Available online at www.sciencedirect.

com

ScienceDirect
Solar Energy 115 (2015) 229–242
www.elsevier.com/locate/solener

Stochastic generation of synthetic minutely irradiance time

series derived from mean hourly weather observation data
J.M. Bright a,⇑,1, C.J. Smith a,1, P.G. Taylor a,b,c, R. Crook a
a
School of Chemical and Process Engineering, Energy Research Institute, University of Leeds, Leeds LS2 9JT, UK
b
School of Earth and Environment, Sustainability Research Institute, University of Leeds, Leeds LS2 9JT, UK
c
Centre for Integrated Energy Research, University of Leeds, Leeds LS2 9JT, UK

Received 22 September 2014; received in revised form 20 February 2015; accepted 21 February 2015

Communicated by: Associate Editor Jan Kleissl

Abstract

Synthetic minutely irradiance time series are utilised in non-spatial solar energy system research simulations. It is necessary that they
accurately capture irradiance fluctuations and variability inherent in the solar resource. This article describes a methodology to generate
a synthetic minutely irradiance time series from widely available hourly weather observation data. The weather observation data are used
to produce a set of Markov chains taking into account seasonal, diurnal, and pressure influences on transition probabilities of cloud
cover. Cloud dynamics are based on a power-law probability distribution, from which cloud length and duration are derived.
Atmospheric transmission losses are simulated with minutely variability, using atmospheric profiles from meteorological reanalysis data
and cloud attenuation derived real-world observations. Both direct and diffuse irradiance are calculated, from which total irradiance is
determined on an arbitrary plane. The method is applied to the city of Leeds, UK, and validated using independent hourly radiation
measurements from the same site. Variability and ramp rate are validated using 1-min resolution irradiance data from the town of
Cambourne, Cornwall, UK. The hourly irradiance frequency distribution correlates with R2 ¼ 0:996 whilst the mean hourly irradiance
correlates with R2 ¼ 0:971, the daily variability indices cumulative probability distribution function (CDF), 1-min irradiance ramp rate
CDF and 1-min irradiance frequency CDF are also shown to correlate with R2 ¼ 0:9903; 1:000, and 0:9994 respectively. Kolmogorov–
Smirnov tests on 1-min data for each day show that the ramp rate frequency of occurrence is captured with a high significance level of
99.99%, whilst the irradiance frequency distribution and minutely variability indices are captured at significances of 99% and 97.5%
respectively. The use of multiple Markov chains and detailed consideration of the atmospheric losses are shown to be essential elements
for the generation of realistic minutely irradiance time series over a typical meteorological year. A freely downloadable example of the
model is made available and may be configured to the particular requirements of users or incorporated into other models.
Ó 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/
licenses/by/4.0/).

Keywords: Irradiance generation; Resource modelling; Minute resolution; Stochastic modelling; Cloud cover

1. Introduction

⇑ Corresponding author. Solar irradiance varies on a minutely time scale (Sayeef

E-mail addresses: pm08jmb@leeds.ac.uk (J.M. Bright), pmcjs@leeds. et al., 2012). The fluctuations are driven by cloud
ac.uk (C.J. Smith), P.G.Taylor@leeds.ac.uk (P.G. Taylor), R.Crook@ dynamics, atmospheric losses (Calinoiu et al., 2014), and
leeds.ac.uk (R. Crook). the transport of airborne pollutants (Vindel and Polo,
1
Address: UK Network of Energy Centres for Doctoral Training,
University of Leeds, UK.
2014a). Changes in irradiance that occur on the same time

http://dx.doi.org/10.1016/j.solener.2015.02.032
0038-092X/Ó 2015 The Authors. Published by Elsevier Ltd.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
230 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

Nomenclature

Latin alphabet Greek alphabet

C cloud coverage fraction ðC=10Þ a coefficient defined by xmax
C8 cloud coverage in okta (0–8) b single power law exponent
epm elements per minute in matrix b tilt angle from horizontal of inclined plane
f white-noise multiplier for k c variations d coefficient defined by xmin
G irradiance, specified by subscript (W m2)  minutely fluctuation from k c
i random start point within row vector hi solar incidence angle normal to angled plane
kc clear-sky index (G=Gcs ) hz solar zenith angle
n number of elements within a re-sampled cloud r std. dev. around hourly mean of k c
length row vector w sampling rate
P1 transition probability matrix
P ðxÞ probability of x to occur Subscript
r random variable between 0 and 1 B beam (direct)
R resolution of primary x (100 m/el) B denoting a Boolean matrix
s number of states in Markov process clear clear minute
t time-step in Markov process cloud cloudy minute
u wind speed (ms1) cs clear sky
u10 u measured 10 m above ground (ms1) C denoting cumulative probability in matrix
x horizontal cloud length (m) D diffuse
X state at point t in Markov process i the ith state at time t  1
x cloud cover row vector of 1’s and 0’s j the jth state at time t
xw cloud cover row vector adjusted by w m minutely
z cloud height (m) max maximum value
z0ref roughness length (m) min minimum value
n the nth order of a Markov process
Numerical P panel (arbitrary alignment)
0 used to represent clear sky minute ref value at a reference measurement
1 used to represent a clouded minute s number of states in Markov process
60 used as a conversion for secs to min

scale as changes in electricity demand will impact the bene- literature (Gueymard, 2012). The methodology of interest
fits of storage and self-consumption in a domestic or com- is a sun obscured type approach. This is where the cloud
munity PV system (Marcos et al., 2014). Integrated cover is predicted or determined, thereby implying when
electricity demand, PV supply, and storage simulations the solar beam irradiance will be obstructed. A more com-
must operate on a minutely time scale to capture these plex methodology is outlined and developed by Morf
effects, and therefore require minutely irradiance time ser- (1998, 2011, 2013) where cloud cover is two-dimensionally
ies as an input (Widen et al., 2015; Sayeef et al., 2012; modelled to replicate sky with certain clouded conditions,
Hummon et al., 2012; Cao and Sirén, 2014). Calibrated whilst a random number generator driven model separates
minutely irradiance datasets are generally the output of iso- irradiance into its subcomponents of diffuse and beam.
lated research projects and are often limited in duration, Atmospheric transmission losses from extraterrestrial
measurement consistency, and location. Hourly weather irradiance to global horizontal irradiance is extensively
data, however, is widely collected and made available detailed in literature, however its inclusion on a time series
through national meteorological offices. This hourly data irradiance generation model is less so. Simplistically,
fails to capture the intermittent nature of solar irradiance clouded periods can be subjected to a random variate to
(Sayeef et al., 2012), therefore some solar irradiance models represent these losses (Ehnberg and Bollen, 2005), however
use hourly weather datasets to artificially generate minutely there is scope for a more sophisticated approach.
irradiance time series. Geographically dependent monthly clearness index dis-
The focus of solar irradiance models can vary from pre- tributions can be used to deterministically select the trans-
dicting the future irradiance, to providing a general expected mission losses during clouded periods (Morf, 2013).
irradiance at any location globally. Many examples of these Probabilistic methods are commonly seen to generate irradi-
models have been reviewed, analysed and validated in ance data by stochastically selecting the clearness index
 J.M. Bright et al. / Solar Energy 115 (2015) 229–242 231

