar photovoltaic systems”, Version
v1.0.0, Zenodo [code], https://doi.org/10.5281/zenodo.8336425,
Competing interests. At least one of the (co-)authors is a member 2023a.
of the editorial board of Atmospheric Measurement Techniques. The Barry, J., Meilinger, S., Pfeilsticker, K., Herman-Czezuch,
peer-review process was guided by an independent editor, and the A., Kimiaie, N., Schirrmeister, C., Yousif, R., Buchmann,
authors also have no other competing interests to declare. T., Grabenstein, J., Deneke, H., Witthuhn, J., Emde, C.,
Gödde, F., Mayer, B., Scheck, L., Schroedter-Homscheidt,
M., Hofbauer, P., and Struck, M.: Datasets for “Irradi-
Disclaimer. Publisher’s note: Copernicus Publications remains ance and cloud optical properties from solar photovoltaic
neutral with regard to jurisdictional claims made in the text, pub- systems” (final version), Version 2.0.0, Zenodo [data set],
lished maps, institutional affiliations, or any other geographical rep- https://doi.org/10.5281/zenodo.8335791, 2023b.

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 5005

Becker, R. and Behrens, K.: Quality assessment of heteroge- Faiman, D.: Assessing the Outdoor Operating Temperature of Pho-
neous surface radiation network data, Adv. Sci. Res., 8, 93–97, tovoltaic Modules, Progress in Photovoltaics: Research and Ap-
https://doi.org/10.5194/asr-8-93-2012, 2012. plications, 16, 307–315, https://doi.org/10.1002/pip.813, 2008.
Boers, R.: Simultaneous retrievals of cloud optical depth and Frank, C. W., Wahl, S., Keller, J. D., Pospichal, B., Hense, A.,
droplet concentration from solar irradiance and microwave liq- and Crewell, S.: Bias correction of a novel European reanalysis
uid water path, J. Geophys. Res.-Atmos., 102, 29881–29891, data set for solar energy applications, Sol. Energy, 164, 12–24,
https://doi.org/10.1029/97JD02494, 1997. https://doi.org/10.1016/j.solener.2018.02.012, 2018.
Boers, R., van Lammeren, A., and Feijt, A.: Accuracy of Cloud Op- Halilovic, S., Bright, J. M., Herzberg, W., and Killinger, S.:
tical Depth Retrievals from Ground-Based Pyranometers, J. At- An analytical approach for estimating the global horizontal
mos. Ocean. Tech., 17, 916–927, https://doi.org/10.1175/1520- from the global tilted irradiance, Sol. Energy, 188, 1042–1053,
0426(2000)017<0916:AOCODR>2.0.CO;2, 1999. https://doi.org/10.1016/J.SOLENER.2019.06.027, 2019.
Buchmann, T.: Potenzial von Photovoltaikanlagen zur Herman-Czezuch, A., Mekeng, A. Z., Meilinger, S., Barry, J., and
Ableitung raum-zeitlich hoch aufgelöster Global- Kimiaie, N.: Impact of aerosols on photovoltaic energy pro-
strahlungsdaten, PhD thesis, Heidelberg University, duction using a spectrally resolved model chain: Case study
https://doi.org/10.11588/heidok.00024687, 2018. of southern West Africa, Renewable Energy, 194, 321–333,
Buras, R., Dowling, T., and Emde, C.: New secondary-scattering https://doi.org/10.1016/j.renene.2022.04.166, 2022.
correction in DISORT with increased efficiency for for- Hess, M., Koepke, P., and Schult, I.: Optical prop-
ward scattering, J. Quant. Spectrosc. Ra., 112, 2028–2034, erties of aerosols and clouds, B. Am. Meteorol.
https://doi.org/10.1016/j.jqsrt.2011.03.019, 2011. Soc., 79, 831–844, https://doi.org/10.1175/1520-
Carrasco, M., Mancilla-David, F., and Ortega, R.: An Estima- 0477(1998)079<0831:OPOAAC>2.0.CO;2, 1998.
tor of Solar Irradiance in Photovoltaic Arrays With Guaran- Holben, B. N., Eck, T. F., Slutsker, I., Tanré, D., Buis, J.
teed Stability Properties, IEEE T. Ind. Electron., 61, 3359–3366, P., Setzer, A., Vermote, E., Reagan, J. A., Kaufman, Y.
https://doi.org/10.1109/TIE.2013.2281154, 2014. J., Nakajima, T., Lavenu, F., Jankowiak, I., and Smirnov,
Črnivec, N. and Mayer, B.: Quantifying the bias of radiative A.: AERONET—A Federated Instrument Network and Data
