Solar Energy 174 (2018) 606–616

Contents lists available at ScienceDirect

Solar Energy
journal homepage: www.elsevier.com/locate/solener

Quantifiers for the solar irradiance variability: A new perspective T

⁎
Robert Blaga, Marius Paulescu
Faculty of Physics, West University of Timisoara, V Parvan 4, 300223 Timisoara, Romania

A R T I C LE I N FO A B S T R A C T

Keywords: This paper focuses on variability in solar irradiance time series. Six quantifiers for the solar irradiance varia-
Solar radiative regime bility, very different in their nature, are analyzed. One of them, based on the cumulative distribution function of
Clearness index the increments of clearness index time series, is developed in this study. The new quantifier is obtained by
Sunshine number integrating the complementary cumulative distribution function over all values of the increments. The same
Variability criteria
level of variability is expressed by different quantifiers of different magnitudes. In order to surpass this obstacle a
normalizing procedure is applied. This is a key task in comparing the output of different quantifiers toward a
unique standard in the evaluation of the solar irradiance variability. As application, a new multi-parameter
ranking procedure for classifying the days according to the solar irradiance variability is introduced.

1. Introduction 13.2-MW PV plant in Nevada. A severe event that changed the output of
a pyranometer by 80% in 60 s led to a 50% change of the output power
The large-scale penetration of photovoltaic (PV) plants into an in the same time. Therefore, understanding the variability of solar ir-
electric grid is firmly limited by the uncontrollable variability of solar radiance at various time scales is a key issue for an efficient perfor-
irradiance at the ground level. This variability may cause erratic var- mance of a solar plant.
iations in the output power of a PV plant at different time scales, which Generally, the actual level of solar irradiance at the Earth’s surface
are further propagated on the grid as flickers of voltage and frequency. results from the synergistic action of two factors: a deterministic one
On the other hand, the fatigue of the PV modules materials may also be associated with the Earth’s movement and a stochastic one associated
accelerated by the solar irradiance fluctuations, which induces alter- with moving clouds. A huge effort has been invested over time for ac-
nating thermal regimes (Tomson 2010). All the above motivate the curately quantifying the stochastic component of solar irradiance. This
importance of understanding the solar irradiance variability at different component is isolated by means of the instantaneous clearness index,
temporal scales. A recent study by Perez et al. (2016) concludes that defined as the ratio of solar irradiance measured at the ground level to
considering the fundamentals of spatial and temporal scales in devel- that at the top of the atmosphere (see Section 2.1 for a detailed defi-
oping mitigating solutions for the variability of solar resources re- nition). Woyte et al. (2007) argue that an analysis of the clearness index
presents a prerequisite in order to maximize effectiveness and minimize variability must focus on the amplitude, persistence and frequency of
costs of PV plants integration in an electrical grid. the fluctuations. Since the time series of the clearness index exhibit no
Compact cloud fields lead to a low frequency variation of solar ir- periodicity, this information cannot be retrieved by means of Fourier
radiance causing a significant steep increase/decrease in the PV output analysis. Instead, a localized spectral analysis based on wavelet trans-
power. The transition occurs between two states, each of them being formation allows the decomposition of the clearness index into ortho-
stable for a quite long period of time. Scattered clouds may cause high normal components, each of them representing a specific scale of per-
frequency variation in solar irradiance inducing a massive fluctuation sistence. Peled and Appelbaum (2013) quantified the solar irradiance
in the output power of a PV plant. For fast moving clouds, changes in fluctuations using a combination of a statistical approach and wavelet
solar irradiance measured by a pyranometer can exceed more than half analysis. These authors reported a tool for converting the decomposed
of its peak in seconds. The time taken for a moving cloud to shade an components of the clearness index into useful forecasts for the operator
entire PV system depends on various factors, including the PV system of the electric grid with high density of embedded PV generation. The
size and cloud speed. Mills et al. (2011) showed that a 75% solar ir- stochastic component of solar irradiance can also be isolated by means
radiance ramp in 10-seconds measured by a pyranometer was asso- of the clear-sky index, defined as the ratio of the solar irradiances
ciated with 20% variation in output power in the same 10-second in a measured at the ground level to that estimated under clear-sky

⁎
Corresponding author.
E-mail address: marius.paulescu@e-uvt.ro (M. Paulescu).

https://doi.org/10.1016/j.solener.2018.09.034
Received 19 June 2018; Received in revised form 29 August 2018; Accepted 13 September 2018
0038-092X/ © 2018 Elsevier Ltd. All rights reserved.
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