based on real observations (Ngoko et al., 2014), or by deter- states for example, an appropriately high-resolution time
mining monthly probability distributions of clearness index scale is required (Darcovich et al., 2015; Torriti, 2014).
(Graham et al., 1988), or stochastically producing cloud This methodology offers an improved input estimation of
cover and applying beam irradiance transmission losses the solar irradiance to these types of studies.
(Aguiar et al., 1988). There are numerous methodologies A validation of the methodology is carried out using
that can be utilised and developed upon. hourly input observation data from Leeds, UK, and with
The methodology presented within this paper is a tem- 1-min input observational data from Cambourne,
poral only sun obscured type model with two distinct sec- Cornwall, UK. Comparisons with the irradiance output
tions, the first generates 1-min resolution cloud cover, have been made against independent radiation observation
and the second calculates the irradiance based on whether data for the same location, and the key features of this
the ensuing minute is clear or cloudy. A distinct aspect of methodology are discussed.
this methodology is the use of a multitude of Markov
chains to stochastically determine future weather condition 2. Methodology
states of pressure, wind speed, cloud height and okta,2
incorporating the weather variation influence of season, An example of the output can be seen in Fig. 1. Fig. 1a is
the diurnal dependency and variations caused by pressure. the mean okta number for the hour, Fig. 1b is the one-di-
Furthermore, the pattern of cloud cover is generated with mensional cloud cover where black indicates the presence
greater complexity than in previous literature by using a of a cloud, and Fig. 1c is the corresponding horizontal glo-
one-dimensional method that considers cloud height, the bal irradiance. The simulated day selected was typical of
speed of the cloud, and the statistical distribution of the June for Leeds, UK.
horizontal cloud length. The irradiance is calculated in The inputs required for this methodology are the hourly
three stages, firstly the atmospheric transmission for each mean weather observation measurements of sea level pres-
minute is determined, secondly the theoretical clear-sky sure (hPa), 10 m wind speed (knots), cloud base height
irradiance is calculated based on earth–sun geometry, and (decameters), and okta number (okta). In order to capture
thirdly the irradiance is broken down into its different com- accurately the probability statistics of a typical
ponents in order to obtain irradiance on an arbitrary plane. meteorological year, a minimum of 10 years of observa-
The most distinct element of the irradiance calculations is tional data are recommended.
the production of clear-sky index probability distributions
based on the okta number, and the deterministic approach 2.1. Cloud size
to applying atmospheric losses minute-by-minute with the
inclusion of an intra-hour dependency. A two-dimensional approach to producing cloud cover
The methodology is geographically flexible, so long as could be used to replicate the measurement type of an okta
the input data exist for the location desired. However,
9
individual simulations at nearby locations would not corre-
Okta Number

6
late due to the non-spatial nature of the methodology. The
model output is a synthetic global solar irradiance time ser- 3
ies upon an arbitrary plane at a 1-min resolution. It is a 0
0 4 8 12 16 20 24
temporal data series only, and does not include a spatial
dimension. Its applications are therefore limited to cases (a)
where the spatial element is not integral, such as small scale
Cloud Cover

cloud
studies where a single high-resolution irradiance data series
input is ideal. The methodology has suggested application
in the improved modelling of small-scale PV supply,
demand, and storage systems, through the calculation of 0 4 8 12 16 20 24

electricity supply on a time scale that matches the demand (b)

flows. Regular demand flows operate with high power
Irradiance (W/m2)

1000
across a small duration, meaning that mean hourly
averages fail to capture electricity peaks in the supply 500
and demand. To appropriately capture electricity flows
within the residence that can accurately calculate efficien- 0
0 4 8 12 16 20 24
cies, self-consumption losses, battery charge and discharge Hours
(c)
2
An okta is the representation of cloud amount reported in eighths, Fig. 1. An example output of a typical day in June for Leeds, UK, that
where zero okta represents the complete absence of cloud, through to 8 visually demonstrates the fundamental working of the methodology. Plot
okta representing full cloud cover; an additional value of 9 okta represents (a) shows the mean okta number for each hour. Plot (b) shows the cloud
full coverage due to fog or other meteorological phenomena (UKMO, cover in a minutely resolution, where black is indicative of cloud present.
2010). Plot (c) shows the corresponding global irradiance on a horizontal plane.
232 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

