Solar Energy: Christian A. Gueymard

Solar Energy xxx (2017) xxxxxx
Contents lists available at ScienceDirect
Solar Energy
journal homepage: www.elsevier.com/locate/solener
Cloud and albedo enhancement impacts on solar irradiance using high-

frequency measurements from thermopile and photodiode radiometers.
Part 2: Performance of separation and transposition models for global
tilted irradiance
Christian A. Gueymard
Solar Consulting Services, Colebrook, NH, USA
a r t i c l e i n f o a b s t r a c t
Article history: Based on high-accuracy irradiance measurements at a high-elevation station frequently affected by snow
Received 7 March 2017 cover, this experimental study evaluates the interannual variability in high global tilted irradiance
Received in revised form 22 April 2017 (>1 sun) incident on either latitude-tilt (40) or vertical-tilt (90) radiometers at 1-min resolution.
Accepted 27 April 2017
Using a 10-year time series, this variability is found substantial, particularly for the 90 tilt. The perfor-
Available online xxxx
mance of five separation and seven transposition models is also analyzed in general, and most specifically
under cloud and/or albedo enhancement events. The separation models performance degrades rapidly
Keywords:
for clearness indices larger than 0.8, to the point that three models tend to predict zero direct normal irra-
Tilted irradiance
Albedo
diance when it is actually high. Only one model (Engerer) can predict acceptable results, even though
Cloud enhancement negatively biased under such conditions. All transposition models are also impacted by enhancement
Solar radiation models events. This is most particularly the case for one of them (Perez), which tends to predict an extremely
low sky-diffuse component during those events, and even negative values in the case of the 90 tilt.
An analysis of the models performance as a function of the clearness index reveals that most models
are affected by a rapid degradation of performance when this index is larger than 0.8. For the 40 tilt
and on average over 10 years, the bias of the CDRS model is found reasonably low and stable even when
the index approaches its maximum value. The maximum recorded value of GTI is 2000 W m2 (2 suns)
for the 40 tilt, using 1-s data. For the 90 tilt, no model has a low and stable bias under all possible con-
ditions, but the CDRS model still performs reasonably well under high clearness index conditions. All
these findings confirm the fact that separation or transposition models that were empirically developed
based on hourly irradiance data do not necessarily respond correctly to transient enhancement
situations.
2017 Elsevier Ltd. All rights reserved.
1. Introduction models in these two categories are empirical and were designed
decades ago based on hourly radiation data. Current measure-
The need for detailed solar resource assessments and photo- ments, however, are now typically done at 1-min resolution or
voltaic (PV) system output simulations at ever-increasing spatial better.
and temporal resolutions is presenting challenges for the existing Part 1 of this study highlighted the relatively frequent occur-
radiation models in common use today for plane-of-array (POA) rence of transient cloud enhancement (CE) effects, potentially aug-
calculations. There are typically two categories of such models, mented by albedo enhancement (AE) effects, when considering
each one performing an essential task in the overall process: (i) measurements of GHI at a time step of 1 min. For instance, at the
separation of the direct and diffuse components from global hori- Golden, Colorado site under scrutiny, GHI was found >1 sun (i.e.,
zontal irradiance (GHI); and (ii) transposition of these components 1 kW m2) 3.8% (5.7%) of the time, and 0.6% (1.1%) larger than its
from horizontal to the tilted POA geometry to ultimately obtain the extraterrestrial counterpart, when using thermopile (photodiode)
global tilted irradiance (GTI). Most, if not all, industry-standard radiometers, respectively, on average over 10 years and at 1-min
resolution. During the same period, an overall maximum GHI of
1.546 kW m2 was recorded with a thermopile pyranometer, and
E-mail address: Chris@SolarConsultingServices.com 1.634 kW m2 with a photodiode sensor, indicating that the type
http://dx.doi.org/10.1016/j.solener.2017.04.068
0038-092X/ 2017 Elsevier Ltd. All rights reserved.
Please cite this article in press as: Gueymard, C.A. Cloud and albedo enhancement impacts on solar irradiance using high-frequency measurements from
thermopile and photodiode radiometers. Part 2: Performance of separation and transposition models for global tilted irradiance. Sol. Energy (2017), http://
dx.doi.org/10.1016/j.solener.2017.04.068
2 C.A. Gueymard / Solar Energy xxx (2017) xxxxxx
of sensor is an important consideration when reporting CE effects. studies did not consider enhancement effects or high-frequency
Similarly, the measurement time step also appears to matter, since data, which justifies additional scrutiny.
even larger maximum values (up to 1.9 suns) were obtained at 1-s In addition to CE effects, GTI can also be impacted by albedo
resolution. Part 1 described three types of circumstances poten- enhancement (AE) effects caused by larger-than-usual ground
tially leading to CE events, with or without contribution from albedo (e.g., over snow-covered ground) and/or large deviations
simultaneous AE events. In this Part 2, the main focus is on high- from the conventionally assumed isotropy of ground reflection. A
GTI cases, irrespective of the circumstances causing them, but with large regional surface albedo increases DIF and GHI through atmo-
additional scrutiny on the impact of high-albedo conditions. spheric backscattering, as discussed in Part 1. A separate phe-
Some important questions remain to be addressed in detail, nomenon is the increase in GTI caused by high local albedo
most particularly: Do transient effects impact GTI more than facing solar collectors and by the associated local non-
GHI? How large GTI can be under very short transient periods? Lambertian reflection processes; see, e.g., (Andrews and Pearce,
How do separation and transposition models of the literature, 2013; Gueymard, 1987, 2009; Ineichen, 1990; Kierkus and
which were designed for coarser hourly simulations, behave at Colbrone, 1989; Skartveit et al., 1998; Temps and Coulson, 1977;
higher frequency? What kind of instrumentation is necessary to Weiser et al., 2016; Yoon et al., 2014). As discussed in Part 1, diffi-
evaluate the performance of such models under rapidly changing culties arise because the local and regional albedos can be widely
conditions? What is the ideal measurement frequency for these different at any moment. So far, the literature does not seem to
measurements? And how do these enhancement events impact have considered the combination of CE and AE effects on GTI,
the power output of PV systems and their inverters? which justifies the present study.
This studys Part 2 intends to address these questions, except The discussion above demonstrates the importance of quantify-
the last one (which is of a very different nature), and in so doing ing how CE, regional-AE and/or local-AE events impact the separa-
provide more background knowledge about enhancement effects tion and transposition models that are customarily used to
in the context of PV applications. Like in Part 1, the main focus is ultimately convert GHI into GTI. To that end, this study critically
on the description of high-irradiance events, since only these can evaluates the high-frequency performance of a variety of both
negatively impact the design, safety and reliability of PV systems types of model under various surface conditions, as well as highly
(Chen et al., 2013; Luoma et al., 2012). The possible impacts (pos- variable cloud conditions. So far the literature has reported a
itive or negative) on overall PV performance are still unclear, due to record GTI of 1.6 suns in Norway (Yordanov et al., 2015). In Califor-
conflicting effects (Burger and Rther, 2006; Luoma et al., 2012; nia, several 1-s periods with GTI > 1.5 suns have been reported by
chsner et al., 2013; Zehner et al., 2011). The reader is referred (Luoma et al., 2012). In contrast, Part 1 of this study has uncovered
to Part 1 for a literature review about the connection between a record 1-s GHI measurement of 1.9 suns, and frequent 1-min
enhancement events and their potential effects on PV systems. GHI measurements above 1.5 suns. Since GTI is generally larger
Recent contributions (Cucumo et al., 2007; Gueymard, 2009; than GHI, this Part 2 investigates the corresponding peaks in GTI,
Lave et al., 2015; Yang et al., 2013) have started to investigate in an effort to estimate the frequency of very high GTI values that
how the different possible combinations of specific separation can impact PV systems, and to evaluate whether routine GHI-to-
and transposition models affected the accuracy of the modeled GTI calculations can provide the correct magnitude of extreme
GTI. Interestingly, a general finding is that the accuracy of the first GTI events.
step (component separation) is what conditions in large part the
uncertainty in the GTI predictions (component transposition). As
2. Separation models
a consequence, the transposition model with the highest intrinsic
performance might not provide the best GTI accuracy, due to the
A recent study (Gueymard and Ruiz-Arias, 2016) evaluated the