conditions. Based on the clear-sky index Perez et al. (2011) presented unstable, the border between stable and unstable is marked by un-
an empirical model for quantifying the short-term variability using certainty.
satellite-derived hourly solar irradiance data. Their model introduces This paper focuses on the characterization of the stochastic nature of
the standard deviation of the global irradiance clear-sky index and the a solar irradiance series from various perspectives. Six different tech-
mean change in clear-sky index between two consecutive intervals as niques for classifying days according to the solar irradiance variability
quantifiers for intra-hour variability of the solar resource. pattern are analyzed. Among the novelties reported in this study we
There are different ways for classifying days from a meteorological mention here the following: (1) A new quantifier is introduced, based
perspective. Common examples are: the daily average temperature la- on the cumulative distribution function of the increments of the
bels a day as “cold” or “warm”, the daily precipitation level labels a day clearness index. The new quantifier is obtained by integrating the
as “wet” or “dry”, the combination of the ambient temperature and complementary cumulative distribution function and integrating over
relative humidity labels a day as “comfortable” or “uncomfortable”. all existing values of the increments. (2) A normalizing procedure of
Many quantifiers have likewise been defined for classifying different different quantifiers is proposed. Note that different quantifiers express
intervals of time (mainly at daily or hourly scales) according to the the same level of variability at different scales (in a mathematical
variability/stability in the solar irradiance time series. A procedure for sense). Thus, the development of an accurate procedure for comparing
accurate classification turns out to be a useful tool for many applica- different quantifiers is a key task toward a unique standard in the
tions: generation of a solar irradiance series like the typical meteor- evaluation of the solar irradiance variability; (3) A new multi-para-
ological year (Cebecauer and Suri 2015), forecasting solar energy meter ranking procedure for classifying days according to the solar ir-
(Paulescu et al. 2014) or the synthetic generation of high temporal radiance variability pattern is introduced.
resolution solar radiation data (Polo et al. 2011). In particular, the The structure of the paper is as follows. In Section 2 the raw data
decomposition of solar irradiance into its direct and diffuse components used in this study are presented and the post-processed quantities
is strongly dependent on the type of the solar radiative regime (clearness index and sunshine number) are defined in detail. In Section
(Paulescu and Blaga 2016, 2018). 3, six quantifiers for the solar irradiance variability pattern (stability
Maafi and Harrouni (2003) introduced fractal measures for daily index, standard deviation of increments, number of fronts, integrated
solar irradiance variability, aiming to classify days into three groups: cumulative distribution function, sunshine stability number and fractal
clear sky, partly-cloudy sky and overcast. Soubdhan et al (2009) clas- dimension) are defined and assessed. In Section 4 the quantifiers are
sified the daily distribution of the clearness index in four classes using a standardized and a multi-parameter ranking procedure for classifying
mixture of Dirichlet distributions. The sequential analysis of the time the solar irradiance variability based on the aggregated behavior of the
series suggested that the solar radiative pattern is governed by a hidden six quantifiers is presented. The ability of the various quantifiers to
Markov chain with four states. Tomson et al. (2008) used criteria based capture different types of variability is analyzed. The main conclusions
on the notion of global solar irradiance increment for characterizing a of this study and an outlook on the standardization of the definition of
given interval of time as stable or unstable. An increment is defined as stable and unstable solar irradiance pattern are given in Section 5.
the difference between two subsequent measurements in a time series.
The magnitude of large-scale fluctuations is assessed by measuring their 2. Data
positive and negative fronts (Tomson 2010). A positive/negative front
is defined as an event with monotonous increase/decrease in the solar Global G and diffuse Gd solar irradiances recorded at the Solar
irradiance series with at least one increment greater than Platform of the West University of Timisoara (Solar Platform 2017) are
50 W m−2 s−1. Thus, the solar irradiance series appears as a stochastic used in this study. The town of Timisoara (latitude 45°46′N, longitude
process including positive and negative fronts and small-scale fluctua- 21°25′E, altitude 85 m asl) has a warm temperate climate, fully humid,
tions in between. Paulescu and Badescu (2011) introduced a binary with warm summer, typical for the Pannonian Basin (Köppen climate
parameter, the sunshine stability number, for assessing the variability a classification Cfb, based on the Kottek et al. (2006) digital world map
solar irradiance time series. Basically, the sunshine stability number on climate classification). Measurements on the Solar Platform are
counts how many times the Sun is covered (or uncovered) by clouds in a performed all day long at equal time intervals of PleaseCheck . Del-
given time interval. Classifying the days from the view-point of the taOHM LP PYRA 02 first class pyranometers which fully comply with
stability of their radiative regime is performed by using the daily ISO 9060 standards and meet the requirements defined by the World
average value of the sunshine stability number. Meteorological Organization are employed. The sensors are connected
In recent years, special attention has been paid to characterizing the to a data acquisition system based on a National Instruments PXI
solar irradiance variability at small temporal scales. Lave et al. (2015) Platform.
introduced a metric for quantifying high-frequency solar irradiance
variability, called the variability score from the ramp rate distribution. 2.1. Clearness index
The ramp rate in a solar irradiance time series is computed using a
definition based on moving averages. Schroedter-Homscheidt et al. The instantaneous clearness index is defined as follows (Liu and
(2018) proposed an algorithm for classifying the hours within a day in Jordan, 1960):
respect to the variability of the 1-minute direct-normal solar irradiance.
G
The algorithm is based on a combination of previously proposed kt = ,
Gext (1)
quantifiers. Lohmann (2018) reviewed recently published studies on
the quantification and small-scale averaging of solar irradiance varia- where G and Gext are the horizontal solar irradiance measured at ground
bility in time and space. The author emphasized that even if there are level and estimated at extraterrestrial level, respectively. Gext can be
many articles dealing with solar irradiance variability at small temporal written as function of the solar elevation angle h: Gext = GSC ε sin h , were
scales, there is an acute need for more high-resolution measurements to GSC = 1366.1 W m−2 (Gueymard 2004) is the solar constant and ε is the
robustly validate the existing models. eccentricity correction factor that can be calculated with Spencer’s
The above summary shows that different quantifiers highlight dif- equation (Spencer 1971).
ferent facets of the variability in a solar irradiance series. Currently
there is no general consensus on which a particular quantifier is the 2.2. Sunshine number
most suitable for classifying different intervals of time according to the
solar irradiance variability. Because every quantifier is associated with For an observer placed on the Earth’s surface, the sunshine number
a set of criteria for classifying a given period of time as stable or SSN (t ) is defined as a time dependent random binary variable, as