number (Morf, 2011). The methodology presented, how- where u is the wind speed (ms1), 60 is a conversion from
ever, requires only a single passage of beam irradiance seconds to minutes, and R is the resolution of x (m/
from sun to the solar panel; a representation of the whole element).
sky is not necessarily required. The typical method for xw is the minutely time series cloud cover row vector as a
measuring cloud cover is to use a cloud base recorder function of the wind speed. Hourly periods are selected at
which uses a vertical laser pulse to track cloud presence random from xw and the hourly mean coverage value, C, is
(UKMO, 2010). This measurement device takes readings determined through Eq. (6). C is a fractional representation
in a single direction, the okta number derived from this of the duration of obscured sun, where 100 is clear sky and 10
10
type of measurement is considered an acceptable represen- is fully obscured;
tation of obscured beam irradiance once it is adapted to a
synthetic temporal representation. Unlike previous !,
X
iþ59
methodologies, the cloud cover will be generated to follow C¼ xw 60 ð6Þ
a size distribution for a single linear dimension. i¼nr
Horizontal cloud size distributions are shown to be well-
represented using a single power-law relationship using an where the overline represents the mean, n is the total num-
exponent of b < 2 (Wood and Field, 2011; Stull, 1988; ber of elements in xw , and i is a random start point along
Leahy et al., 2012; Pressel and Collins, 2012; Ray and the length of xw .
Engelhardt, 1992) Hours are selected from xw until there is a cloud cover
database consisting of 1000 examples of an hour of cloud
P ðxÞ ¼ axb ð1Þ
cover for each possible u at each C. During each hour of
where x is the horizontal cloud length, b is the exponent simulation, an hour of cloud cover is selected from the
taken at 1.66 between 0.1 and 1000 km (Wood and Field, database using the stochastically determined values of u
2011), a is a constant and P ðxÞ is the probability that x will and C as a reference.
occur.
Eq. (1) represents the probability of x without horizon- 2.2. Markov chains
tal cloud length limits. Eq. (2) introduces these limits and
to allow for pseudo-random number extraction To implement the stochastic element to the modelling, a
1 Markov process is used. Markov models are a very popular
x ¼ ða þ drÞ 1b
ð2Þ method of stochastic data generation and have been used in
many applications, from wind estimates (Masseran, 2015),
where r is a random variable that is uniformly distributed
solar energy estimations (Bhardwaj et al., 2013; Vindel and
between 0 and 1, and a and d are coefficients defined by
Polo, 2014b; Hocaoglu, 2011), and in weather generation
the upper and lower limits of x as
(Yang et al., 2011).
a ¼ x1b
max ð3Þ A Markovian process is a probabilistic mathematical
method whereby transitions from one state to the next
and are directed by discreet probabilities taken from the statis-
d ¼ x1b
min  a ð4Þ tics of real-world processes. In the case of this methodol-
where xmin is the minimum cloud length and xmax is the ogy, statistics are developed from real observational
maximum cloud length. transitions of mean hourly okta, wind speed and cloud
These equations allow a 1 metre resolution row vector of height. From these statistics, transition matrices can be
cloud cover, x, to be generated with 1s and 0s representing constructed.
single metres of cloudy and clear intervals respectively. A This methodology uses a single order Markov model
signal poly-phase filtering technique (implemented using whereby only one previous time-step, t  1, influences the
the built in Matlab ’resample’ function (Oppenheim transition process from t  1 to t. Higher order transitions
et al., 1999; Matlab, 2012a)) is applied that decimates vec- exist, from first order t  1, through to the nth order t  n.
tor, x, by a sampling rate, w. This converts the cloud length Ngoko et al. (2014) describes how a Markov process
row vector (in m) into a smaller vector that is representative ðX t ; t ¼ 0; 1; 2; . . .Þ that has s allowable states ð1; 2; . . . ; sÞ,
of increasing cloud speeds, and so the resultant vector, xw , is in state j at time t if X t ¼ j. In this first order Markov
process, given that the process is in state i at time t  1,
is a time series.
the probability that it will transition to state j at time t is
The clouds are assumed to be travelling at the concur-
given by discreet probability P ij , this is calculated as
rent wind speed and cloud height for each hour of sim-
ulation. The sampling rate used in the signal poly-phase
P ij ¼ P ðX t ¼ jjX t1 ¼ iÞ ð7Þ
filtering technique is
R These probabilities are stored in a transition matrix P1
w¼ ð5Þ
u  60 and can be represented as
 J.M. Bright et al. / Solar Energy 115 (2015) 229–242 233

2 3
P 11 P 12 . . . P 1s based on the conditions of the current hour. The variables
6P P 22 ... P 2s 7 selected through the Markov process are the okta number
6 21 7
P ¼6
1
6 .. .. .. .. 7 7 for each season during both above and below average pres-
4. . . . 5 sures (8 matrices), wind speed for each season (4 matrices),
P s1 P s2 . . . P ss cloud height for each season (4 matrices), and the diurnal
changes for each season (4 matrices). The transition proba-
Constructing a transition matrix requires the conversion
bilities are calculated through statistical analysis of the
of a variable’s range of magnitudes into discreet states.
transitions between hours within the observation data.
Fortunately for this study all the variables are extracted in
The wind speed and cloud height Markov chains are pro-
appropriate states, for example each okta number
duced accounting for seasonal variations. A Markov chain
ð0; 1; . . . ; 9Þ is its own state. The linear time series of observed
is used for each variable representing each of the four sea-
states must be converted into a statistical representation of
sons, capturing the variability at different times of the year,
transitions. Firstly, the frequency of transitions f st1 ;st must
totalling four chains each. The okta number Markov chains
be found. This is done by tallying every transition of states
also consider the effect of season, with the inclusion of
found in the observational data in the appropriate f st1 ;st
impacts from pressure and diurnal variation. Eight okta
location. Secondly, the frequencies are converted into Markov chains are produced that are split by above and
probabilities through dividing f st1 ;st by the total number below average pressure for each season, and four additional
of occurrences that st1 transitioned, f st1 . In the case where morning okta Markov chains are produced to capture the
the observed frequency of state i transitioning to state j is f ij , diurnal variation for okta transitions between 00:00 and
and where f i is the total number times state i has transi- 05:00 am for each season. The intent is to capture the varia-
tioned. The probability of this transition can be expressed as tion in transition probability that occurs as a result of
P ij ¼ f ij =f i ð8Þ weather changes due to the presence of solar energy. 5 am
is considered the cut-off because it is a typical sunrise in
A pseudo-random number generator is used to deter-
the summer for the applied study locations. 5 h represents
mine st from st1 within the simulation. To do this, the
5 okta transitions and is considered an appropriate duration
cumulative sum of the probabilities P1C is found as for the slight propensity to shift towards an increased okta
2 X s 3
to be represented, Fig. 8 demonstrates the diurnal transition
6 P 11 P 11 þ P 12 . . . P 1n 7
differences. Fig. 2 visually demonstrates the mean okta
6 n¼1 7
6 7 Markov chain for the entire year, whilst the effect of season
6 X s 7
6P P 21 þ P 22 . . . P 2n 7 can be seen in Fig. 11.
6 21 7
P1C ¼ 6
6 n¼1 7
7 The hourly mean sea level pressure is considered to be
6. . . . 7 either above or below average at any hour, this is to
6. . . . 7
6. . . . 7 account for weather condition variation between high
6 X 7
s
4 5 and low pressure systems. The mean pressure is calculated
P s1 P s1 þ P s2 . . . P sn
n¼1 from the observation data and the frequency distribution
To implement the Markov process in Matlab r2012a, a of below and above average pressure durations are
random variate r, evenly distributed between 0 and 1, is
introduced at each transition. P1C is queried against r in a 0.7
0
logical statement, resulting in a Boolean matrix
1 0.6
P1C < r ! P1B ¼ 1 ð9Þ
2
Probability of Transition
Present Okta Number

0.5
where ! denotes a logical if statement, and P1B is the resul- 3
tant Boolean matrix. For example, if each of the cumula- 0.4
4
tive transition probabilities are less than r, a value of 1 is
assigned, else a value of 0 is set. The future state st is then 5
0.3
given by 6
X 0.2
st ¼ 1 þ ðP1B;st1 Þ ð10Þ 7