potentially incorrect balance between the estimated direct and dif-
performance of 140 separation models under various climatic con-
fuse components of GHI resulting from the first step. Since com-
ditions. Using irradiance data at 54 research-class stations, two
pensations of errors might be highly location specific, it is argued
important findings of that study were that (i) some separation
here that the best avenue, with the highest chances for universal
models that had been developed from hourly data showed signifi-
validity, is to separately look for the most accurate models in each
cant issues when 1-min input data were used instead; and (ii) most
category. This means that the first step (component separation)
of these issues were triggered by CE and/or AE effects. This justifies
must provide the direct horizontal irradiance (DHI)or, alterna-
the more specific analysis undertaken here. However, to maintain
tively, the direct normal irradiance (DNI)and the diffuse horizon-
it within reasonable limits, only five separation models are consid-
tal irradiance (DIF) with minimal error.
ered in what follows. For clarity, reference to these models utilizes
From that perspective, a recent study (Gueymard and Ruiz-
the specific notation (SMALL CAPS) that was introduced in (Gueymard
Arias, 2016) has shown that hourly separation models do not
and Ruiz-Arias, 2016) and later also adopted in (Yang, 2016) for the
always produce correct results when using input data of GHI at
case of transposition models. The popular ERBS, MAXWELL, PEREZ1 and
1-min resolution, most particularly under transient cloud situa-
PEREZ2 models (Erbs et al., 1982; Maxwell, 1987; Perez et al.,
tions or over highly reflective surfaces. What has just been men-
tioned justifies further investigation into the effects of rapid 1992, 2002) were developed from hourly data, whereas the ENGERER
fluctuations in GHI on the modeled GTI due to CE and/or AE epi- model (Engerer, 2015) was developed from 1-min data. The latter
sodes, which can typically last from a fraction of a second to sev- demonstrated the best performance results overall in the afore-
eral minutes. The present study intends to bring a new mentioned study (Gueymard and Ruiz-Arias, 2016). ERBS and
perspective and a quantitative assessment on the impact of CE ENGERER provide an estimate of the diffuse fraction, K = DIF/GHI,
and/or AE effects on the modeled GTI. So far, most of the literature from which DHI can be simply derived using the closure equation,
has focused on the CE effects on GHI, as discussed in Part 1. How- DHI = GHIDIF = GHI (1 K). In contrast, MAXWELL, PEREZ1 and PEREZ2
ever, the ramping effects on inverters and PV power output are evaluate the direct transmittance, Kn = DNI/ETN, where ETN is the
actually dependent on GTI rather than GHI, since the vast majority extraterrestrial irradiance at normal incidence, and then finally
of solar panels are mounted at a tilt rather than horizontally. The DHI = DNI cos Z and DIF = GHI DHI, where Z is the suns zenith
horizontal-to-tilted solar irradiance transposition has been shown angle. Both ENGERER and PEREZ2 require estimates of the clear-sky
to directly impact the modeling of PV array performance (Polo irradiances, which are simply obtained here through the Perez-
et al., 2016; Roberts et al., 2017; Yoshida et al., 2013), but these Ineichen clear-sky radiation model (Perez et al., 2002), per the
C.A. Gueymard / Solar Energy xxx (2017) xxxxxx 3
current practice (Gueymard and Ruiz-Arias, 2016; Perez et al., The simple formulas described by Eqs. (3) and (5) became pop-
2002). As stated in Part 1, this fast and simple model is operated ular following the seminal work of Liu and Jordan (1963), to whom
with mean-monthly Linke turbidity coefficients from a high- they are usuallybut incorrectlyattributed. The isotropic trans-
resolution worldwide database (SoDa service, http://www.soda- position model was mentioned earlier for solar applications
is.com/eng/services/climat_free_eng.php#c5). (Kondratyev and Manolova, 1960), but its first demonstration
appears to date back two decades earlier, targeting building illumi-
3. Transposition models nation applications (Moon and Spencer, 1942). The formula in
Eq. (5) was confirmed recently using numerical methods (Xie and
The transposition procedure consists in evaluating the direct Sengupta, 2016). In contrast to the simplicity of ISOTROPIC, and even
and diffuse components incident on the tilted plane of array, pro- of most of the other models considered here, PEREZ appears much
vided their horizontal counterparts are known. The sum of the more intricate, considering its larger number of equations and its
tilted components is the desired result, GTI, hereafter symbolized voluminous quantity of numerical coefficients. To avoid any misin-
as ES. The direct component on the tilted plane is terpretation, the original Fortran code was obtained from its author
and used here directly.
EbS Ebn Maxcos h; 0 Eb Maxcos h= cos Z; 0 1
where Ebn stands for DNI, Eb for DHI, and h stands for the suns angle 4. Instrumentation
of incidence on the tilted surface.
The ground-reflected component is conventionally estimated The experimental data considered here originates from the
from same site as in Part 1, i.e., the Solar Radiation Research Laboratory
ErS q Rr E 2 (SRRL), located on South Table Mountain (latitude 39.742N, longi-
tude 105.179W, elevation 1829 m), a mesa overlooking the cam-
where q is the local surface albedo facing the collector (within pus of the National Renewable Energy Laboratory (NREL) in
10 m from it), E is GHI, and Rr is the reflected diffuse transposition Golden, Colorado.
function. In the vast majority of transposition calculations, the In addition to the instrumentation described in Part 1, some
reflection process is assumed Lambertian, the foreground is tilted radiometers are also considered here. These include pyra-
assumed infinite, and no shading is assumed to affect the compo- nometers to observe the 1-min latitude-tilt (40) global irradiance
nents of GHI incident on the reflective surface. Under those highly (GTI40) and the vertical global irradiance (GTI90), as well as photo-
ideal conditions, one obtains diode sensors to observe the 1-min and 1-s GTI40 irradiance. The
Rr 1 cos S=2; 3 1-min dataset covers the 10-year period 20062015, whereas the
1-s dataset covers only the 27-month period between Feb. 2010
where S is the tilt angle from horizontal.
and Apr. 2012, with however a few additional days in 2015. Details
Finally, the tilted sky diffuse component is obtained from
on all the instrumentation used here are summarized in Table 1.
EdS Rd Ed 4 Note the changes of line-up that occurred over time. The redun-
dancy in DNI and DIF sensors serves an important purpose in the
where Ed represents DIF and Rd the sky diffuse transposition func-
stringent quality control procedure (detailed in Part 1) that is
tion. Contrarily to the case described in Eq. (1), Rd is typically not
applied to the GHI, DIF and DNI data. After quality control, only
a pure geometric function, and is exactly what transposition models
data from the primary radiometer is used, and GHI is calculated
attempt to evaluate. The present experimental setup does not pro-
from the DNI and DIF observations from the closure equation. All
vide a way to measure EdS directly. Nevertheless, its observed value
tilted sensors considered here are oriented toward the equator
can be approximated using EdS = ES EbS ErS.
(i.e., due south), and face a horizontal foreground. The albedo con-
A large selection of transposition models that provide Rd has
tribution from the slopes of the Rocky Mountains (mostly to the
been recently reviewed in (Yang, 2016). A subset of these models,
west) can be assumed negligible because of distance, even if they
which also showed good performance in a previous study
are snow covered. The albedo is obtained by dividing the upwelling
(Gueymard, 2009), is considered here. It consists of the models
by the downwelling global irradiance at close proximity from the
referred to below as HAY (Hay, 1979), MUNEER (Muneer, 2004), PEREZ
GTI40 and GTI90 sensors. All low irradiance values (less than
(Perez et al., 1990), HDKR (Duffie and Beckman, 1991), and CDRS
100 W m2 for GHI and GTI40, and 10 W m2 for GTI90) are elimi-
(Gueymard, 1987). Two input options are considered for the latter,
nated here, since they are not relevant to the goal of this study.
using either just irradiance data to estimate the opaque cloud frac-
In the absence of any known specific procedure for the quality con-
tion (CDRS1), or rather using the alternate 1-min determinations
trol of tilted irradiance observations, simple quality checks, doubled
from a co-located sky imager (CDRS2). In addition, the conven-
with visual inspection of daily traces of all components, helped
tional and simpler Isotropic model is also considered for compara-
detect and eliminate a number of malfunction periods, particularly
tive purposes:
in the 1-s GTI (photodiode) data.
Rd 1 cos S=2: 5 After application of all these quality-control steps, the observa-
tional uncertainty of the cleaned datasets is estimated at 12% for
Table 1
SRRLs thermopile radiometers used during each measurement period with 1-min data. For redundant DNI and DIF observations, the primary instrument is indicated in bold. The
time response at 95% (in seconds) is given between brackets. K&Z: Kipp & Zonen.
Instrument 20062013 2014 2015