607
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

follows (Badescu and Paulescu 2011): Table 1

Summary statistics of two clearness index datasets, built with data from August
0 if the sun is covered by clouds at time t
SSN (t ) = ⎧ . and December 2010, respectively.
⎨
⎩1 otherwise (2)
Date Mean Standard Median 25% 75% percentile
Series of SSN values are derived from the series of measured solar deviation percentile
irradiance values by using the World Meteorological Organization
08/2010 0.575 0.107 0.601 0.510 0.653
sunshine criterion (WMO 2008): “the Sun is shining” at time t if direct 12/2010 0.548 0.155 0.534 0.408 0.662
solar irradiance exceeds 120 W m−2. In our notation:
G (t ) − Gd (t )
⎧1 if > 120 W m−2 In this case the standard deviation reads:
SSN (t ) = sin h ,
⎨0 otherwise (3)
⎩ N N 2
1 ⎛ ⎞
where G (t ) and Gd (t ) denote the global and diffuse solar irradiance at σ (dkt ) =
N
∑ |dkt , i |2 −⎜∑ |dkt , i | ⎟
time t, and h is the solar elevation angle. i=0 ⎝ i=0 ⎠ (7)
Raw data, measured in a period of 31 days, from 1 to 31 August N represents the number of data measured in the considered time
2010, were used to conduct this study. Fig. A1 from Appendix A illus- interval.
trates visually the temporal variability of solar irradiance in these Unlike the increments defined in terms of clear-sky index, the in-
31 days. A total number of 96 279 instantaneous recordings were post- crements defined in terms of clearness index may be influenced by the
processed for obtaining the instantaneous values of clearness index and seasonality of the atmospheric transmittance. In order to evaluate this
sunshine number. A larger dataset containing 309,718 lines, built by effect, we compared two clearness index datasets, one recorded in
post-processing 708,480 raw data measured in March, July, September August 2010 and the other in December 2010. The two data sets were
and December 2010, was used to test the main conclusions. All the built only with data measured in sunny conditions (SSN = 1). Note that
measurements associated with h < 5° have been removed from the without clouds on the Sun direction, clearness index is a measure of the
database. There are two reasons for this: (i) the pyranometer’s accuracy atmospheric transmittance. Table 1 presents a summary statistic for the
around sunrise and sunset is questionable and (ii) sin h tends towards two datasets, emphasizing no significant differences between them.
zero for small values of h and, consequently, the measurement errors Therefore, the atmospheric transmittance may add only a small un-
can be amplified above the threshold in Eq. (3), yielding a positive certainty when the increments are evaluated on basis of clearness
SSN (t ) , regardless of the values of G (t ) and Gd (t ) . Raw data and post- index. Beyond this empirical evaluation, we would like to stress that the
processed data used in this study, along with other meteorological increments defined in terms of clearness index (thus influenced by the
parameters, are available online at http://solar.physics.uvt.ro/srms/ atmospheric transmittance) capture naturally the change in solar irra-
index.php?target=download. diance: when a cloud covers/uncovers the Sun, the change in solar ir-
radiance is more intense in the summer with the Sun closer to zenith
3. Quantifiers for the solar irradiance variability than in the winter with the Sun closer to horizon.

Six quantifiers for solar irradiance are defined in this section:

3.2. Stability index
standard deviation of the increments, stability index, number of fronts,
integrated complementary cumulative distribution function, sunshine
In the simplest way, the variability in a clearness index time series
stability number and fractal dimension. These quantifiers are very dif-
can be quantified in terms of the cumulated increments in absolute value
ferent in nature and each of them captures a distinct facet of the solar
over a larger time interval than the sampling interval of the measure-
irradiance variability.
ments:
N
3.1. Standard deviation of increments
dkΔt = ∑ |dkt,i|.
i=1 (8)
The most popular quantifier for characterizing the solar irradiance
variability is the standard deviation of the increments (Lave et al. A time interval characterized by the increment |dkt , i| is considered
2013). An increment is defined as the value of the difference between unstable if the increment is larger than a threshold value kc , i.e.
subsequent measurements: |dkt , i | > kc . Following the analysis of Tomson and Tamm (2006), which
established a threshold for the increments in solar irradiance time
dktc = kcs, i + 1−kcs, i (4) series, we take this value to be kc = 0.025. This threshold can be ex-
Eq. (4) written in terms of clear-sky index, since the analysis is often tended easily to cumulated increments by multiplying it with the
performed on a time series of the clear-sky index (Lauret et al. 2016). number of increments within Δ t , kc,Δt = N ·kc .
The clear-sky index is defined as: For quantifying the degree of instability of the solar irradiance
pattern on a given day, we introduced a new quantifier, the stability
G
kcs = , index, defined as the percentage of cumulated increments over an in-
Gcs (5) terval Δ t , which are unstable in a given day (Lucaciu et al. 2017):
where Gcs represents the global horizontal irradiance estimated under SI = N (dkΔt > kc,Δt ) N . (9)
clear-sky at the same moment of time when G was measured. Gcs is
usually calculated using a clear-sky solar irradiance model, which As an example, Fig. 1 illustrates the quantities defined by Eqs. (1),
naturally introduces additional uncertainty (see e.g Badescu et al. 2013 (6) and (8), necessary for computing the stability index for
for a discussion on the performance of the clear-sky solar irradiance Δ t = 10 min , using the measurements from day 08/20/2018. Out of a
models). total of 3069 increments, 923 are unstable, while the stability index is
In this study instead of the clear-sky index we use the clearness SI = 31.5%.
index aiming to avoid the uncertainty introduced by the clear-sky
model. The increments dkt are defined similarly to Eq. (4): 3.3. Number of fronts
dki = kt , i + 1−kt , i (6)
Tomson (2010) argues that considering only the largest increments