heating rates in numerical weather prediction models for shal- Archive for Aerosol Characterization, Remote Sens. Envi-
low cumulus clouds, Atmos. Chem. Phys., 19, 8083–8100, ron., 66, 1–16, https://doi.org/10.1016/S0034-4257(98)00031-5,
https://doi.org/10.5194/acp-19-8083-2019, 2019. 1998.
De Soto, W., Klein, S. A., and Beckman, W. A.: Im- Holm, L. M.: Photovoltaik in Deutschland, Strom-Report, https://
provement and validation of a model for photo- strom-report.com/photovoltaik/ (last access: 17 October 2023),
voltaic array performance, Sol. Energy, 80, 78–88, 2023.
https://doi.org/10.1016/j.solener.2005.06.010, 2006. Hu, Y. X. and Stamnes, K.: An Accurate Parameterization of the Ra-
Deneke, H.: Influence of clouds on the solar radiation diative Properties of Water Clouds Suitable for Use in Climate
budget, PhD thesis, Rheinische Friedrich-Wilhelms- Models, J. Climate, 6, 728–742, https://doi.org/10.1175/1520-
Universität Bonn, https://www.knmi.nl/research/publications/ 0442(1993)006<0728:AAPOTR>2.0.CO;2, 1993.
influence-of-clouds-on-the-solar-radiation-budget (last access: Huld, T., Friesen, G., Skoczek, A., Kenny, R. P., Sample, T., Field,
16 October 2023), 2002. M., and Dunlop, E. D.: A power-rating model for crystalline
Dows, R. N. and Gough, E. J.: PVUSA model technical spec- silicon PV modules, Sol. Energ. Mat. Sol. C., 95, 3359–3369,
ification for a turnkey photovoltaic power system, Depart- https://doi.org/10.1016/j.solmat.2011.07.026, 2011.
ment of Energy, Technical Report no. DOE/AL/82993-27, Kato, S., Ackerman, T. P., Mather, J. H., and Clothiaux, E. E.: The k-
https://doi.org/10.2172/172103, 1995. distribution method and correlated-k approximation for a short-
Duffie, J. A. and Beckman, W. A.: Radiation Transmission through wave radiative transfer model, J. Quant. Spectrosc. Ra., 62, 109–
Glazing: Absorbed Radiation, John Wiley & Sons, Ltd, Chap. 5, 121, https://doi.org/10.1016/S0022-4073(98)00075-2, 1999.
202–235, https://doi.org/10.1002/9781118671603.ch5, 2013. Killinger, S., Braam, F., Müller, B., Wille-Haussmann, B.,
Elsinga, B., van Sark, W., and Ramaekers, L.: Inverse photovoltaic and McKenna, R.: Projection of power generation between
yield model for global horizontal irradiance reconstruction, En- differently-oriented PV systems, Sol. Energy, 136, 153–165,
ergy Sci. Eng., 5, 226–239, https://doi.org/10.1002/ese3.162, https://doi.org/10.1016/J.SOLENER.2016.06.075, 2016.
2017. King, D. L., Boyson, W. E., and Kratochvil, J. A.: Pho-
Emde, C., Buras-Schnell, R., Kylling, A., Mayer, B., Gasteiger, J., tovoltaic array performance model, Technical Report
Hamann, U., Kylling, J., Richter, B., Pause, C., Dowling, T., Number: SAND2004-3535, Sandia National Laboratories,
and Bugliaro, L.: The libRadtran software package for radia- https://doi.org/10.2172/920449, 2004.
tive transfer calculations (version 2.0.1), Geosci. Model Dev., 9, King, D. L., Gonzalez, S., Galbraith, G. M., and Boyson, W. E.:
1647–1672, https://doi.org/10.5194/gmd-9-1647-2016, 2016. Performance Model for Grid-Connected Photovoltaic Inverters,
Engerer, N. and Mills, F.: KPV: A clear-sky in- Technical Report Number: SAND2007-5036, Sandia National
dex for photovoltaics, Sol. Energy, 105, 679–693, Laboratories, https://doi.org/10.2172/920449, 2007.
https://doi.org/10.1016/J.SOLENER.2014.04.019, 2014. Klein, S.: Calculation of the monthly-average transmittance-
Evans, D. and Florschuetz, L.: Cost studies on terrestrial photo- absorptance product, Sol. Energy, 23, 547–551,
voltaic power systems with sunlight concentration, Sol. Energy, https://doi.org/10.1016/0038-092X(79)90083-5, 1979.
19, 255–262, https://doi.org/10.1016/0038-092X(77)90068-8, Klüser, L., Killius, N., and Gesell, G.: APOLLO_NG –
1977. a probabilistic interpretation of the APOLLO legacy for