DNI pyrheliometer K&Z CH1 (7 s), Eppley NIP (5 s) EKO MS-56 (1 s), K&Z CH1 (7 s) EKO MS-56 (1 s), K&Z CHP1 (5 s)
DIF pyranometer K&Z CM22 (5 s), Eppley 8-48 (30 s) K&Z CM22 (5 s), Eppley 8-48 (30 s) K&Z CM22#1 (5 s), K&Z CM22#2 (5 s)
GHI pyranometer K&Z CM22 (5 s) K&Z CM22 (5 s) K&Z CM22 (5 s)
GTI40 pyranometer Eppley PSP (5 s) Eppley PSP (5 s) K&Z CMP11 (5 s)
GTI90 pyranometer Eppley PSP (5 s) Eppley PSP (5 s) Eppley PSP (5 s)
Albedo pyranometer Eppley PSP (5 s) Eppley PSP (5 s) K&Z CM21 (5 s)
pyrheliometers, 35% for pyranometers and 5% for silicon that the frequency of enhancement events is more sensitive to sen-
sensors. sor type when the tilt is steeper and albedo effects are stronger.
The results presented in Tables 2 and 3 are meant to provide the
necessary comparative basis for subsequent investigations to be
5. Frequency of enhancement episodes
conducted independently under different climates. So far, only
very limited results of this nature have been reported in the liter-
Some useful statistics describing the enhancement-related fea-
ature (Almeida et al., 2014; de Andrade and Tiba, 2016; Emck and
tures of these datasets appear in Table 2, based on 1.8 million valid
Richter, 2008; Ramgolam and Soyjaudah, 2014; Schade et al., 2007;
points for the 1-min dataset and 23.6 million data points for the 1-
Tapakis and Charalambides, 2014; Yordanov et al., 2013). More-
s dataset. Interestingly, the mean and maximum irradiances are
over, they were presented in a different (mostly graphical) form
somewhat higher when observed with photodiodes than with
only, so that a direct comparison of the frequency of extreme
thermopiles, as previously noted in Part 1. The highest observed
events at the present test site with those occurring at other sites
values of GTI40 and GTI90 are 1740 W m2 and 1812 W m2 when
is virtually impossible.
measured at 1-min resolution with a thermopile or a photodiode,
respectively. When using a photodiode at 1-s resolution, the max-
imum observed GTI40 value culminates at 1998 W m2 (2 suns)
6. Results for separation models
during a snow-covered spring day when GHI also reached a record
1.9 suns, as described in Part 1. Even the 1-min GTI90 can reach val-
Separation models are usually validated with respect to diffuse
ues as high as 1519 W m2 or 1604 W m2, depending on sensor
irradiance, apparently because most of them evaluate the diffuse
type.
fraction. Here, this validation is done in terms of DNI because
The unusually long 10-year span of the available 1-min dataset
direct irradiance is a major contributor to both GHI and GTI at
offers the novelty of studying the interannual variability in mea-
the SRRL test site. More specifically, the mean annual direct frac-
sured GTI, whose results appear in Table 3. Previous results
tion is 69% with regard to GHI, 70% with regard to GTI40 and 64%
(Gueymard and Wilcox, 2011) have shown that the modeled GTIs
with regard to GTI90, assuming a 1-min time step. These mean
interannual variability is intermediate between those of GHI and
annual fractions even reach much higher values (82%, 81%, and
DNI, at least for the annual optimum tilti.e., 40 in the present
72%, respectively) when KT 0.8. Such large values may explain
case. To select periods most affected by CE or AE effects, a 1-sun
why an imprecise separation of DNI from GHI can lead to signifi-
threshold criterion is selected for either GTI40 or GTI90, based on
cant errors in modeled GTI (Gueymard, 2009).
the consideration that PV cells or modules are tested at a nominal
In this section, the performance of the five separation models
incident irradiance of 1 sun. In Part 1, other criteria were addition-
described in Section 2 is evaluated on an annual basis, using 1-
ally used, based on the clearness index KT, i.e., KT 0.8 and
min data from the 20062015 test period. Many different perfor-
KT 1.0. The tilted clearness index, KS = GTI/ETT (where ETT is
mance statistics can be selected for such a task (Gueymard,
the extraterrestrial irradiance on the tilted surface), could have
2014). Among all the possible statistical indicators described there,
been potentially used here in the same way. However, ETT is 0
and for conciseness, only the usual mean bias difference (MBD) and
when the sun is behind the tilted receivers surface. When the sun-
root mean squares difference (RMSD) of the predicted DNI are
beams reach that surface at grazing angles (i.e., for incidence
reported below, in addition to a frequency analysis as a function
angles close to 90), KS can become extremely large, yet meaning-
of KT. Separate MBD and RMSD results are provided in Table 4 for
less, thus creating numerical instability. For this reason, it will not
the whole 10-year dataset (ALL) and different subsets (KT -
be employed here to characterize enhancement episodes on tilted
0.80, GTI 1 sun, GHI 1 sun) intended to describe the
surfaces.
CE and/or AE periods in increasing order of strength and decreasing
The results in Table 3 indicate that the variability in GTI90 is lar-
population of samples (see also Table 3 in Part 1). Moreover, these
ger than that in GHI or GTI40, but depends on the instrument type.
results separately consider the type of GHI instrumentation (ther-
Indeed, based on the ratio photodiode/thermopile (P/T), a small
mopile or photodiode). In addition to the 10-year mean MBD and
instrumentation-induced difference (1.4%) exists for GTI40, but is
RMSD values, the standard deviation of their annual values is indi-
much more pronounced (6.1%) in the case of GTI90, likely because
cated to quantify the interannual variability of the models
of the much stronger impact of AE effects on GTI90 than on GTI40
performance.
or GHI. Additionally, the number of cases with GHI 1 sun or
Table 4 indicates that ERBS consistently generates larger random
GTI90 1 sun is about 50% more when detected by photodiodes
deviations (higher RMSD) than the other models, with additionally
than by thermopiles, which confirms the impact of the detector
a high bias in most cases. In contrast, MAXWELL, PEREZ1 and PEREZ2 have
type, previously noted in Part 1. Overall, GTI40 1 sun events
a different, but common, behavior. This could be expected since
occur 19% of the time when using either thermopile or photodi-
these three models are tightly linked by design. However, whereas
ode sensors at the test site. In contrast, GTI90 1 sun events
occur 3.8% of the time when monitored with a thermopile, and PEREZ1tends to overestimate, PEREZ2 tends to underestimate. This
5.7% when monitored with a photodiode. Even though not all these appears to be caused by the clear-sky correction that transforms
cases are directly attributable to CE/AE effects, it can be construed PEREZ1 into PEREZ2: The clear-sky radiation model used here has only
Table 2
General statistics on the 1-min (period 20062015) and 1-s (period 20102012) quality-controlled observational datasets, according to the GHI/GTI instrumentation type. All
irradiances are expressed in W m2. N represents the number of quality-controlled observations.
Dataset Thermopiles Photodiodes

Time N Mean Mean Max. Mean Max. Mean Max. N Mean Max. Mean Max. Mean Max.
step DNI GHI GHI GTI40 GTI40 GTI90 GTI90 GHI GHI GTI40 GTI40 GTI90 GTI90
1-min 1,810,662 589.9 502.5 1545.9 614.1 1740.1 418.7 1519.3 1,810,752 508.3 1634.3 622.5 1812.2 445.0 1604.5
1-s 23,595,599 589.4 498.3 1501.4 22,626,684 503.9 1891.5 631.6 1998.1
Table 3
Mean annual irradiances (GTI40 and GTI90, in W m2) for the 1-min period (20062015), according to the GTI instrumentation type, and total number of occurrences of extreme
events. SD: standard deviation; CV: coefficient of variation; P/T: ratio photodiode/thermopile.
Year Thermopiles Photodiodes

Mean GTI40 Mean GTI90 GTI40 1 sun GTI90 1 sun Mean GTI40 Mean GTI90 GTI40 1 sun GTI90 1 sun
2006 620.8 429.0 34,596 8165 630.5 456.8 35,818 12,425
2007 616.6 426.0 34,816 9887 620.5 448.1 34,423 12,373
2008 629.7 427.2 38,325 7991 631.5 450.8 35,505 11,633
2009 618.9 426.2 36,971 6919 626.9 452.4 37,057 11,649
2010 620.3 414.0 36,071 5747 634.5 442.6 38,152 9562
2011 629.0 426.4 36,796 8435 634.3 448.5 36,483 11,179
2012 617.2 435.8 30,487 6017 623.7 459.3 30,558 8488
2013 611.0 419.0 32,154 5633 619.9 442.8 32,493 10,123
2014 584.4 386.1 28,854 4359 607.3 419.7 34,527 8318
2015 591.8 399.3 29,473 5596 595.4 424.5 30,031 8362
ALL 614.1 418.7 338,543 68,749 622.5 445.0 345,047 104,112
SD 14.8 15.3 3386 1698 12.6 13.0 2710 1653
CV (%) 2.4 3.6 10.0 24.7 2.0 2.9 7.9 15.9
P/T 1.014 1.061 1.019 1.514
Table 4
Performance results of five separation models using 1-min data at the SRRL site in Golden, Colorado. The MBD and RMSD statistics are expressed in percent of the observed mean
DNI over the whole 10-year period (20062015). Their mean interannual variability is expressed as one standard deviation. Results refer to two different methods of GHI
observation, based on either a thermopile pyranometer or a photodiode sensor.
Statistic Thermopile global pyranometer Photodiode global sensor