608
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Fig. 2. Illustration of two positive fronts in terms of (a) clearness index (b)
increments of clearness index. The two instances show different types of fronts:
in the left side the highly-unstable increment is located at the end of the front
while on the right side the highly-unstable increment is located in the middle of
the front.

3.5. Sunshine stability number

Fig. 1. (a) Time series of the clearness index measured on day 08/20/2010 (Eq.
(5)); (b) Time series of the increments of the clearness index (Eq. (6)); (c) The A quantifier for the variability in the solar irradiance variability
cumulated increments over a time interval of Δ t = 10 min (Eq. (8)). The sta- based on the number of changes that SSN (see Section 2.2) exhibits
bility index represents the percent of cumulated increments above the critical during a time interval Δt has been defined by Paulescu and Badescu
value marked with the dashed line. (2011). This is the sunshine stability number SSSN, which is defined as
a random binary variable:
is not sufficient for characterizing the variability at low frequency in a
1 if SSN (t ) > SSN (t −1)
solar irradiance time series. The author suggests to use the number of SSSNΔt (t ) = ⎧ .
⎨
⎩ 0 otherwise (11)
fronts instead of the percentage of the unstable increments. A front is
defined as a series of positive increments with at least one of them being A few comments about the definition of Eq. (11) are useful. SSN
larger than a given threshold. This definition counts the increasing defined by Eq. (2) is a binary variable, with its values equal to zero
fronts only. The total number of fronts in a given period is denoted by when the Sun is covered by clouds, and one when the Sun shines in the
Fr. We set the threshold value at kc = 0.15. The increments above the sky. Thus, SSSN defined by Eq. (11) indicates only the Sun’s appearance
threshold are highly unstable. With the choice of this threshold, only in the sky while the entire sky may or may not be clear of clouds. The
the fronts containing extreme jumps in the solar irradiance are isolated. sum of SSSNΔt (t ) counts the number of times the Sun appears in the sky
In this study, unlike Tomson (2010) we evaluated the number of fronts during Δt. The average value of SSSN during Δt is denoted by SSSN . It
of the clearness index instead of global solar irradiance. ranges between zero (when the values of SSN(t) are all 0 or 1, respec-
Fig. 2 illustrates the above definition of fronts using a sample of data tively, for all time moments t during Δt ) and 1/2 (when the in-
recorded on day 08/20/2010. The number of fronts during the entire stantaneous values of SSN change every two consecutive moments
day is Fr = 38. during Δt ). The solar radiative regime is fully stable in the first case and
fully unstable in the last case. Since SSSN is a measure of the frequency
3.4. Integrated complementary cumulative distribution function of the Sun occurrence in the sky, it quantifies straightforwardly the
solar irradiance variability. Of course, SSSN may be defined equiva-
The cumulative distribution function (CDF) of the cumulated in- lently for counting the Sun’s disappearance from the sky.
crements gives an overall picture of the solar irradiance variability. Fig. 4 illustrates how SSSN operates. Data from the same day 08/20/
Fig. 3a shows the CDF for Δt = 10 min using data from the same day 2010 are used. Broken clouds were recorded in the first half of the day
08/20/2010. One can easily read-off the percentage of stable incre- while the sky was clear in the second half of the day. The number of
ments as the value of CDF at dkΔt = kc,Δt , which in this case is 68.4%. The changes of SSN during the first half of the day is 43 indicating a high
complementary CDF gives the percentage of increments which are variability of the solar radiative regime while the number of changes of
above a given value. Integrating this quantity, we obtained a new SSN during the second half was 4 indicating a stable regime. For the
quantifier for the variability of the solar irradiance variability. This is whole day, SSSN is equal to 0.0153.
the integrated complementary cumulative distribution function ICCDF,
with the shorthand notation I, expressed as follows: 3.6. Fractal dimension

I= ∫ (1−CDF ) d (dkΔt ) (10) The fractal dimension D of an object depends on how much surface
area (or volume) the object “fills” (Dubuc et al. 1989). A fractal di-
where the integration is taken over all available values of the cumulated mension D(Δt) may be assigned to a solar irradiance G(t) time series
increments. This represents the shaded area in Fig. 3b. For day 08/20/ during a time interval Δt. Because the time series represents a discrete
2010, we have I = 0.76, as Fig. 3b shows. set, the fractal dimension will be a fractional number between 1 and 2.