8 0.1
where the subscript st1 indicates the matrix row to sum
using the previously determined state. In the subsequent 9
iteration st becomes st1 and the process is repeated. 0 1 2 3 4 5 6 7 8 9
Future Okta Number
2.3. Okta, wind speed, pressure and cloud height
Fig. 2. Colour plot illustrating the Markov chain for the mean annual
transition probability of okta number from the present hour to the future
Twenty different Markov chains are utilised to stochas- hour. (For interpretation of the references to colour in this figure legend,
tically select weather variables for the subsequent hour, the reader is referred to the web version of this article.)
234 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

produced. The duration of the pressure system is then Theoretical clear-sky irradiance based on earth–sun geome-
selected using a pseudo-random number generator that fol- try is then calculated and the atmospheric transmission fac-
lows the duration probability distribution. The pressure tor applied to obtain simulated global irradiance. Lastly,
systems always alternate between periods when they are simulated all-sky direct and diffuse irradiance is calculated
above and below average. and then applied to obtain irradiance on an arbitrary
Wind speed measured by weather stations is typically plane.
recorded at 10 m above ground level. For clouds in the
atmospheric boundary layer (below 1 km) the wind speed 2.4.1. Generation of clear sky indices
can be extrapolated using the stochastically selected cloud The atmospheric transmission is parameterised in terms
base height as the target reference height, zref . The extrapo- of the clear-sky index, k c ¼ G=Gcs , where Gcs is a theoretical
lated wind speed is assumed to be the speed at which the cloud-free irradiance value. To account for the varying
cloud travels. To extrapolate the wind speed from the optical thicknesses of clouds, it is appropriate to choose
10 m measurement to a reference cloud height, the follow- k c from a distribution rather than to assign a fixed value.
ing logarithmic profile extrapolation (Best et al., 2008) is To this end, observations of G from the 91 MIDAS3 sta-
used tions that provided observations of both cloud okta and
lnðzref =z0ref Þ hourly global irradiance in the year of 2012 were used.
uref ¼ u10 ð11Þ Hours in which the mean solar zenith angle was greater
lnð10=z0ref Þ
than 80 were rejected to minimise the impact of horizon
where uref is the wind speed at the reference height (ms1), effects and exclude hours during which the sun rose or set.
u10 is the wind speed at 10 m above ground level (ms1), zref Values of Gcs were calculated by the DISORT radiative
is the vertical reference height, and z0ref is the roughness transfer solver (Stamnes et al., 2000) in the libRadtran
length of the location (set to 0.14 for rural locations package (Mayer and Kylling, 2005). In the radiative trans-
(Best et al., 2008)). fer calculation, atmospheric profiles of temperature, O3
Above 1 km, the geostrophic wind speeds in the free concentration, precipitable water vapour and surface
atmosphere are influenced by pressure and thermal gradi- albedo were obtained for each month of 2012 from the
ents (UKMO, 1997). Estimating geostrophic wind speeds European Centre for Medium-Range Weather Forecasts
is difficult using mean hourly surface observational data; (ECMWF) ERA-interim reanalysis data (ECMWF, 2014)
presenting one of the limitations of the current state of at a resolution of 1:5  1:5 . Monthly aerosol data was
the methodology. Without knowledge of pressure and ther- provided by the GLOMAP model at a resolution of
mal gradients, alternatives must be used. Excellent methods 2:8  2:8 , specifying scattering and absorption coeffi-
exist such as the wavelet variability model (Lave and cients and asymmetry parameters for 6 bands in the short-
Kleissl, 2013), however the data input required is an irradi- wave spectrum for sulphate, sea-salt, black carbon and
ance profile, an input avoided in the rationale of this meth- particulate organic matter aerosols in four size modes
odology once clear sky statistical analysis has been (Scott et al., 2014). As each simulated hour is to be aligned
performed. Future development is required to truly address with actual data to obtain k c , it is critical to obtain the solar
the cloud speed, the validation results however were found geometry to high accuracy. The latitude and longitude of
to be relatively insensitive to this aspect, and so at this time each MIDAS station was therefore provided to four deci-
a guided estimation is utilised. Using the same range as mal places, from which the effective cosine-weighted solar
used in the wavelet variability model ð0  25 msÞ and typi- zenith angle for each hour was calculated using the
cal geostrophic wind speeds of 12.5 km h1 (Mathiesen Blanco-Muriel et al. (2001) algorithm (accurate to within
et al., 2013), the estimated wind speed at cloud heights 0.01 ) included in libRadtran.
above 1 km is determined allowing variation between the Clear-sky index observations were grouped by
suggested range with the mean found at the typical free corresponding cloud okta for the hour and histograms of
atmosphere wind speed as u>1km ¼ Gammað2:69; 2:14Þ. the observations created (Fig. 3). There were 111090 obser-
For each hour of simulation, the wind speed, cloud vations in total. From this, the distributions from okta 0, 6,
height and okta are stochastically selected from the appro- 7 and 8 were kept. Periods of okta 9, which represents sun
priate Markov chain. From here, an hour of cloud cover at obscured due to fog or other meteorological phenomena,
a minutely resolution can be selected from the cloud cover were treated as overcast hours. For hours where okta is
database, resulting in a temporal Boolean vector represent- in the 1–6 range, the distribution for okta 6 is used, as
ing 1 min cloud cover. the aim is to obtain a distribution for cloud optical thick-
ness for partially cloudy hours which differs from cloud
2.4. Irradiance calculation optical thickness in overcast or nearly overcast hours, but
is relatively unaffected by periods of clear sky. As okta
Once the cloud cover is generated for the entire time ser-
ies desired, the irradiance is calculated. The irradiance 3
MIDAS: the UK Met Office Integrated Data Archive System who
model is in three parts. Firstly, atmospheric transmission have maintained and updated meteorological observations from land and
for each cloudy and cloud-free minute is determined. marine surface stations since 1853 (UKMO, 2014).
 J.M. Bright et al. / Solar Energy 115 (2015) 229–242 235

2 8 distribution. This is justified because the clear-sky index

okta 0 okta 6
4 of overcast skies does not vary extensively (Skartveit and
Olseth, 1992). For partially cloudy hours (okta 1–7), the
1
2 hourly k c;cloud value is allowed to change each hour and is
drawn from the okta 6 distribution if the hourly okta C 8
Probability density