DNI Obs. DNI ERBS DNI MAXWELL DNI PEREZ1 DNI PEREZ2 DNI ENGERER DNI Obs. DNI ERBS DNI MAXWELL DNI PEREZ1 DNI PEREZ2 DNI ENGERER
ALL
Mean (W m2) 590.7 608.5 615.8 614.3 568.3 573.6 590.7 611.2 618.1 616.5 570.4 576.5
MBD (%) 3.0 0.9 4.3 0.9 4.0 0.9 3.8 1.0 2.9 0.6 3.5 1.3 4.7 0.8 4.4 1.0 3.4 1.0 2.4 0.6
RMSD (%) 20.0 1.2 19.4 1.5 18.1 1.4 18.9 1.1 17.7 1.1 22.7 1.3 20.9 1.2 19.7 1.2 20.3 1.0 18.3 1.0
KT 0.80
Mean (W m2) 924.0 957.9 920.4 936.6 876.0 901.1 913.4 965.2 919.8 934.5 872.1 902.6
MBD (%) 3.7 1.2 0.4 2.3 1.4 1.9 5.2 2.0 2.5 2.7 5.7 1.0 0.7 0.8 2.3 0.7 4.5 0.7 1.2 0.8
RMSD (%) 15.9 3.9 12.2 2.9 11.5 2.9 12.4 3.0 11.8 3.1 17.4 1.2 13.6 0.9 13.3 0.9 13.6 0.9 12.5 0.9
GTI 1 sun
Mean (W m2) 951.8 944.3 934.4 955.5 902.4 927.5 950.7 959.3 934.4 955.8 903.7 929.6
MBD (%) 0.8 0.6 1.8 0.4 0.4 0.4 5.2 0.4 2.5 0.3 0.9 0.9 1.7 0.4 0.5 0.5 4.9 0.5 2.2 0.4
RMSD (%) 11.2 0.6 9.9 0.5 9.2 0.6 10.5 0.5 9.3 0.5 12.9 0.9 11.3 0.8 10.9 0.8 11.7 0.8 9.9 0.7
GHI 1 sun
Mean (W m2) 927.3 965.0 934.2 949.9 885.8 933.2 920.4 984.9 928.4 945.1 881.9 931.5
MBD (%) 4.1 2.4 0.7 1.5 2.4 1.2 4.5 1.2 0.6 1.3 7.0 2.5 0.9 1.3 2.7 1.3 4.2 1.2 1.2 1.3
RMSD (%) 13.9 1.4 11.3 0.9 11.2 1.2 11.8 0.8 10.0 1.1 16.4 2.0 13.6 1.3 13.6 1.6 13.7 1.5 10.9 1.3
limited performance (Gueymard and Ruiz-Arias, 2015), and uses whereas ENGERER tends to slightly underestimate. In the region of
turbidity data that are available only as a climatology (long-term potentially large CE effects (KT 0.8), which is the focus here, all
mean monthly values), hence on a much coarser temporal basis models clearly tend to diverge. ERBS predicts unphysically high val-
than the 1-min time step considered here. The same limitation ues of DNI for KT > 1, as noted before (Gueymard and Ruiz-Arias,
appears to hurt ENGERER too, but its low bias is much less than that 2014, 2016). Simultaneously, MAXWELL, PEREZ1 and PEREZ2 (the latter
of PEREZ2. Overall, ENGERER performs consistently better than all other not shown for clarity) tend to predict an extremely low (and even
models, thus confirming an earlier study (Gueymard and Ruiz- frequently null) DNI, when it is actually high. The widely different
Arias, 2016). The interannual variability of the bias is relatively individual 1-min DNI predictions of ERBS, PEREZ2 and ENGERER under
small, which should guarantee the possibility to generalize results such over-unity situations are shown in Fig. 2. PEREZ2 (as well as
even if only one year of data is available at any siteat least if its PEREZ1, not shown) predicts null values (or actually negative values
radiative climate is similar to SRRLs. that are artificially floored to 0) when KT > 1.15. This problem was
For all models, the MBD and RMSD statistics vary significantly also noted before (Gueymard and Ruiz-Arias, 2014, 2016). The
according to the type of GHI instrumentation, and also depending divergence of ENGERER is much less accentuated, and might be
on the more or less intensive filtering of the database to pinpoint caused in part by the imperfect clear-sky DNI estimates. This could
CE and AE effects. To elucidate the magnitude of these effects be corrected with a better clear-sky radiation model and/or better
impact on the accuracy of the modeled DNI, the distribution of input turbidity data. The generally better results of ENGERER makes it
the mean residuals is plotted against KT in Fig. 1. (The residuals more suitable to high-frequency applications when KT can be fre-
are averaged over equidistant KT bins.) The results for the ther- quently >1. Fig. 1 also shows the mean DNI in each KT bin. The max-
mopile and photodiode cases are similar, but a larger maximum imum mean DNI (940 W m2) is reached for 0.8 KT < 0.85. Such
value of KT is reached in the latter case (1.49 compared to 1.30 in situations typically correspond to spring days at SRRL, when the
the thermopile case). Over mid-KT values (between 0.3 and 0.8), sun is relatively high and the atmosphere is still very dry and clean.
where most of the cases are concentrated, MAXWELL, PEREZ1 and When KT is above 0.85, the mean DNI slightly decreases. Such cases
PEREZ2 tend to overestimate (by up to 100 W m2 on average), correspond to CE and AE effects, which mostly occur during late
Fig. 1. Binned mean residuals (left Y-axis, colored lines) of the predicted DNI with ERBS, MAXWELL, PEREZ1 and ENGERER as a function of KT. The thick grey lines represent the binned
mean observed DNI (right Y-axis). Solid lines correspond to the use of a thermopile pyranometer to measure GHI, and the dashed lines to the alternate case of using a
photodiode sensor. The MAXWELL and PEREZ1 mean residuals are indistinguishable over most of the domain. (For interpretation of the references to colour in this figure legend,
the reader is referred to the web version of this article.)
90), 7 transposition models, 2 alternate inputs for DNI and DIF