609
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Fig. 3. (a) The cumulative distribution function of the increments and (b) the complementary cumulative distribution function of the cumulated increments
(Δt = 10 min ). Data from day 08/20/2010 are considered.

In this study we calculated D(Δt) using the method proposed in

Harrouni (2008), where the curve described by G(t) is covered by
rectangles (Fig. 5). For N different time moments tj separated by the
same interval, the area S(Δτ, Δt) of this coverage is given by:
N −1
S (Δτ , Δt ) = ∑ |G (t + Δτ )−G (t )| Δτ
t=0 (12)

The fractal dimension D(Δt) represents the slope of the following

straight line:

S (Δτ , Δt ) ⎞ 1
ln ⎛ = D (Δt ) ln ⎛ ⎞ + C , Δτ → 0.
⎝ Δτ 2 ⎠ ⎝ Δτ ⎠ (13)

Usage of Eq. (13) requires choosing different values of Δτ and

computation of the associated surface area S(Δτ, Δt).
Fig. 6 illustrates the computation of the fractal dimension for the
day 20. We plot the left and right-hand side of Eq. (13) for different
time intervals Δτ, ranging from the instantaneous time step to the half
of the day. A line was fitted on the resulting scatter plot, using the

Fig. 4. (a) Global solar irradiance (b) Sunshine number and (c) Sunshine sta-
bility number during the day 09/20/2010.

Fig. 5. Covering the global solar irradiance signal with rectangles.

Fig. 6. Illustration of the fractal dimension evaluation for the day 08/20/2010.
Δ t is expressed in quanta of 15 s.

610
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

smallest squares method. The fractal dimension D results as the slope of such by at least 4 of the 6 indicators. Otherwise, the day is classified as
this line. Note that as we increase the measuring frequency, the fractal intermediate. Table 2 shows the classification of the days of August
dimension converges to one, which is the correct dimension for a line. 2010 according to the proposed multi-parameter ranking procedure
This can be seen already for sub-minute intervals (last four dots in against the classification made by the individual quantifiers.
Fig. 6), which suggests that the fractal dimension is not efficient for A stable day may would be either sunny or overcast. The solar ir-
capturing high frequency fluctuations in solar irradiance time series. radiance pattern in these days varies smoothly (smoother in the sunny
days) but at different magnitudes (Fig. 8). A question arising at this
point is whether the quantifiers make distinction between these two
4. A new multi-parameter ranking procedure for classifying days
very different states of the sky? The answer is in Table 3, where the
values of the quantifiers are listed for the four days whose solar irra-
4.1. Standardizing the quantifiers
diance patterns are plotted in Fig. 8. It can be seen that there are no
significant differences between the values taken by a quantifier in a
The quantifiers defined in Section 3 are of different magnitudes. In
sunny day and in an overcast day.
order to compare the quantifiers, we have to bring them to the same
scale. This is typically done by standardizing the distributions resulting
4.3. Quantifying the types of solar radiation variability
from applying the quantifiers on a dataset. The procedure consists of
subtracting the mean from the values of the quantifiers and dividing the
Fig. 9 shows a sample of six days from our dataset presented in
result by the standard deviation. For a generic quantifier A, the stan-
Appendix A, that illustrate different aspects of the solar irradiance
dardizing procedure runs as follows (see e.g. Montgomery and Runger
variability. There are three distinct types of variability experienced by
2011): (1) Computing the mean value of the quantifier
N the solar radiative regime, which are clearly visible in Fig. 8:
1
〈A〉 = N
∑ Aj ≡ μ . In our case N signifies the number of days (N = 31
j=1
for August); (2) Computing the standard deviation σ = 〈A2 〉−μ2 where • Intermittency. The state of the sky is rapidly alternating between
N
two distinct regimes: the Sun is shining and the Sun is covered by
1 ∼
〈A2 〉 = N ∑ Aj2 and (3) Standardizing the quantifier A = (A−μ) σ . clouds. Generally, the intermittency is caused by scattered opaque
j=1 clouds (day 9 in Fig. 9).
It is worth noting that the standardization process is not fully effi-
cient when the initial distribution is not a normal one. Among the six • Fluctuations. The state of the sky changes at high frequency but
solar irradiance does not increase/decrease so steeply like in the
quantifiers, only the fractal dimension is characterized by a symmetric
intermittency class. It is produced by broken clouds which, are
distribution while the others are characterized by skewed distributions.
somewhat translucent. Fluctuations are often accompanied by cloud
A Box-Cox transformation could be used to bring the data closer to a
reflections (see e.g, Inman et al. 2016) (day 28 in Fig. 9).
normal distribution (see e.g. Osborne, 2010). Nevertheless, as we will
see further, the distribution of the standardized quantifiers is close • Spikes. The state of the sky is generally overcast with occasional
broken clouds producing low frequency but high amplitude varia-
enough to the normal one. Thus, the normalized quantifiers provide a
tions in the solar irradiance at ground level (day 25 in Fig. 9).
reasonable framework for classifying days.
Generally, a day experiences both stable and unstable periods.
4.2. The day degree of instability During an unstable period, the solar irradiance pattern is a result of a
combination of the three types of variations (intermittency, fluctuations
In this section, we propose a multi-parameter ranking procedure for and spikes). For example, day 1 starts with a relatively stable overcast
classifying days according to the stability of the solar radiative regime. sky, then it passes through a highly fluctuating state, then it returns to a
This procedure classifies the days into three categories: stable, inter- stable overcast state, ending in a mostly stable clear-sky state. A dif-
mediate and highly-unstable. The intermediate category includes days ferent example is day 13, which starts with a stable period in the
characterized by different degree of instability. Differently, the days morning, followed by a period with fluctuations, a stable interval with
from the last category are characterized by a high degree of instability mild variations, a mostly overcast period with spikes, an intermediate
only. In order to clearly differentiate the two categories and to avoid mixed variability period and ends with a stable interval. This illustrates
any confusion we have named the last category “highly-unstable” and that in order to accurately characterize the solar irradiance pattern
not simply “unstable”, as intuition would suggest. In the following the during a day, a quantifier must be sensitive to all types of variability. In
new procedure is introduced and illustrated using the entire dataset the following, the sensitivity of the six quantifiers considered in this
from August 2010. study is assessed.
Fig. 7 shows in the same plot all six quantifiers in a standardized Firstly, it is useful to compare the values of the quantifiers on day 9
form. The threshold values below/above which a standardized quanti- (intermittency type) and on day 28 (fluctuations type). Table 2 shows
fier classifies a day as stable/highly-unstable were established empiri- that the day 9 is classified by σ(dk) and Fr as highly unstable and as
cally at Q− = −0.675/Q+ = +0.675 i.e. ∓ 67.5% of one standard devia- intermediate by SSSN and SI. The day 28 is classified as highly unstable
tion. This way, roughly 25% of days are classified as being highly- by all four quantifiers. Note, however, the large difference in the values
unstable, 25% are classified as stable, while 50% of days are designated of SSSN and SI for day 28 in comparison with those on day 9, while
as having intermediate variability. The same classification can be ob- σ(dk) and Fr take roughly the same values for both days, as it can be
tained if the stability index thresholds in non-standardized form are set seen in Fig. 7. This demonstrates that σ(dk) and Fr are suitable quan-
to SI = 0.05 and SI = 0.15, respectively, meaning that stable days are tifiers for characterizing days with high-frequency variability and often
those which have unstable increments below 5%, while highly-unstable small amplitude fluctuations, while SSSN and SI capture very well the
days are those which have unstable increments more than 15%. intermittency type of variability with lower frequency variability and
Visual inspection of Fig. 7 shows that the stable and highly-unstable large amplitude fluctuations.
days are classified as such by most quantifiers. Some days are classified There are also specific days when one of the quantifiers takes sig-
in different categories by different quantifiers. For example, the day 4 is nificantly higher values than all others. An example is day 25, which is
categorized as highly-unstable by four quantifiers (I, SI, D(G) and SSSN ) characterized by low frequency, slowly growing and high amplitude
and of intermediate variability by two (σ (dk ) , Fr). The multi-parameter variations (spikes). The SSSN classifies this day as highly unstable,
ranking procedure proposed in this work postulates a day as being while the other quantifiers as intermediate. The fractal dimension
certainly in the stable or highly-unstable category if it is classified as seems completely inadequate for characterizing the high-frequency