0 0 is 6 or less, and the okta 7 distribution if C 8 ¼ 7. If the

0 0.5 1 1.5 0 0.5 1 1.5
value of k c;cloud selected from the distribution exceeds 1, a
2 3 new value from the distribution is drawn.
okta 7 okta 8
To move from hourly averages of k c;cloud to minutely
2
instantaneous values, a linear interpolation between two
1
1
values of k c;cloud was performed to obtain baseline minutely
values k c;cloud;m . To produce more realistic irradiance pro-
0 0 files, the k c;cloud;m were allowed to fluctuate between intervals
0 0.5 1 1.5 0 0.5 1 1.5
of 6 min. The model allows fluctuations at intervals that are
Clear-sky index
a factor of 60. Allowing fluctuations >10 min intervals did
Fig. 3. Distributions of k c broken down by okta, with analytical not offer similarities in terms of ramp rate occurrences or
distributions fitted that are used in the model. the variability indices, whilst fluctuations at <5 min intervals
saw too much variability. The intention is to capture the gen-
1–5 h include a significant amount of clear sky, it is difficult tle rolling of the k c;cloud;m observed in real profiles. Following
to estimate a sun-obscured k c distribution for these hours. this, minutely clear-sky index variations were introduced for
Furthermore, observations of okta 1–5 are comparatively both cloudy and clear minutes by using a Gaussian white
rare. The distributions were fitted using maximum likeli- noise multiplier based on the hourly cloud okta:
hood estimation (Table 1), except in the case of okta 0
where a normal distribution with mean of 0.99 and stan- k c;m ¼ fk c;m ; f  Nð1; rÞ ð12Þ
dard deviation of 0.08 was selected to provide a good visi- where
ble fit to the peak of the histogram. In the okta 0
histogram, there are several k c values which are much less r ¼ 0:01 þ 0:003C 8 ð13Þ
than 1, indicating that even when no cloud is registered, for cloudy minutes, and
the irradiance reported can be much lower than expected.
This can be explained by the fact that the cloud sensor is r ¼ 0:001 þ 0:0015C 8 ð14Þ
pointing directly upwards whilst the sun is in a different for clear minutes. Eq. (13) gives the greatest variation in
portion of the sky and may be obscured by cloud. This is cloudy minutes for higher okta; Eq. (14) gives the greatest
a particular issue in the UK where the sun is always at least variation in clear minutes with increasing cloud cover.
27 from zenith. To ensure realistic clear-sky values where
the sun is shining, these low values have been rejected and a 2.4.3. Generation of irradiance time series
normal distribution fitted visually to the spike of the his- Clear-sky irradiance values, and their direct and diffuse
togram, which represents genuinely clear hours. components, were obtained from the HELIOSAT method
(Hammer et al., 2003). HELIOSAT is fully described by
2.4.2. Construction of minutely clear-sky index Hammer et al. (2003) and so is not repeated here. Whilst
For each day, a baseline cloudless clear-sky index, clear-sky irradiance from the radiative transfer code can
k c;clear , is selected from the okta 0 distribution, and applied be used, the HELIOSAT method is flexible enough to calcu-
to all minutes where the sun is not obscured. For cloudy late minutely clear-sky direct and diffuse irradiances GB;cs
minutes, the baseline cloud-obscured clear-sky index, and GD;cs for any location of choice in very little com-
k c;cloud , is selected from the appropriate distribution. If putational time.
the sky is overcast for an extended period (hourly okta 8 The global clear sky horizontal irradiance is the sum of
for > 6 hours), the model sets the baseline value of k c;cloud the components
for each interval within the overcast period from the okta
Gcs ¼ GB;cs þ GD;cs ð15Þ
Table 1
Sky states used in the weather generator model and the distribution and the global all-sky irradiance is given by
driving the initial value of k c for each change of state. The normal G ¼ k c;m Gcs : ð16Þ
distribution is parameterised by ðl; rÞ where l is the mean and r is the
standard deviation. The Weibull and Gamma distributions are parame- A further two adjustments are made to ensure that
terised by ða; bÞ where a is a shape parameter and b is a scale parameter.
unrealistic values of G are not seen. If the value of k c;m is
Hourly okta k c distribution for cloud cover less than 0.01, it is set equal to 0.01. At the other end of
66 k c  Nð0:6784; 0:2046Þ the scale, there are many situations where k c;m exceeds 1
7 k c  Weibullð0:5577; 2:4061Þ in the distribution tails. This is most likely to happen at
8 k c  Gammað3:5624; 0:0867Þ
low solar elevations where horizon and ground effects are
236 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

more pronounced, and the division of a small G by a very BADC MIDAS data sets to produce the Markov chains.
small Gcs leads to high values of k c . The largest values of k c In order to validate the model, 12 years of radiation obser-
from the observed irradiances were found to obey the vation data were taken for the same monitoring station for
relationship given by (R2 ¼ 0:9931) comparison. Missing data points are assumed to insignifi-
cantly impact the validation. In cases where duplicate
k c;max ðhz Þ ¼ 27:21expð114 coshz Þ . .. þ 1:665 expð4:494cos hz Þ þ 1:08
hourly measurements exist, the hour that has undergone
ð17Þ
quality control by MIDAS is selected.
so that k c;m is set equal to k c;max if the value drawn from the Figs. 4 and 5 illustrate direct comparisons between mod-
distribution exceeds this upper threshold. elled and observational mean hourly irradiance, averaged
across 10 years. The 10 year mean annual irradiance is also
2.4.4. Decomposition of all-sky irradiance into direct and indicated by the + scatter point at 112.5 W m2 observa-
diffuse components tional and 113.5 W m2 modelled, giving approximately a
The decomposition of global irradiance into direct and 0.9% yearly irradiance overestimate for the location of
diffuse components is necessary in order to calculate the Leeds. Whilst the intent of the model is not to produce
irradiance on a plane of arbitrary alignment. Direct irradi- hourly mean data in this format, averaging the minutely
ance under all sky conditions is related to clear-sky index irradiance generated by the model over hourly timesteps
and clear-sky direct irradiance by an adjustment (Müller shows a strong correlation when compared with the obser-
and Trentmann, 2010) such that vational data ðR2 ¼ 0:9715Þ.
8 Fig. 6 displays the mean 10-year frequency distribution
<0
> k c < 19
69
; of the mean hourly irradiance, normalised to a single year.
2:5
GB ¼ GB;cs ðk c  0:38ð1  k c ÞÞ 19
6 k c 6 1; ð18Þ The correlation between the modelled and observational
>
: 69
GB;cs k c k c > 1: outputs is R2 ¼ 0:9963.

Diffuse all-sky irradiance is the difference between the glo-

bal and direct horizontal values, GD ¼ G  GB . The direct 3.2. Minutely validation for Cambourne, UK
and diffuse horizontal irradiances are translated to the
irradiance incident on a panel of arbitrary inclination and Validation against a minutely irradiance dataset is
aspect GP using the Klucher model (Klucher, 1979) which necessary to confirm that the model can achieve sta-
is found to perform well even when compared to more tistically representative minutely resolution. To demon-
recent models (Gueymard, 2009): strate this, 7 years of minutely radiation data was taken
   from the World Radiation Monitoring Centre – Baseline
1 þ cos b 3 b Surface Radiation Network (WRMC-BSRN) from
GP ¼ GD 1 þ F sin ð1 . . . þ F cos2 hi sin3 hz Þ
2 2 BSRN station number 50, located in Cambourne,
cos hi Cornwall, UK (WRMC-BSRN, 2014). Missing data points
þ GB ; ð19Þ were ignored and deemed to not significantly impact the
cos hz
distributions for comparison. The model inputs were
where hi is solar incidence angle normal to the angled adjusted for the location of the Cambourne weather station
2 using a latitude of 50:2178, longitude of 5:32656, and
plane, and F ¼ 1  ðGD =ðGB þ GD ÞÞ .