(measured or modeled), 2 databases (all observations and the
KT 0.80 subset), and 2 types of instrumentation (thermopiles or
photodiodes). For simplification, it is assumed that the same type
of instrumentation is used for both GHI and GTI, which is typically
the case in practice. (Preliminary tests indicate that the results
obtained below would not change significantly because of differing
types of instrumentation.) The interannual variability in the results
is similar to that observed in Section 5 (Table 3), so that only the
10-year average results are reported for conciseness.
Tables 5 and 6 provide the results for the 40 and 90 tilts,
respectively. The 40 results in Table 5 show that, on average,
the thermopile and photodiode GTI observations are in reasonable
agreement for both the whole database and the KT 0.80 subset.
(Note, however, that the number of occurrences in the latter case
is much larger when measuring GTI with a photodiode than with
a thermopile instrument, per Table 3). Conversely, the model pre-
dictions appear to be sensitive to the nature of their inputs, i.e.,
GHI obtained from thermopile or photodiode instruments, and
DNI and DIF obtained from measured or modeled values. The per-
Fig. 2. Measured vs. predicted DNI by three separation models (ENGERER, ERBS, and
formance of all transposition models (MBD and RMSD) tends to be
PEREZ2) when KT > 1, using 1-min data over the period 20062008.
better when using photodiodes than thermopiles, except when
using ENGERER-modeled DNI and DIF and KT 0.80, in which case
spring or summer when the sun is highest, the atmosphere is more the bias (represented by MBD) is reduced. The bias is within 3%
turbid, and CE effects might obscure the sun during a fraction of in all cases for all models, except ISOTROPIC. The latter tends to signif-
the 1-min period. When KT > 1.2, the mean observed DNI is in the icantly underestimate, which could be expected since it predicts
800860 W m2 range, whereas MAXWELL, PEREZ1 and PEREZ2 all pre- lower diffuse irradiance on sunlit surfaces than anisotropic models.
dict null values, but such cases are quite infrequent. Similar results The random deviations (represented by RMSD) tend to decrease
are obtained when using 1-s data, but are not shown for appreciably when KT 0.80, which can be explained by the lower
conciseness. contribution from the modeled sky diffuse irradiance to GTI, and
the concomitant larger fraction from the deterministic direct irradi-
ance contribution. Surprisingly, PEREZ tends to have a negative bias
7. Results for transposition models and a larger RMSD than nearly all other models (except ISOTROPIC),
which was unexpected. The reasons for this are discussed below.
The transposition models described in Section 3 are validated The 90 results in Table 6 show some amplification of differ-
here against 1-min GTI measurements obtained for both 40 and ences compared to those in Table 5. For instance, the mean
90 tilted surfaces. The DNI and DIF components are supposed to observed GTI90 is 6.3% higher when measured with a photodiode
be either measured (reference case) or derived from the measured than with a thermopile pyranometer, which is significant. In many
GHI through a separation model. Based on the results of the previ- cases, this experimental artifact overshadows the potential model-
ous section, and for conciseness, only ENGERER is used in the latter to-model differences in GTI90 predictions. Additionally, the MBD
case, since no good GTI prediction can be expected at high KT when and RMSD results tend to indicate larger model errors than in
DNI is too far off, as would be the case if using the other separation Table 5. This can be explained in part by the lower contribution
models. Overall, results are obtained for a total of 1120 possible of direct irradiance to GTI90 already discussed. Hence, sky diffuse
cases: 10 separate years (20062015), 2 different tilts (40 and transposition errors are magnified, as well as those due to depar-
Table 5
Mean annual performance statistics for the modeling of 1-min GTI40 with seven transposition models over the period 20062015. The observed GHI and GTI40 are measured with
either thermopiles or photodiodes. The DNI and DIF inputs are measured, or alternatively evaluated with ENGERERs separation model. The top rows refer to all data points, whereas
the bottom rows are for the subset restricted to the KT 0.8 cases.
Statistic Thermopiles Photodiodes

GTI40 Obs. CDRS1 CDRS2 HAY ISOTROPIC MUNEER PEREZ HDKR Obs. CDRS1 CDRS2 HAY ISOTROPIC MUNEER PEREZ HDKR
ALL points, N = 1,810,662
Measured DNI and DIF 614.1 603.4 610.1 603.4 584.9 612.3 603.7 604.9 622.5 603.5 609.9 603.2 584.8 612.5 603.6 604.7
MBD (%) 1.7 0.7 1.7 4.7 0.3 1.7 1.5 3.0 2.0 3.1 6.1 1.6 3.0 2.9
RMSD (%) 5.2 4.9 5.5 7.9 5.3 6.6 5.4 5.6 5.1 5.6 8.7 5.0 7.0 5.4
Modeled DNI and DIF 614.1 601.9 608.6 602.3 580.6 612.2 603.0 603.9 622.5 603.5 614.0 607.6 586.0 617.9 607.6 609.3
MBD (%) 2.0 0.9 1.9 5.5 0.3 1.8 1.7 2.4 1.4 2.4 5.9 0.7 2.4 2.1
RMSD (%) 6.5 5.8 6.6 9.3 6.4 7.7 6.5 7.8 6.9 7.3 10.3 7.4 8.5 7.3
KT 0.80, N = 445,171
MBD (%) 1.6 1.6 1.2 4.5 0.5 2.2 1.1 2.1 2.1 1.8 4.8 0.1 2.6 1.7
RMSD (%) 3.5 3.6 3.9 5.8 4.0 4.2 3.9 3.6 3.6 3.7 6.1 3.4 4.6 3.6
MBD (%) 2.1 2.0 1.7 5.4 0.2 2.6 1.5 0.4 0.3 0.1 3.6 1.9 1.2 0.0
RMSD (%) 4.1 3.9 3.9 7.0 3.8 4.8 3.9 4.5 4.1 4.0 6.5 4.7 5.4 4.0
Table 6
Mean annual performance statistics for the modeling of 1-min GTI90 with seven transposition models over the period 20062015. The observed GHI and GTI90 are measured with
either thermopiles or photodiodes. The DNI and DIF inputs are measured, or alternatively evaluated with ENGERERs separation model. The top rows refer to all data points, whereas
the bottom rows are for the subset restricted to the KT 0.8 cases.
Statistic Thermopiles Photodiodes