611
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Fig. 7. Classification of days by the six normalized quantifiers. The threshold values for a stable (Q− ) and a highly-unstable (Q+ ) day are indicated with dashed lines.
Data from August 2010 are used.

Table 2
Classification of the days of August 2010 into stable (green), intermediate (yellow) and highly unstable (red) by the individual quantifiers and
the multi-parameter ranking procedure (MR).

fluctuations, in turn capturing low-frequency variations (on the time-

scale of hours, i.e. alternations between clear and overcast sky), as is the
case of days 13 and 18 (Fig. 9).
Schroedter-Homscheidt et al. (2018) showed that, at hourly time
scale, some quantifiers present similar behavior in pairs or groups. In
order to assess the relative behavior of our six quantifiers at daily time
scale, we computed the correlation matrix of the quantifiers values,
obtained from the entire dataset. The results are listed in Table 4. It can
be observed that the fractal dimension does not have a significant
correlation with any of the other quantifiers. The SSSN and SI show a
large cross-correlation. Similarly, σ(dk) and Fr show also a large cross-
correlation. The new quantifier I shows a strong correlation (over 0.9,
see Table 4) with all the above quantifiers (SSSN, SI, σ(dk) and Fr). Out
of all quantifiers, the newly introduced quantifier I presents the most
reliable behavior, classifying all highly unstable and stable days prop-
erly, in the same way as the more complex multi-parameter procedure
proposed in this work.
The above conclusions are confirmed by analyzing the days classi-
Fig. 8. Global solar irradiance in two sunny days (07/02 and 07/16) and two
overcast days (03/11 and 03/23) in 2010.
fication in other four months of 2010: March, July, September and
December. The classification tables (similar to Table 2) for all

Table 3
Values of the quantifiers in two sunny days and two overcast days.
Date State-of-the-sky I Fr SI σ(dk) D(G) SSSN

03/11/2010 Overcast −0.952 −0.571 −1.005 −1.191 −1.297 −1.081

03/23/2010 Overcast −0.900 −0.571 −0.949 −1.067 −1.436 −0.890
07/02/2010 Sunny −0.871 −0.499 −0.976 −0.837 −2.013 −0.735
07/16/2010 Sunny −0.833 −0.499 −0.915 −0.810 −1.034 −0.930

612
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Fig. 9. Days from August 2010 illustrating different types of variability of global solar irradiance. The days 9, 25 and 28 are characterized by well-defined types of
variability while the days 1, 13 and 18 are characterized by a mix of the three types of variability, described in Section 4.3.