3. Validation 500

450
All data processing was performed using the commercial
400
software package Matlab r2012a (Matlab, 2012b). Hourly
Irradiance (W/m )

weather observational data are taken from the British 350 Spr mod
2

Spr obs
Atmospheric Data Centre (BADC) (BADC, 2013) which 300 Sum mod
Sum obs
stores data from MIDAS’ land surface monitoring sta- 250 Aut mod
Aut obs
tions, which are well geographically dispersed across the Win mod
200 Win obs
UK. As monitoring stations are occasionally taken off-line All mod
150
for repairs or upgrades for months at a time, 12 years of All obs

data sets are used to allow at least 10 years data for each 100
variable. 50

0
3.1. Hourly validation in Leeds, UK 5 10 15 20
Time (hours)
For the application of the model, 12 years of hourly Fig. 4. Comparison of the seasonal mean diurnal irradiance profiles and
weather observations for the monitoring site at Leeds the annual mean diurnal irradiance profile between 10-years of modelled
Church Fenton (Source ID: 533) were taken from the output and 10-years of validation data from Leeds, UK.
 J.M. Bright et al. / Solar Energy 115 (2015) 229–242 237

700
have a slightly increased frequency of mildly variable days
between variability indices 10–25 as is indicated by the stee-
Mod. Mean Hourly Irradiance (Wm )
−2

600 per slope, whilst having a slightly reduced frequency of

extremely variable days. Table 2 indicates that 94:38% of
500
days analysed with the 2 sample K–S test reject the null
400 hypothesis that the minutely variability indices of each
day of the year observed and modelled cdfs are not from
300
the same sample, when using a significance level of 99%.
200 No annual temporal bias was observed from the days that
Mean Hourly accept the null hypothesis. The comparison of the observa-
100 Mean Annual
tional ramp rate occurrence CDF and that of the model
x=y
0 have an excellent correlation, as shown in Fig. 7b with
0 100 200 300 400 500 600 700
R2 ¼ 1. Both the 1-min interval of ramp-up and ramp-down
−2
Obs. Mean Hourly Irradiance (Wm ) events are captured excellently using this methodology. This
Fig. 5. Comparison scatter plot of each hour of the year, averaged over is furthermore demonstrated with the results from the day-
10 years, for the model output against the observational data. The yearly by-day K–S tests shown in Table 2. All days tested reject the
mean is also included. An x ¼ y line is included for reference. null hypothesis across all tested significance levels with
mean asymptotic p-value of 6:7365e29 . Fig. 7c shows the
1
comparison of the CDFs of the irradiance occurrence fre-
quency between the 1-min generated data and the 1-min
observational data across 7 years. The probability dis-
0.8
tribution functions correlate well at R2 ¼ 0:994 with a very
Cumulative Probability

slight underestimation of the mid-range irradiance occur-

0.6
rences. A possible cause for this is combining the clear
sky indices of okta values 1–6 with the same distribution.
0.4 The day-by-day irradiance frequency distribution compare
well with observational data as seen in Table 2, which shows
0.2 that over 95% of days tested reject the null hypothesis at a
Modelled Hourly significance level of 99%.
Observed Hourly
0
0 200 400 600 800 1000 4. Application and discussion
Mean Hourly Irradiance (Wm−2)
Irradiance is generated over a 10 year simulation using
Fig. 6. Comparison of the mean hourly irradiance frequency distribution
Leeds, UK, input observational data to complete the meth-
across 10 years of modelled and observational data from Leeds, UK.
odology, the same simulation was used in the validation. A
height above sea level of 87 m. 12 years of hourly weather distinctive element of this methodology is the use of 20 dif-
observations for the monitoring site at Cambourne (Source ferent Markov chains in order to capture seasonal, diurnal
ID: 1395) were taken from the BADC MIDAS data sets to and pressure based variations.
produce the appropriate Markov chains. Combined, the Markov chains successfully replicated
Three metrics are used to validate the intermittent nature the okta frequency distribution of the 12 years of observa-
of the model output, the variability index CDF, the irradi- tion data. The mean percentage error between the modelled
ance CDF and the ramp rate CDF. The variability index is and observed okta transitions is 0:03%, which demon-
the ratio between the rate of change of the measured irradi- strates that the overall statistics for Leeds are retained even
ance and the rate of change of the reference clear sky irradi- with 20 separate Markov chains in use, whilst capturing
ance (Stein et al., 2012). The more intermittent the day’s more detailed transition characteristics at certain times of
irradiance, the higher the variability index. Furthermore, a year. The annual correlation of modelled and observa-
the 2-sample Kolmogorov–Smirnov (K–S) test was carried tional okta frequency distribution is R2 ¼ 0:9956, showing
out for each metric, for each day of the year. The subset of a highly accurate reproduction of the location’s okta
each K–S test consisted of 7 of the same day’s minutely statistics.
data. For example, 7 modelled samples of the 1st-January Each type of okta Markov chain (diurnal, pressure, and
represents one subset and is compared against 7 samples season) is analysed through examining the variation away
of the same day from the observational data. from the annual mean okta Markov chain (Fig. 2) to assess
The comparison of the CDFs of the mean daily variabil- the impact each of the different types have on the transition
ity indices for the model output and the measured val- probability. The mean okta Markov chain is produced with
idation data is shown in Fig. 7a. The two curves correlate every transition in the observational data, and not sepa-
well at R2 ¼ 0:9903. The model generated variability indices rated for season, pressure and season.
238 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

(a) (b) (c)

1 1 1

Cumulative Probability
0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2 Modelled

Observed
0 0 0
0 20 40 0 1 2 3 0 500 1000
Daily Variability Index Minutely Ramp Change (frac.) Irradiance Frequency (Wm−2)

Fig. 7. Cumulative probability distribution plots of the (a) daily variability index, (b) the minutely irradiance ramp-up and ramp-down occurrences, and
the (c) minutely global horizontal irradiance frequencies for both the modelled data and the measured data from Cambourne, UK.