GTI90 Obs. CDRS1 CDRS2 HAY ISOTROPIC MUNEER PEREZ HDKR Obs. CDRS1 CDRS2 HAY ISOTROPIC MUNEER PEREZ HDKR
ALL points, N = 1,810,662
MBD (%) 5.6 6.3 1.0 4.3 3.4 1.7 0.8 0.5 0.1 6.7 9.8 2.6 7.4 5.1
RMSD (%) 11.3 11.3 8.2 10.6 9.9 11.1 8.0 11.4 11.1 10.7 15.2 9.1 13.7 9.6
MBD (%) 5.3 5.9 1.3 5.5 3.5 1.7 0.6 0.3 0.3 6.8 10.5 2.0 7.3 4.9
RMSD (%) 13.4 12.6 10.3 13.9 11.7 12.9 10.2 15.0 13.9 13.4 18.8 12.4 16.2 12.8
KT 0.80, N = 445,171
MBD (%) 2.5 2.8 0.0 3.3 5.7 0.7 1.2 2.6 2.4 5.6 8.1 0.2 6.0 4.4
RMSD (%) 6.4 6.6 6.4 7.7 9.3 6.5 6.6 7.1 7.1 8.3 11.7 7.3 9.5 7.6
MBD (%) 1.6 2.0 0.9 5.0 5.2 1.4 0.3 1.5 1.2 4.9 7.9 1.8 5.6 3.5
RMSD (%) 6.9 6.7 6.2 10.5 8.8 8.0 6.3 8.9 8.4 8.7 13.8 8.6 11.9 8.2
ture from the assumptions behind Eqs. (2) and (3). The latter issue fits and handle extreme combinations of inputs caused by CE
means that a part of the prediction errors attributed to the models or AE effects. This issue is structurally similar to that of the PEREZ1
can actually be caused by an incorrect estimate of the reflected separation model discussed above. Whereas the model is known
irradiance. Additional instrumentation (pyranometers with shades for its generally excellent performance in hourly calculations, its
obstructing the sky and/or the ground view) would be necessary to performance is seriously hampered by this problem in the case
resolve this issue. of 1-min calculations. An overhaul of this model to accommodate
The unexpected low performance of PEREZ requires further high-KT situations is necessary, but is beyond the scope of this
examination. It is found that, under specific circumstances related study. More details are provided in Section 8.
to CE and/or AE effects, the PEREZ-predicted diffuse tilted irradiance, Plots of the mean residuals of each transposition model appear
derived from Eq. (4), can be extremely low, and even negative. in Fig. 3 for the 40 tilt and Fig. 4 for the 90 tilt. Similarly to Fig. 1,
(Note that its original Fortran code, used here to avoid any misin- the observed values and model residuals of GTI are averaged for
terpretation, does not prevent such negative values.) This very each bin of KT, separately for the thermopile and photodiode cases.
rarely happens for the 40 tilt, but occurs each year during a sizeble The results for only four transposition models are shown for legi-
number (2001000) of 1-min periods for the 90 tilt, depending on bility. To analyze their intrinsic performance, all models are oper-
instrumentation type and whether DNI and DIF are measured or ated here with measured DNI and DIF values. Figs. 3 and 4 indicate
modeled. The statistics in Tables 5 and 6 correspond to the cor- that PEREZ overestimates for KT 0.35, and underestimates for
rected PEREZ output, i.e., after flooring all unphysically negative KT > 0.4 (and more intensely so when KT reaches large values for
tilted diffuse predictions to 0. The cause for the unexpected behav- the reason just explained). In contrast, MUNEER and HDKR strongly
ior of PEREZ, which has not been reported in the literature until now, overestimate for KT > 1. (HAYs behavior is almost identical to
appears related to the structure of the model, which might overfit HDKRs by design.) In Fig. 3, CDRS1 (and CDRS2, not shown) remain
under normal conditions, while being unable to extrapolate these essentially unbiased over the whole domain, particularly when
Fig. 3. Dependence on KT of the mean residuals (left Y-axis, colored curves) of the predicted irradiance on a 40 tilt using four transposition models: CDRS1 (blue curves),
HDKR (green curves), MUNEER (pink curves) and PEREZ (brown curves). The grey curves represent the mean measured global tilted irradiance (right Y-axis). Solid curves refer to
thermopile measurements of both GHI and GTI, whereas dashed curves refer to photodiode measurements. Measured DNI and DIF components are used by the models. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 4. Dependence on KT of the mean residuals of the predicted irradiance on a 90 tilt using four transposition models: CDRS1 (blue curves), HDKR (green curves), MUNEER
(pink curves) and PEREZ (brown curves). The grey curves represent the mean measured global tilted irradiance (right Y-axis). Solid curves refer to thermopile measurements of
both GHI and GTI, whereas dashed curves refer to photodiode measurements. Measured DNI and DIF components are used by the models. (For interpretation of the references
to colour in this figure legend, the reader is referred to the web version of this article.)
using photodiodes. Note also the steep change in GTI40 when KT intelligence technique) could provide the best solution for any tilt
increases from 0.65 to 0.80, possibly due to AE effects. and any KT.
For the 90 tilt, Fig. 4 shows similar features as in Fig. 3, but Using 1-s data in lieu of 1-min data obviously results in an
with more amplitude. PEREZ now substantially underestimates for amplification of the problems just described. Considering that 1-s
KT in the range 0.350.80 and above 0.90, particularly with photo- data are are only available during 27 months, and not available
diodes. MUNEER and HDKR (as well as Hay, not shown) considerably for GTI90, no specific results are pictured for conciseness.
overestimate for KT above 0.801.0, depending on model and
instrumentation, but are basically unbiased below 0.80. In contrast, 8. Case studies
CDRS1 (and CDRS2, not shown) significantly overestimates for
KT < 0.7, but provides almost unbiased results for KT > 0.90 when Some enhancement episodes are illustrated in this section.
using photodiodes. Their impact on the modeled GTI for either the 40 or 90 tilt is
Figs. 3 and 4 exemplify the complexity of the dependence on KT evaluated by showing the residuals (modeled GTI minus measured
of the performance results of transposition models, which has GTI) for a few transposition models. Since, in general, CDRS2 and
apparently not been studied before. This complexity is such that
HAY behave very closely to CDRS1 and HDKR, respectively, the
cancellation of significant errors along the whole KT domain may
results from only four models are shown for legibility: CDRS1,
result in an artificially low overall bias. Only one model (CDRS)
provides nearly unbiased results for the 40 tilt, but no model MUNEER, PEREZ, and HDKR.
succeeds to do the same for the 90 tilt. All this suggests that
improvements in transposition models are necessary to make 8.1. Results for the 40 tilt
them able to provide accurate predictions under all possible situa-
tions on a 1-min time step. Alternatively, an optimum combination A first case is that of a spring day, 2014-05-12, just after a heavy
of existing models (using, e.g., neural networks or other artificial snowfall (Fig. 5). The snow accumulation (measured with a Texas
Fig. 5. Impacts of cloud and albedo enhancement events on the measured 1-min GTI40 and its predictions from four transposition models during a high-albedo spring day.
The bottom panel shows the 1-min observations of total cloud fraction (cover), opaque cloud fraction and local surface albedo.
Electronics TR-525i tipping-bucket gauge) reached 15 cm in early 1596 W m2 (KT = 1.32) at the end of a sequence of four CE/AE
morning, and was still about 10 cm during the episode depicted events, which lasted 1230 min each. All four transposition models
in Fig. 5, which explains the concomitant high albedo (0.8). A display different residual patterns, with HDKR and MUNEER now clo-
short period (15 min) of snow flurries even occurred after the ser to the ideal base line. During the GTI40 peaks, CDRS1 underes-
main CE/AE-related peak value of GTI (1375 W m2, KT = 1.04) at timates by up to 200 W m2, while PEREZ underestimates by up to
12:15 LST. The GTI40 residuals are indicated for four transposition 881 W m2a 55% deviation.
models. Remarkably, CDRS1 and HDKR are in perfect agreement,
with residuals typically in the 50 W m2 range. MUNEER is close,
8.2. Results for the 90 tilt
with only a few spikes. In contrast, PEREZ significantly underesti-
mates during the peak periods, by up to 300 W m2.
Figs. 911 provide similar results and for the same days as those
The case of another spring day, 2015-04-17, is shown in Fig. 6
in Figs. 57, respectively, but for GTI90 and only for 1-min data,
(and was also discussed in Part 1). This case is relatively similar
since 1-s measurements do not exist for that tilt.
to the previous one in the sense that significant snowfall had just
During 2014-05-12, the two GTI peaks that are shown in Fig. 5
occurred and that a high albedo of 0.8 was recorded in early morn-
occur at the same time as in Fig. 9, even though their relative mag-
ing. However, the snow melted fast in the present case, so that the
nitude is now reversed. During this high-albedo CE/AE event, HDKR
local albedo was only 0.45 during the CE/AE period shown in
performs best, closely followed by MUNEER. CDRS1 overestimates by
Fig. 6. Despite this moderate albedo, the 1-s GTI40 culminated at
up to 200 W m2, whereas PEREZ underestimates by up to
a record 2 suns, whereas the 1-min peak GTI40 was only
475 W m2.
1544 W m2 and occurred some 50 min later. The CDRS1 and
Fig. 10, for 2015-04-17, shows marked differences with Fig. 6,
HDKR models behave similarly during the whole period and their
partly because of the availability of 1-min data only. The peak
residuals are generally very close to the baseline. In contrast, PEREZ
GTI90 is 1107 W m2a remarkably high value for a vertical plane
underestimates most of the time, by up to 609 W m2.
during spring and only moderate albedo. CDRS1 tends to slightly
Fig. 7 illustrates the case of another high-albedo day, but this
overestimate here, whereas HDKR and MUNEER behave similarly
time in winter (2012-02-10). Snow accumulation had started a
with somewhat lower deviations. That of PEREZ reaches
week earlier. The albedo during this test day is only 0.7, due to