Table 4 standard deviation of increments σ(dk), the number of fronts Fr, the
Correlation matrix of the six quantifiers. integrated complementary cumulative distribution function of the in-
I SSSN SI σ(dk) Fr D(G) crements I, the sunshine stability number SSSN and the fractal dimen-
sion D(G). I is a new quantifier developed in this study. Generally, the
I 1 0.931 0.971 0.966 0.935 0.795 results showed that the SSSN and SI are suitable in capturing the high-
SSSN 0.931 1 0.957 0.870 0.798 0.763 frequency fluctuations, while σ(dk) and Fr are suitable in capturing the
SI 0.971 0.957 1 0.918 0.838 0.823
σ(dk) 0.966 0.870 0.918 1 0.941 0.855
intermittent type of variability. Only the new quantifier I proves its
Fr 0.935 0.798 0.838 0.941 1 0.667 ability of capturing the different facets of the solar irradiance varia-
D(G) 0.795 0.763 0.823 0.855 0.667 1 bility. Since the individual quantifiers express the same level of varia-
bility at different magnitudes, a normalizing procedure of the quanti-
fiers values (inspired from the standardization of the normal
Table 5 distribution) was applied. This was a key task in comparing the output
Percentage of similarity between each individual quantifier and the proposed of the six quantifiers, which opens a way toward a unique standard in
procedure in classifying the test days.
the evaluation of the solar irradiance variability. A new multi-para-
Quantifier I Fr SI σ(dk) D(G) SSSN meter ranking procedure for classifying days according to the solar ir-
radiance variability was introduced. The procedure postulates a day as
Similarity percentage 91.5% 66.1% 83.0% 86.4% 75.4% 78.8%
being certainly in a given category of the solar the irradiance variability
pattern (stable or highly-unstable) if it is classified as such by at least
quantifiers and the proposed procedure are presented in Appendix B. four of the six quantifiers. Out of the six quantifiers, only the newly
Table 5 summarizes the percentage of similarity between each in- introduced quantifier I presented a reliable behavior, classifying all
dividual quantifier and the proposed procedure in classifying the test highly-unstable and stable days correctly according to the procedure
days. It can be seen that the new quantifier I reaches the greatest per- proposed in this work.
centage of similarity, confirming its high ability in classifying the days
according to the solar irradiance variability pattern.
Acknowledgments
5. Conclusions
The authors thank the reviewers for useful comments and sugges-
This study focused on the ability of different quantifiers to capture tions. This work was supported by a grant of the Romanian National
the peculiarities of the solar irradiance variability. Six quantifiers, very Authority for Scientific Research and Innovation, CNCS/CCCDI –
different in their nature, were analyzed: the stability index SI, the UEFISCDI, project number PN-III-P2-2.1-PED-2016-0592.

613
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Appendix A. Variation of solar irradiance in every day of August 2010

See Fig. A1.

Fig. A1. Variation of the global solar irradiance for all days of August 2010 recorded at the Solar Platform of the West University of Timisoara, Romania.

614
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Appendix B. Classification tables

See Table B1.

Table B1
Classification of the days from four months of 2010 into stable (green), intermediate (yellow) and highly unstable (red) by the individual
quantifiers and the multi-parameter ranking procedure (MR). White square indicates no recorded data.