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023

5006 J. Barry et al.: Irradiance and atmospheric optical properties from PV power

AVHRR heritage channels, Atmos. Meas. Tech., 8, 4155–4170, P., and Rindt, B.: Entwicklung innovativer satellitengestützter
https://doi.org/10.5194/amt-8-4155-2015, 2015. Methoden zur verbesserten PV-Ertragsvorhersage auf ver-
Kriebel, K. T., Gesell, G., Kästner, M., and Mannstein, schiedenen Zeitskalen für Anwendungen auf Verteilnetzebene
H.: The cloud analysis tool APOLLO: Improvements BT – Schlussbericht, 1–180, https://doi.org/10.18418/opus-
and validations, Int. J. Remote Sens., 24, 2389–2408, 5955, 2021b.
https://doi.org/10.1080/01431160210163065, 2003. Mertens, K.: Photovoltaics: Fundamentals, Technology and Prac-
Laudani, A., Fulginei, F. R., Salvini, A., Carrasco, M., and tice, 1st edn., John Wiley & Sons Ltd, ISBN 978-1-118-63416-5,
Mancilla-David, F.: A fast and effective procedure for sens- 2014.
ing solar irradiance in photovoltaic arrays, in: 2016 IEEE 16th Nakajima, T. and King, M. D.: Determination of the Opti-
International Conference on Environment and Electrical Engi- cal Thickness and Effective Particle Radius of Clouds from
neering (EEEIC), Florence, Italy, 7–10 June 2016, IEEE, 1–4, Reflected Solar Radiation Measurements. Part I: Theory,
https://doi.org/10.1109/EEEIC.2016.7555541, 2016. J. Atmos. Sci., 47, 1878–1893, https://doi.org/10.1175/1520-
Leontyeva, E. and Stamnes, K.: Estimations of Cloud 0469(1990)047<1878:DOTOTA>2.0.CO;2, 1990.
Optical Thickness from Ground-Based Measure- Nespoli, L. and Medici, V.: An unsupervised method for
ments of Incoming Solar Radiation in the Arctic, estimating the global horizontal irradiance from photo-
J. Climate, 7, 566–578, https://doi.org/10.1175/1520- voltaic power measurements, Sol. Energy, 158, 701–710,
0442(1994)007<0566:EOCOTF>2.0.CO;2, 1994. https://doi.org/10.1016/J.SOLENER.2017.10.039, 2017.
Mandija, F., Chavez-Perez, V. M., Nieto, R., Sicard, M., Ohmura, A., Dutton, E. G., Forgan, B., Fröhlich, C., Gilgen, H.,
Danylevsky, V., Añel, J. A., and Gimeno, L.: The climatol- Hegner, H., Heimo, A., König-Langlo, G., McArthur, B., Müller,
ogy of dust events over the European continent using data G., Philipona, R., Pinker, R., Whitlock, C. H., Dehne, K., and
of the BSC-DREAM8b model, Atmos. Res., 209, 144–162, Wild, M.: Baseline Surface Radiation Network (BSRN/WCRP):
https://doi.org/10.1016/j.atmosres.2018.03.006, 2018. New Precision Radiometry for Climate Research, B. Am.
Marion, B. and Smith, B.: Photovoltaic system derived data for Meteorol. Soc., 79, 2115–2136, https://doi.org/10.1175/1520-
determining the solar resource and for modeling the perfor- 0477(1998)079<2115:BSRNBW>2.0.CO;2, 1998.
mance of other photovoltaic systems, Sol. Energy, 147, 349–357, Perez, R., Ineichen, P., Maxwell, E., Seals, R., and Zelenka, A.: Dy-
https://doi.org/10.1016/J.SOLENER.2017.03.043, 2017. namic global-to-direct irradiance conversion model, ASHRAE
Matricardi, M.: The inclusion of aerosols and clouds in RTIASI, Tran., 98, 354–369, 1992.
the ECMWF fast radiative transfer model for the infrared atmo- Petty, G. W.: A First Course in Atmospheric Radiation,
spheric sounding interferometer, ECMWF, Technical memoran- 2nd edn., Sundog Publishing, Madison, Wisconsin, ISBN-