1235 W m2 due to its prediction of a highly negative tilted dif-
the somewhat aged snow. The peak 1-s GTI40 for that day is
fuse component. Flooring the latter to 0 reduces the underestima-
1652 W m2 (KT = 1.36). The 1-s residuals that are also shown indi-
tion of the corrected PEREZ GTI to 453 W m2.
cate larger deviations than in the previous case. CDRS1, MUNEER and
Finally, a stronger CE/AE event is depicted in Fig. 11 for 2012-
HDKR now deviate by up to 200 W m2. PEREZ again displays sys-
02-10, when the peak GTI90 reached 1599 W m2. CDRS1, HDKR,
tematic and considerable underestimation during the whole CE/
and MUNEER tend to overestimate by up to 200 W m2. CDRS1 and
AE event, up to 850 W m2.
HDKR have very similar performance and are closest to the ideal
Finally, 1-s results for another high-albedo day (2011-11-02)
baseline. In contrast, PEREZ underestimates by up to 943 W m2a
are depicted in Fig. 8. The measured GTI40 culminates at
deviation of 59% at that specific moment.
Fig. 6. Impacts of cloud and albedo enhancement events on the measured 1-min and 1-s GTI40 and its predictions from four transposition models during a spring day with
moderate to high albedo. See Fig. 5 for a description of the bottom panel.
Fig. 7. Impacts of cloud and albedo enhancement events on the measured 1-min and 1-s GTI40 and its predictions from four transposition models during a high-albedo winter
day. See Fig. 5 for a description of the bottom panel.
Fig. 8. Impacts of cloud and albedo enhancement events on the measured 1-s GTI40 and its predictions from four transposition models during a high-albedo winter day. See
Fig. 5 for a description of the bottom panel.
Fig. 9. Impacts of cloud and albedo enhancement events on the measured 1-min GTI90 and its predictions from four transposition models during a high-albedo spring day.
See Fig. 5 for a description of the bottom panel.
Fig. 10. Impacts of cloud and albedo enhancement events on the measured 1-min GTI90 and its predictions from four transposition models during a spring day with moderate
to high albedo. See Fig. 5 for a description of the bottom panel.
Fig. 11. Impacts of cloud and albedo enhancement events on the measured 1-min GTI90 and its predictions from four transposition models during a high-albedo winter day.
See Fig. 5 for a description of the bottom panel.
8.3. Discussion case. The interannual variability of these events also appears sub-
stantial, again particularly in the 90 case. As a long-term average,
The cases illustrated above show the key role of a high local sur- GTI is >1 sun 19% of the time at the test site for the 40 tilt and
face albedo as an amplifier for CE effects. It is known that snows 3.85.7% of the time (depending on instrument type) for the 90
reflectance has substantial anisotropy (specially icy snow), thus tilt. These statistics complement those pertaining to GHI that were
resulting in enhanced forward reflections that can substantially provided in Part 1 of this contribution. For the 40 tilt, the maxi-
increase GTI when the sun faces a tilted surface. Unfortunately, mum recorded value of GTI is 2000 W m2 (2 suns), using 1-s
Eqs. (2) and (3) cannot model these anisotropic effects, which data11% more than when using 1-min data. For the 90 tilt with
can trigger significant modeling errors in GTI, and may artificially 1-min data, it is 1605 W m2. The GTI peaks are not always in
improve or deteriorate any models performance, at least tem- phase when considering 1-min vs. 1-s data, and they are typically
porarily. In general, it can be anticipated that the highest possible much larger in the latter case than the former, thus confirming the
GTI values result from the combination of CE and snow-induced AE findings in Part 1.
effects. The performance of five separation models of the literature has
It is stressed that the local albedo can substantially differ from been evaluated as a function of KT. This performance is found to
the regional albedo, which triggers AE effects on GHI, as discussed degrade substantially for a clearness index (KT) above 0.8, when
in Part 1. Because of this difference, AE or CE/AE effects on GHI and enhancement effects become prominent. This is particularly the
GTI may occur at different times and may vary in strength. The pre- case for MAXWELL, PEREZ1 and PEREZ2, which tend to predict a null
sent findings underline the boosting albedo effect of CE events. DNI when it is actually very high, as well as for ERBS, which con-
The precise validation of transposition models during days that versely tends to predict unphysically high DNI values. The best
include CE/AE situations is difficult in the present experimental results are obtained with ENGERER, but it could still benefit from a
setup because of (i) the additional dependence on instruments better clear-sky radiation model and better input data. All this con-
time response, and (ii) the lack of radiometers equipped with sky firms previous findings (Gueymard and Ruiz-Arias, 2016).
or ground shades to separate the sky diffuse from the ground- Comparatively to the case of GHI analyzed in Part 1, it is more
reflection contributions, such as those described by e.g. (Behr, difficult to detect enhancement effects on historical time series
1997). Hence, in the absence of precise experimental determina- of GTI measurement. In particular, the tilted-plane equivalent of
tion of EdS, transposition models (which normally only predict KT is found inappropriate for that purpose, due to its numerical
the sky diffuse component of GTI, i.e. EdS in Eq. (4)) can only be instability during specific periods. The performance of the seven
tested for their overall performance relative to GTI. It is thus con- transposition models considered here is thus analyzed as a func-
ceivable that, at any instant, a good transposition model might tion of KT only. The overall latitude-tilt (40) results are comparable
appear biased if the reflected component (ErS in Eq. (2)) is incor- between five of the six anisotropic transposition models, whereas
rectly estimated for that instant, and vice versa. The magnitude PEREZ tends to underpredict significantly during enhancement
of this source of uncertainty, and how much it affects the overall events. The vertical-tilt (90) results show substantial differences,
model predictions and their long-term performance is still particularly considering PEREZs unexpected tendency to severely
unknown. In any case, it is obvious that AE effects will have stron- mispredict GTI overall, due in particular to many occurrences of
ger impacts on steeper tilts. unphysically negative sky-diffuse predictions that occur during
In addition to the difficulties of validating diffuse transposition CE/AE events. Consequently, its use is not recommended with
models with incomplete experimental setups, the case studies sub-hourly data or whenever KT > 1. For the 40 tilt, the CDRS
described above clearly indicate that the exact magnitude or time model is found the most stable with low bias overall, even at large
of occurrence of GTI peaks during CE/AE events cannot be easily KT. For the 90 tilt, the situation is more complex, since each model
extrapolated from just the combination of GHI measurements experiences significant biases in various parts of the KT range, most
and existing separation/transposition models. This is most spe- particularly over 0.8. On average, CDRS is less affected by high-KT
cially the case if the transposition is based on PEREZ, since basically CE/AE events than the other models.
all strong CE/AE events would then be missed. Improvements in Based on the present findings, the frequency of appearance of
both experimental methods and transposition modeling are there- CE/AE episodes appears to be a determining factor in the overall
fore required to prepare more in-depth studies of the magnitude GTI uncertainty at any specific location, particularly over snow-
and frequency of CE/AE events wherever their impact on PV sys- impacted areas and for steep tilts. Further research is needed to
tems needs to be investigated. develop new or improved separation and transposition models
that can withstand the impact of such extreme conditions. Their
validation and/or improvement would require advanced experi-
9. Conclusion mental setups to separate the sky-diffuse contribution from the
ground-reflected contribution to GTI. This is because of the current
Based on a long 10-year time series of high-accuracy irradiance uncertainty in the latters estimation, which is highly site- and
measurements at a high-elevation station frequently impacted by time-specific. An optimum combination of the best existing models
snow cover, this experimental study has evaluated the perfor- could also be obtained with the help of appropriate artificial intel-
mance of separation and transposition models in general, and most ligence techniques. Finally, it is stressed that the current models
specifically under cloud and/or albedo enhancement situations. that are routinely used to transform GHI into GTI cannot always
High-irradiance situations that can negatively impact PV installa- predict the precise magnitude or period of occurrence of those
tions constituted the focus of this study. Two types of sensorther- intense, short-lived GTI peaks caused by CE/AE events.
mopile and photodiode radiometers with widely different time
response characteristicshave been used at either 1-min resolu-
tion (10-year period) or 1-s resolution (27-month period). The fre- Acknowledgements
quency of 1-min periods such that the global total irradiance (GTI)
is >1 sun, for either a latitude-tilt (40) or a vertical tilt (90), has This study would not have been possible without the dedicated
been determined over the period 20062015. It is found to vary work of NRELs personnel who maintains the SRRL stations and