Table 2
The percentage of days that reject the null hypothesis when performing the 2-sample Kolmogorov–Smirnov test on 7 of the same day of modelled and
observational CDF profiles of the minutely variability indices, minutely ramp rate occurrences, and the minutely irradiance frequency, tested at increasing
confidence limits.
K–S test significance level 90% 92.5% 95% 97.5% 99%
Variability index 98.63% 98.15% 97.81% 96.51% 94.38%
Ramp rate occurrence 100% 100% 100% 100% 100%
Irradiance frequency 99.11% 98.70% 98.29% 97.40% 95.89%

4.1. Diurnally weighted okta Markov chains with a decrease of between 0.01 and 0.25 for okta 0–6. Okta
7, 8 and 9 however have an increased probability of
The most significant deviation of transition probabilities remaining the same meaning that cloudier weather states
away from the mean is seen with the morning diurnal are longer and more prevalent. Significant probability
dependency type Markov chain. A comparison of the mean increases for an okta state transitioning to a higher okta
of the diurnal Markov chains (so that variations caused by number are seen across all okta states. There is also a slight
pressure and season are excluded) minus the mean okta increase for okta 3, 4 and 8 to transition to okta 0. The
Markov chain is shown in Fig. 8. It illustrates the most sig- physical implication of this is that during the morning per-
nificant deviations from the mean, as the differences in iod (00:00–05:00 am), the current okta value has an
probability below 0:01 are removed from the plot. increased tendency to transition towards an okta value of
There is a very distinct pattern for the probability of an 6–8 or 0, and so cloud state is most likely to develop into
okta number to remain the same from one state to the next, either fully obscured or complete clear sky. This increased
tendency towards both okta extremes during the morning

0
0
Difference in Transition Probability

0.2
BA>AA| Dif. in Prob. |AA>BA

1 0.1
1
2
Present Okta Number

2
Present Okta Number

0.1
3 0.05
3
4 4
0
5 5 0
6 6
−0.1
7 7 −0.05
8 8
−0.2
9 9
−0.1
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Future Okta Number Future Okta Number

Fig. 8. Colour plot indicating the deviation of the diurnally subjective Fig. 9. Colour plot indicating the difference between the above average
okta Markov chain minus the annual yearly mean okta Markov chain. pressure Markov chain (AA) minus the below-average pressure Markov
White space indicates a 6 0:01 change in transition probability. (For chain (BA). White space indicates a 6 0:01 change in transition
interpretation of the references to colour in this figure legend, the reader is probability. (For interpretation of the references to colour in this figure
referred to the web version of this article.) legend, the reader is referred to the web version of this article.)
 J.M. Bright et al. / Solar Energy 115 (2015) 229–242 239

period is not captured when using a single Markov chain is simplified to be either above or below the average pres-
for all times of the year. sure. Fig. 9 shows a slightly different analysis to the diurnal
comparison for simplicity. The mean below average (BA)
transition probabilities are subtracted from the mean
4.2. Pressure weighted okta Markov chains above average probabilities (AA). The darker shade (blue)
represents occurrences where the BA transition probability
The reason for the inclusion of transitions based on is greater than that of AA, for that particular transition
pressure is to attempt to capture variations in weather that only. The lighter shade (green) represents where the AA
are caused by high and low pressure systems. The approach transition probability is greater than that of BA. The most
significant key point to note is that the probabilities for
1.5 okta 0–3 to transition to a cloudier state are consistently
and significantly greater during periods of BA pressure.
The transitions are more stable during AA pressure, with
an increased tendency to remain at the same okta state.
Frequency density

1 For okta 1–2 during AA pressure, there is a slight increase

in probability to become less cloudy. Okta 7–9 during BA
pressure has a higher tendency, þ0:1, to remain cloudy.

0.5
4.3. Seasonally weighted okta Markov chains

The variation due to the different seasons is detailed in

0 Fig. 11, where the deviation from the mean okta Markov
0 0.5 1 1.5
is shown for each of the four seasons. Again, in order to
Clear−sky index
observe the most significant of differences, deviations of
Fig. 10. Histogram of the mean hourly observed values of k c (bar) 0:015 are removed. Autumn variations are the closest to
compared to the values of k c from the model (line). the mean. The frequency of occurrence of okta 2–6 are

Spring Summer
0.06
0 0
1 1
2 2
0.04
3 3
4 4
5 5 0.02
6 6
Difference in Transition Probability

7 7
8 8 0
9 9
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
−0.02

Autumn Winter
−0.04
0 0
1 1
Present Okta Number

2 2
−0.06
3 3
4 4
5 5 −0.08
6 6
7 7
8 8 −0.1
9 9
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Future Okta Number

Fig. 11. Colour plot indicating the deviation of the individual season Markov chains from the yearly mean okta Markov chain. The colour indicates the
difference in transition probability; white space indicates a 6 0:015 change in transition probability. (For interpretation of the references to colour in this
figure legend, the reader is referred to the web version of this article.)
240 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

the lowest and so differences in this region offer the least Model Output Example
impact on the overall model. The tendency for okta 1 to 1000

Irradiance (Wm−2)
remain the same is significantly larger for autumn. Spring 800
transitions are similar to those of autumn, in that the most 600
significant deviations occur during the least sensitive okta 400
range. There is however a distinct shift to favour lower okta 200
values. Summer and winter contain the most significant
0
variations from the mean. This is most important in sum- Day 176 Day 177 Day 178
mer as this is the period receiving the majority of the year’s
irradiance. Summer sees an increase in the probabilities Observational Data for Corresponding Days
1000
along the diagonal (x ¼ y), hinting towards more stable

Irradiance(Wm−2)
transitions, directly contrasting the pattern observed in win- 800

ter. Interestingly, winter has an increased tendency to move 600

towards the two extremes of clear sky and fully obscured. It 400
is worth noting that the differences reported here are specific 200
to the location of Leeds, whilst the differences may be simi- 0
lar in pattern, the seasonal, diurnal and pressure differences Day 176 Day 177 Day 178
are specific to the input observational data.
Fig. 13. Three days of example irradiance profile outputs from the model
(top), and from observational data (bottom). Note this is neither a
4.4. Irradiance predictive method nor a forecasting model and so the synthetic profiles are
not intended to match or recreate observational data. The purpose of this
figure is to demonstrate the model’s output and to show real 1-min
Hourly averages of clear sky indices derived from the resolution observational data irradiance profiles.
model were compared with k c values from the MIDAS data
(Fig. 10). It can be seen that the model recreates bi-modal
structure of the real-world distribution of clear-sky indices.
The peaks from the simulated data are of similar height to Typical un-obscured irradiance outputs during these times
the MIDAS data, although the intermodal spread is lower are around 900 W m2 at Cambourne. The RMSe in
and in particular there are fewer extreme high or low values Fig. 12 at midday in summer is at 14 W m2, offering poten-
in the simulated data. tial mean hourly irradiance error of 1:5%. The vast
Fig. 12 displays the diurnal root mean square error majority of hourly RMSe values fall below 15 W m2. The
(RMSe) performance of the simulated data for each season correlation of 10 years observational and generated mean
and annually, averaged across 10 years. The RMSe is cal- hourly irradiance profiles is R2 ¼ 0:9715.
culated as a direct comparison of 10 years of modelled
mean hourly irradiance data aligned with 10 years of mean
5. Conclusion
hourly observational data. The time of day of most impor-
tance is around midday during the summer season during
This paper presented a stochastic sun obscured type syn-
maximum clear-sky irradiance as this is when the potential
thetic irradiance generation methodology with the inten-
for the largest ramp-rates and peak outputs are found.
tion of taking readily available hourly resolution input
data and generating statistically accurate 1-min resolution
irradiance profiles. One-dimensional cloud cover has been
15 Spr generated so that the horizontal cloud length follows a
Sum power law relationship. Twenty different Markov chains
Irradiance RMSe (Wm )
−2