References preprints201806.0308.v1.
Lucaciu, S., Blaga, R., Stefu, N., Paulescu, M., 2017. Quantification of the solar radiative
regime variability based on the clearness index. Ann. West Univ. Timisoara – Phys. 59
Badescu, V., Paulescu, M., 2011. Statistical properties of the sunshine number illustrated (1), 13–17. https://doi.org/10.1515/awutp-2016-0003.
with measurements from Timisoara (Romania). Atmos. Res. 101, 194–204. Maafi, A., Harrouni, S., 2003. Preliminary results of the fractal classification of daily solar
Badescu, V., Gueymard, C.A., Cheval, S., Oprea, C., Baciu, M., Dumitrescu, A., Iacobescu, irradiances. Sol. Energy 75, 53–61.
F., Milos, I., Bada, C., 2013. Accuracy analysis for fifty-four clear-sky solar radiation Mills, A., Ahlstrom, M., Brower, M., Ellis, A., George, R., Hoff, T., Kroposki, B., Lenox, C.,
models using routine hourly global irradiance measurements in Romania. Renew. Miller, N., Milligan, M., Stein, J., Wan, Y.-h., 2011. Dark shadows. understanding
Energy 55, 85–103. variability and uncertainty of photovoltaics for integration with the electric power
Cebecauer, T., Suri, M., 2015. Typical Meteorological Year data: SolarGIS approach. system. IEEE Power Energy M 9, 33–41.
Energy Proc. 69, 1958–1969. Montgomery, D.C., Runger, G.C., 2011. Applied statistics and probability for engineers.
Dubuc, B., Quiniou, J.F., Roques-Carmes, C., Tricot, C., Zucker, S.W., 1989. Evaluating John Wiley & Sons, pp. 501–502.
the fractal dimension of profiles. Phys. Rev. A 39, 1500–1512. Osborne, J.W., 2010. Improving your data transformations: Applying the Box-Cox
Gueymard, C.A., 2004. The sun’s total and spectral irradiance for solar energy applica- transformation. Practical Assessment, Research & Evaluation 15/Nr. 12. Available
tions and solar radiations models. Sol. Energy 76, 423–453. online: http://pareonline.net/getvn.asp?v=15&n=12 (Accessed in May 2018).
Harrouni, S., 2008. Fractal classification of typical meteorological days from global solar Paulescu, M., Badescu, V., 2011. New approach to measure the stability of the solar ra-
irradiance: Application to five sites of different climates. In: Badescu, V. (Ed.), diative regime. Theor. Appl. Climatol. 103, 459–470.
Modeling Solar Radiation at the Earth Surface. Springer, Berlin, pp. 29–54. Paulescu, M., Mares, O., Paulescu, E., Stefu, N., Pacurar, A., Calinoiu, D., Gravila, P., Pop,
Inman, R.H., Chu, Y., Coimbra, C.F.M., 2016. Cloud enhancement of global horizontal N., Boata, R., 2014. Nowcasting solar irradiance using the sunshine number. Energ.
irradiance in California and Hawaii. Sol. Energy 130, 128–138. Conv. Manage. 79, 690–697.
Kottek, M., Grieser, J., Beck, C., Rudolf, B., Rubel, F., 2006. World Map of the Köppen- Paulescu, E., Blaga, R., 2016. Regression models for hourly diffuse solar radiation. Sol.
Geiger climate classification updated. Meteorol. Z. 15, 259–263. Energy 125, 111–124.
Lauret, P., Perez, R., Aguiar, L.M., Tapaches, E., Diagne, H.M., David, M., 2016. Paulescu, E., Blaga, R., 2018. A simple and accurate empirical model for estimating dif-
Characterization of the intraday variability regime of solar irradiation of climatically fuse fraction at 1-minute resolution, submitted to Sol. Energy.
distinct locations. Sol. Energy 125, 99–110. Peled, A., Appelbaum, J., 2013. Evaluation of solar radiation properties by statistical tools
Lave, M., Kleissl, J., Stein, J., 2013. Quantifying and simulating solar-plant variability and wavelet analysis. Renew. Energy 59, 30–38.
using irradiance data. In: Kleissl, J. (Ed.), Solar energy forecasting and resources Perez, R., Kivalov, S., Schlemmer, J., Hemker Jr., K., Hoff, T., 2011. Parameterization of
assessment. Academic Press, Oxford, pp. 149–169. site-specific short-term irradiance variability. Sol. Energy 85, 1343–1353.
Lave, M., Reno, M.J., Broderick, R.J., 2015. Characterizing local high-frequency solar Perez, R., David, M., Hoff, T.E., Jamaly, M., Kivalov, S., Kleissl, J., Lauret, P., Perez, M.,
variability and its impact to distribution studies. Sol. Energy 118, 327–337. 2016. Spatial and temporal variability of solar energy. Foundations and Trends®.
Liu, B.Y., Jordan, R.C., 1960. The interrelationship and characteristic distribution of di- Renew. Energy 1 (1), 1–44.
rect, diffuse and total solar radiation. Sol. Energy 4, 1–19. Polo, J., Zarzalejo, L.F., Marchante, R., Navarro, A.A., 2011. A simple approach to the
Lohmann, G.M., 2018. Irradiance variability quantification and small-scale averaging in synthetic generation of solar irradiance time series with high temporal resolution.
space and time: a short review. Preprints 2018060308. https://doi.org/10.20944/ Sol. Energy 85, 1164–1170.

615
R. Blaga, M. Paulescu Solar Energy 174 (2018) 606–616

Schroedter-Homscheidt, M., Kosmale, M., Jung, S., Kleissl, J., 2018. Classifying ground- Tomson, T., 2010. Fast dynamic processes of solar radiation. Sol. Energy 84, 18–323.
measured 1 minute temporal variability within hourly intervals for direct normal Tomson, T., Tamm, G., 2006. Short-term variability of solar radiation. Sol. Energy 80,
irradiances. PrePub. Meteorologische Zeitschrift. https://doi.org/10.1127/metz/ 600–606.
2018/0875. Tomson, T., Russak, V., Kallis, A., 2008. Dynamic behavior of solar radiation. In: Badescu,
Solar Platform, 2017. Solar Platform of the West University of Timisoara. http://solar. V. (Ed.), Modeling Solar Radiation at the Earth Surface. Springer, Berlin, pp.
physics.uvt.ro/srms. (Accessed on April 2018). 257–281.
Soubdhan, T., Emilion, R., Calif, R., 2009. Classification of daily solar radiation dis- WMO, 2008. Guide to Meteorological Instruments and Methods of Observation, WMO-
tributions using a mixture of Dirichlet distributions. Sol. Energy 83, 1056–1063. No.8/2008. http://www.wmo.int/ (Accessed in April 2018).
Spencer, J.W., 1971. Fourier series representation of the position of the Sun. Search 2, Woyte, A., Belmans, R., Nijs, J., 2007. Fluctuations in instantaneous clearness index:
172–175. Analysis and statistics. Sol. Energy 81, 195–206.

616