dum no. 474, https://doi.org/10.21957/1krvb28ql, 2005. 13: 9780972903318, 2006.
Mayer, B.: Radiative transfer in the cloudy atmosphere, EPJ Web Qu, Z., Oumbe, A., Blanc, P., Espinar, B., Gesell, G., Gschwind,
Conf., 1, 75–99, https://doi.org/10.1140/epjconf/e2009-00912-1, B., Klüser, L., Lefèvre, M., Saboret, L., Schroedter-Homscheidt,
2009. M., and Wald, L.: Fast radiative transfer parameterisation for as-
Mayer, B. and Kylling, A.: Technical note: The libRadtran soft- sessing the surface solar irradiance: The Heliosat-4 method, Me-
ware package for radiative transfer calculations – description teorol. Z., 26, 33–57, https://doi.org/10.1127/metz/2016/0781,
and examples of use, Atmos. Chem. Phys., 5, 1855–1877, 2017.
https://doi.org/10.5194/acp-5-1855-2005, 2005. Rhodes, B.: PyEphem: Scientific-grade astronomy routines
McArthur, L. J. B.: Baseline Surface Radiation Network (BSRN), for Python, GitHub [code], https://github.com/brandon-rhodes/
Operations Manual, Version 2.1, World Climate Research pyephem (last access: 17 October 2023), 2022.
Programme, Downsview, Ontario, WCRP-121, WMO/TD- Rivera Aguilar, M. J. and Reise, C.: Silicon Sensors vs.
No. 1274, 1–176, https://bsrn.awi.de/fileadmin/user_upload/ Pyranometers – Review of Deviations and Conversion of
bsrn.awi.de/Publications/McArthur.pdf (last access: 26 June Measured Values, in: 37th European Photovoltaic Solar
2023), 2005. Energy Conference and Exhibition, Online, 1449–1454,
Meilinger, S., Herman-Czezuch, A., Kimiaie, N., Schirrmeister, C., https://doi.org/10.4229/EUPVSEC20202020-5BV.3.3, 2020.
Yousif, R., Geiss, S., Scheck, L., Weissmann, M., Gödde, F., Rodgers, C. D.: Inverse Methods for Atmospheric Sounding, World
Mayer, B., Zinner, T., Barry, J., Pfeilsticker, K., Kraiczy, M., Scientific, 2, 256 pp., https://doi.org/10.1142/3171, 2000.
Winter, K., Altayara, A., Reise, C., Rivera, M., Deneke, H., Wit- Saint-Drenan, Y.-M.: A Probabilistic Approach to the Estimation of
thuhn, J., Betcke, J., Schroedter-Homscheidt, M., Hofbauer, P., Regional Photovoltaic Power Generation using Meteorological
and Rindt, B.: Development of innovative satellite-based meth- Data Application of the Approach to the German Case, Disser-
ods for improved PV yield prediction on different time scales tation, University of Kassel, http://nbn-resolving.de/urn:nbn:de:
for distribution grid level applications (MetPVNet), Hochschule hebis:34-2016090550868 (last access: 17 October 2023), 2015.
Bonn-Rhein-Sieg, Internationales Zentrum für Nachhaltige En- Saint-Drenan, Y. M., Bofinger, S., Fritz, R., Vogt, S.,
twicklung (IZNE), Sankt Augustin, IZNE Working Paper Series Good, G. H., and Dobschinski, J.: An empirical ap-
No. 21/4, https://doi.org/10.18418/978-3-96043-094-0, 2021a. proach to parameterizing photovoltaic plants for power
Meilinger, S., Herman-Czezuch, A., Kimiaie, N., Schirrmeister, C., forecasting and simulation, Sol. Energy, 120, 479–493,
Yousif, R., Geiss, S., Scheck, L., Weissmann, M., Gödde, F., https://doi.org/10.1016/j.solener.2015.07.024, 2015.
Mayer, B., Zinner, T., Barry, J., Pfeilsticker, K., Kraiczy, M., Sauer, D. U.: Untersuchungen zum Einsatz und Entwicklung
Winter, K., Altayara, A., Reise, C., Rivera, M., Deneke, H., von Simulationsmodellen für die Auslegung von Photovoltaik-
Witthuhn, J., Betcke, J., Schroedter-Homscheidt, M., Hofbauer,