significantly depending on instrument type, with photodiode sen- radiometers. This author was granted special access to the internal
sors recording more high-irradiance periods, particularly in the 90 1-sec NREL grid data, which is highly appreciated. The author also
thanks Nathan Kamphuis for sharing information about the his- Luoma, J., Kleissl, J., Murray, K., 2012. Optimal inverter sizing considering cloud
enhancement. Sol. Energy 86, 421429. http://dx.doi.org/10.1016/
toric origin of the isotropic model.
j.solener.2011.10.012.
Maxwell, E.L., 1987. A quasi-physical model for converting hourly global horizontal
References to direct normal insolation. Tech. Rep. SERI/TR-215-3087, National Renewable
Energy Lab., Golden, CO. <http://rredc.nrel.gov/solar/pubs/PDFs/TR-215-3087.
Almeida, M.P., Zilles, R., Lorenzo, E., 2014. Extreme overirradiance events in So pdf>.
Paulo, Brazil. Sol. Energy 110, 168173. http://dx.doi.org/10.1016/ Moon, P., Spencer, D.E., 1942. Illumination from a non-uniform sky. Trans. Illum.
j.solener.2014.09.012. Eng. Soc. 37, 707726.
Andrews, R.W., Pearce, J.M., 2013. The effect of spectral albedo on amorphous Muneer, T.E., 2004. Solar Radiation and Daylight Models. Elsevier.
silicon and crystalline silicon solar photovoltaic device performance. Sol. Energy chsner, P., Zehner, M., Lang, F., Rauscher, T., Weizenbeck, J., Weigl, T., Becker, G.,
91, 233241. http://dx.doi.org/10.1016/j.solener.2013.01.030. Bettenwort, G., Giesler, B., Betts, T.R., Gottschalg, R., 2013. Spatial modelling of
Behr, H.D., 1997. Solar radiation on tilted south oriented surfaces: validation of grid connected PV plants with 3D irradiance values. In: Proc. 28th European
transfer-models. Sol. Energy 61, 399413. http://dx.doi.org/10.1016/S0038- Photovoltaic Solar Energy Conference and Exhibition, Paris, France.
092X(97)00081-9. Perez, R., Ineichen, P., Maxwell, E.L., Seals, R., Zelenka, A., 1992. Dynamic global-to-
Burger, B., Rther, R., 2006. Inverter sizing of grid-connected photovoltaic systems direct irradiance conversion models. ASHRAE Trans. 98 (1), 354369.
in the light of local solar resource distribution characteristics and temperature. Perez, R., Ineichen, P., Moore, K., Kmiecik, M., Chain, C., George, R., Vignola, F., 2002.
Sol. Energy 80, 3245. A new operational model for satellite-derived irradiances: description and
Chen, S., Li, P., Brady, D., Lehman, B., 2013. Determining the optimum grid- validation. Sol. Energy 73, 307317. http://dx.doi.org/10.1016/S0038-092X(02)
connected photovoltaic inverter size. Sol. Energy 87, 96116. http://dx.doi.org/ 00122-6.
10.1016/j.solener.2012.09.012. Perez, R., Ineichen, P., Seals, R., Michalsky, J., Stewart, R., 1990. Modeling daylight
Cucumo, M., De Rosa, A., Ferraro, V., Kaliakatsos, D., Marinelli, V., 2007. availability and irradiance components from direct and global irradiance. Sol.
Experimental testing of models for the estimation of hourly solar radiation on Energy 44, 271289.
vertical surfaces at Arcavacata di Rende. Sol. Energy 81, 692695. Polo, J., Garcia-Bouhaben, S., Alonso-Garcia, M.C., 2016. A comparative study of the
de Andrade, R.C., Tiba, C., 2016. Extreme global solar irradiance due to cloud impact of horizontal-to-tilted solar irradiance conversion in modelling small PV
enhancement in northeastern Brazil. Renew. Energy 86, 14331441. http://dx. array performance. J. Renew. Sustain. Energy 8, 053501.
doi.org/10.1016/j.renene.2015.09.012. Ramgolam, Y.K., Soyjaudah, K.M.S., 2014. Enhanced insolation and global irradiance
Duffie, J.A., Beckman, W.A., 1991. Solar Engineering of Thermal Processes. Wiley, in near-tropic region. In: Proc. EuroSun 2014 Conf. International Solar Energy
New York. Soc, Aix-les-Bains, France.
Emck, P., Richter, M., 2008. An upper threshold of enhanced global shortwave Roberts, J.J., Mendiburu Zevallos, A.A., Cassula, A.M., 2017. Assessment of
irradiance in the troposphere derived from field measurements in tropical photovoltaic performance models for system simulation. Renew. Sustain.
mountains. J. Appl. Meteorol. Clim. 47, 28282845. Energy Rev. 72, 11041123. http://dx.doi.org/10.1016/j.rser.2016.10.022.
Engerer, N.A., 2015. Minute resolution estimates of the diffuse fraction of global Schade, N.H., Macke, A., Sandmann, H., Stick, C., 2007. Enhanced solar global
irradiance for southeastern Australia. Sol. Energy 116, 215237. http://dx.doi. irradiance during cloudy sky conditions. Meteorol. Zeit. 16, 295303.
org/10.1016/j.solener.2015.04.012. Skartveit, A., Olseth, J.A., Tuft, M.E., 1998. An hourly diffuse fraction model with
Erbs, D., Klein, S.A., Duffie, J.A., 1982. Estimation of the diffuse radiation fraction for correction for variability and surface albedo. Sol. Energy 63, 173183.
hourly, daily and monthly average global radiation. Sol. Energy 28, 293302. Tapakis, R., Charalambides, A.G., 2014. Enhanced values of global irradiance due to
Gueymard, C.A., 1987. An anisotropic solar irradiance model for tilted surfaces and the presence of clouds in Eastern Mediterranean. Renew. Energy 62, 459467.
its comparison with selected engineering algorithms. Sol. Energy 38, 367386. http://dx.doi.org/10.1016/j.renene.2013.08.001.
Gueymard, C.A., 2009. Direct and indirect uncertainties in the prediction of tilted Temps, R.C., Coulson, K.L., 1977. Solar radiation incident upon slopes of different
irradiance for solar engineering applications. Sol. Energy 83, 432444. orientations. Sol. Energy 19, 179184.
Gueymard, C.A., 2014. A review of validation methodologies and statistical Weiser, U., Olefs, M., Schner, W., Weyss, G., Hynek, B., 2016. Correction of
performance indicators for modeled solar radiation data: towards a better broadband snow albedo measurements affected by unknown slope and sensor
bankability of solar projects. Renew. Sustain. Energy Rev. 39, 10241034. tilts. The Cryosphere 10, 775790.
Gueymard, C.A., Ruiz-Arias, J.A., 2014. Performance of separation models to predict Xie, Y., Sengupta, M., 2016. Diagnosing model errors in simulations of solar
direct irradiance at high frequency: validation over arid areas. In: Proc. EuroSun radiation on inclined surfaces. In: Proc. 43rd IEEE Photovoltaic Specialists Conf.,
2014 Conf., Aix-les-Bains, France, International Solar Energy Soc. <http://www. Portland, OR <http://www.nrel.gov/docs/fy16osti/66489.pdf>.
ises.org>. Yang, D., 2016. Solar radiation on inclined surfaces: corrections and benchmarks.
Gueymard, C.A., Ruiz-Arias, J.A., 2015. Validation of direct normal irradiance Sol. Energy 136, 288302. http://dx.doi.org/10.1016/j.solener.2016.06.062.
predictions under arid conditions: a review of radiative models and their Yang, D., Dong, Z., Nobre, A., Khoo, Y.S., Jirutitijaroen, P., Walsh, W.M., 2013.
turbidity-dependent performance. Renew. Sustain. Energy Rev. 45, 379396. Evaluation of transposition and decomposition models for converting global
Gueymard, C.A., Ruiz-Arias, J.A., 2016. Extensive worldwide validation and climate solar irradiance from tilted surface to horizontal in tropical regions. Sol. Energy
sensitivity analysis of direct irradiance predictions from 1-min global 97, 369387. http://dx.doi.org/10.1016/j.solener.2013.08.033.
irradiance. Sol. Energy 128, 130. http://dx.doi.org/10.1016/ Yoon, K., Yun, G., Jeon, J., Kim, K.S., 2014. Evaluation of hourly solar radiation on
j.solener.2015.10.010. inclined surfaces at Seoul by Photographical Method. Sol. Energy 100, 203216.
Gueymard, C.A., Wilcox, S.M., 2011. Assessment of spatial and temporal variability http://dx.doi.org/10.1016/j.solener.2013.11.011.
in the US solar resource from radiometric measurements and predictions from Yordanov, G.H., Midtgrd, O.-M., Saetre, T.O., Nielsen, H.K., Norum, L.E., 2013.
models using ground-based or satellite data. Sol. Energy 85, 10681084. Overirradiance (cloud enhancement) events at high latitudes. IEEE J.
Hay, J.E., 1979. Calculation of monthly mean solar radiation or horizontal and Photovoltaics 3, 271277.
inclined surfaces. Sol. Energy 23, 301307. Yordanov, G.H., Saetre, T.O., Midtgrd, O.-M., 2015. Extreme overirradiance events
Ineichen, P., 1990. Ground-reflected radiation and albedo. Sol. Energy 44, 207214. in Norway: 1.6 suns measured close to 60N. Sol. Energy 115, 6873. http://dx.
Kierkus, W.T., Colbrone, W.G., 1989. Diffuse solar radiationdaily and monthly doi.org/10.1016/j.solener.2015.02.020.
values as affected by snow cover. Sol. Energy 42, 143147. Yoshida, S., Ueno, S., Kataoka, N., Takakura, H., Minemoto, T., 2013. Estimation of
Kondratyev, K.J., Manolova, M.P., 1960. The radiation balance of slopes. Sol. Energy global tilted irradiance and output energy using meteorological data and
4, 1419. performance of photovoltaic modules. Sol. Energy 93, 9099.
Lave, M., Hayes, W., Pohl, A., Hansen, C., 2015. Evaluation of global horizontal Zehner, M., Weigl, T., Hartmann, M., Thaler, S., Schrank, O., Czakalla, M., Knig-
irradiance to plane of array irradiance models at locations across the United Langlo, G., Giesler, B., Becker, G., Mayer, O., 2011. Energy loss due to irradiance
States. IEEE J. Photovoltaics 5, 597606. enhancement. In: Proc. 26th EU PVSEC Conf., Hamburg, Germany.
Liu, B.Y.H., Jordan, R.C., 1963. A rational procedure for predicting the long-term
average performance of flat-plate solar-energy collectors. Sol. Energy 7, 5374.

Solar Energy: Christian A. Gueymard

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Solar Energy: Christian A. Gueymard

Uploaded by

Copyright:

Available Formats

Solar Energy xxx (2017) xxxxxx

Contents lists available at ScienceDirect

Cloud and albedo enhancement impacts on solar irradiance using high-

Instrument 20062013 2014 2015

Dataset Thermopiles Photodiodes

Year Thermopiles Photodiodes

Statistic Thermopile global pyranometer Photodiode global sensor

90), 7 transposition models, 2 alternate inputs for DNI and DIF

Statistic Thermopiles Photodiodes

Statistic Thermopiles Photodiodes

You might also like