Atmos. Meas. Tech., 16, 4975–5007, 2023 https://doi.org/10.5194/amt-16-4975-2023

J. Barry et al.: Irradiance and atmospheric optical properties from PV power 5007

Systemen, Diplomarbeit, Technische Hochschule Darmstadt, Stamnes, K., Tsay, S.-C., Wiscombe, W., and Jayaweera, K.: Nu-
https://doi.org/10.13140/RG.2.1.1833.7366, 1994. merically stable algorithm for discrete-ordinate-method radiative
Schade, N. H., Macke, A., Sandmann, H., and Stick, C.: En- transfer in multiple scattering and emitting layered media, Appl.
hanced solar global irradiance during cloudy sky condi- Optics, 27, 2502, https://doi.org/10.1364/AO.27.002502, 1988.
tions, Meteorol. Z., 16, 295–303, https://doi.org/10.1127/0941- Stephens, G. L., Ackerman, S., and Smith, E. A.: A Short-
2948/2007/0206, 2007. wave Parameterization Revised to Improve Cloud Absorp-
Scheck, L., Weissmann, M., and Mayer, B.: Efficient Methods to tion, J. Atmos. Sci., 41, 687–690, https://doi.org/10.1175/1520-
Account for Cloud-Top Inclination and Cloud Overlap in Syn- 0469(1984)041<0687:ASPRTI>2.0.CO;2, 1984.
thetic Visible Satellite Images Efficient Methods to Account TamizhMani, G., Ji, L., Tang, Y., Petacci, L., and Osterwald,
for Cloud-Top Inclination and Cloud Overlap in Synthetic Vis- C.: Photovoltaic Module Thermal/Wind Performance: Long-
ible Satellite Images, J. Atmos. Ocean. Tech., 35, 665–685, Term Monitoring and Model Development for Energy Rating,
https://doi.org/10.1175/JTECH-D-17-0057.1, 2018. NCPV and Solar Program Review Meeting Proceedings, Den-
Schroedter-Homscheidt, M., Azam, F., Betcke, J., Hanrieder, N., ver, Colorado 24–26 March 2003, CD-ROM, https://www.osti.
Lefèvre, M., Saboret, L., and Saint-Drenan, Y.: Surface solar ir- gov/biblio/15006842 (last access: 17 October 2023), 2003.
radiation retrieval from MSG/SEVIRI based on APOLLO Next Urraca, R., Huld, T., Gracia-Amillo, A., Martinez-de Pison, F. J.,
Generation and HELIOSAT-4 methods, Meteorol. Z., 31, 455– Kaspar, F., and Sanz-Garcia, A.: Evaluation of global horizontal
476, https://doi.org/10.1127/metz/2022/1132, 2022. irradiance estimates from ERA5 and COSMO-REA6 reanalyses
Scolari, E., Sossan, F., and Paolone, M.: Photovoltaic-model-based using ground and satellite-based data, Sol. Energy, 164, 339–354,
solar irradiance estimators: Performance comparison and appli- https://doi.org/10.1016/j.solener.2018.02.059, 2018.
cation to maximum power forecasting, IEEE T. Sustain. Energ., Yang, D. and Bright, J. M.: Worldwide validation of 8
9, 35–44, https://doi.org/10.1109/TSTE.2017.2714690, 2018. satellite-derived and reanalysis solar radiation prod-
Shockley, W.: The theory of p-n junctions in semiconductors ucts: A preliminary evaluation and overall metrics for
and p-n junction transistors, Bell Syst. Tech. J., 28, 435–489, hourly data over 27 years, Sol. Energy, 210, 3–19,
https://doi.org/10.1002/j.1538-7305.1949.tb03645.x, 1949. https://doi.org/10.1016/j.solener.2020.04.016, 2020.
Sjerps-Koomen, E. A., Alsema, E. A., and Turkenburg, W. C.: A Zubler, E. M., Lohmann, U., Lüthi, D., and Schär, C.: Inter-
simple model for PV module reflection losses under field condi- comparison of aerosol climatologies for use in a regional cli-
tions, Sol. Energy, 57, 421–432, https://doi.org/10.1016/S0038- mate model over Europe, Geophys. Res. Lett., 38, L15705,
092X(96)00137-5, 1996. https://doi.org/10.1029/2011GL048081, 2011.
Skoplaki, E. and Palyvos, J.: On the temperature dependence
of photovoltaic module electrical performance: A review
of efficiency/power correlations, Sol. Energy, 83, 614–624,
https://doi.org/10.1016/J.SOLENER.2008.10.008, 2009.

https://doi.org/10.5194/amt-16-4975-2023 Atmos. Meas. Tech., 16, 4975–5007, 2023