Aut are used in setting the hourly conditions that determine

10 Win the cloud cover. These Markov chains allow the stochastic
Tot selection of wind speed and cloud height whilst accounting
for seasonal variability, and the selection of cloud cover
accounting for seasonal, diurnal and pressure variability.
5 The use of diurnal Markov chains is shown to have the
most significant influence on the output. The variability
of okta transition due to pressure system and season also
0 allows for greater accuracy in capturing the weather condi-
4 8 12 16 20 24 tions of a geographical location. The distribution of hourly
Hour of Day (hour) clear-sky indices are recreated whilst allowing for minutely
Fig. 12. The diurnal mean hourly irradiance root mean square error clear sky index fluctuations.
(W m2), averaged across 10 simulations. The RMSe is displayed This methodology is validated on an hourly basis using
diurnally and separated by season, the yearly diurnal RMSe is also shown. radiation observations from Leeds Church Fenton, UK;
 J.M. Bright et al. / Solar Energy 115 (2015) 229–242 241

and on a 1-min basis using radiation observations from References

Cambourne, Cornwall, UK. The model was shown to cap-
ture the irradiance frequency distribution with R2 ¼ 0:9963 Aguiar, R., Collares-Pereira, M., Conde, J., 1988. Simple procedure for
generating sequences of daily radiation values using a library of
and the 10-year mean hourly irradiance with R2 ¼ 0:9715. Markov transition matrices. Sol. Energy 40 (3), 269–279.
The 10-year mean diurnal comparison was shown to per- BADC, 2013. British Atmospheric Data Centre – National Centre for
form well for each season, the maximum RMSe during sum- Atmospheric Science – Natural Environment Research Council.
mer midday is at 14 W m2 for a single hour. The largest <http://badc.nerc.ac.uk>.
Best, M., Brown, A., Clark, P., Hollis, D., Middleton, D., Rooney, G.,
RMSe equates to an error during summer height outputs
Thomson, D., Wilson, C., 2008. Small-scale Wind Energy – Technical
of 1:5%. The use of clear sky index frequency distributions Report. Tech. Rep., UK Met Office.
as a function of the okta number allows the typical cloud Bhardwaj, S., Sharma, V., Srivastava, S., Sastry, O., Bandyopadhyay, B.,
characteristics inherent during okta to be produced. The Chandel, S., Gupta, J., 2013. Estimation of solar radiation using a
1-min validation in Cambourne showed the same perfor- combination of Hidden Markov Model and generalized Fuzzy model.
Sol. Energy 93 (0), 43–54.
mance on an hourly resolution as shown for Leeds, however
Blanco-Muriel, M., Alarcón-Padilla, D., López-Moratalla, T., Lara-Coira,
it has been demonstrated that the 1-min irradiance ramp- M., 2001. Computing the solar vector. Sol. Energy 70 (5), 431–441.
rates, irradiance frequencies and the variability indices dis- Bright, J.M., Smith, C.J., Taylor, P.G., Crook, R., 2015. The Bright Solar
tributions correlated very well at R2 values of 1:000; 0:9994, Resource Model. <http://jamiembright.github.io/BrightSolarModel/>.
and 0:9903 respectively. Day by day 2-sample K–S tests Calinoiu, D.-G., Stefu, N., Paulescu, M., Trif-Tordai, G., Mares, O.,
Paulescu, E., Boata, R., Pop, N., Pacurar, A., 2014. Evaluation of
were carried out on the modelled and obscured CDFs of errors made in solar irradiance estimation due to averaging the
the variability index, ramp rate occurrences, and the irradi- Angstrom turbidity coefficient. Atmos. Res. 150 (0), 69–78.
ance frequency; the percentage of days that reject the null Cao, S., Sirén, K., 2014. Impact of simulation time-resolution on the
hypothesis at significance levels of 99% are 94.38%, 100%, matching of PV production and household electric demand. Appl.
and 95.89% respectively. Whilst not a predictive model, Energy 128 (0), 192–208.
Darcovich, K., Kenney, B., MacNeil, D., Armstrong, M., 2015. Control
visual inspection of both the 1-min resolution data dis- strategies and cycling demands for Li-ion storage batteries in residen-
played in Fig. 13 shows how the observational and mod- tial micro-cogeneration systems. Appl. Energy 141 (0), 32–41.
elled irradiance profiles have very similar characteristics ECMWF, 2014. European Centre for Medium-Range Weather Forecasts
during overcast and intermittent moments. – ERA-Interim. Tech. Rep., ECMWF. <http://www.ecmwf.int/en/
This model can be used to generate statistically accurate research/climate-reanalysis/era-interim>.
Ehnberg, J., Bollen, M., 2005. Simulation of global solar radiation based
1-min time series irradiance data from readily available 1-h on cloud observations. Sol. Energy 78, 157–162.
meteorological input data. The generated irradiance can be Graham, V., Hollands, K., Unny, T., 1988. A time series model for Kt
utilised as an input into other applications such as tem- with application to global synthetic weather generation. Sol. Energy 40
poral intermittency studies, small scale grid impacts mod- (2), 83–92.
elling associated with residential PV, and residential PV Gueymard, C., 2009. Direct and indirect uncertainties in the prediction of
tilted irradiance for solar engineering applications. Sol. Energy 83,
integrated battery storage modelling. Furthermore, this 432–444.
methodology is not limited to a minutely resolution and Gueymard, C.A., 2012. Clear-sky irradiance predictions for solar resource
can theoretically be applied to produce secondly time ser- mapping and large-scale applications: Improved validation methodol-
ies, however, in order to have confidence in the accuracy ogy and detailed performance analysis of 18 broadband radiative
of the output, higher quality resolution observation data models. Sol. Energy 86 (8), 2145–2169.
Hammer, A., Heinemann, D., Hoyer, C., Kuhlemann, R., Lorenz, E.,
is required. This methodology could be extended by the Müller, R., Beyer, H., 2003. Solar energy assessment using remote
inclusion of a spatial dimension using geographic smooth- sensing technologies. Rem. Sens. Environ. 86, 423–432.
ing techniques such as those demonstrated in Lave et al. Hocaoglu, F.O., 2011. Stochastic approach for daily solar radiation
(2012), as well as exploring potential improved accuracy modeling. Sol. Energy 85 (2), 278–287.
through the use of higher order Markov models. Hummon, M., Ibanez, E., Brinkman, G., Lew, D., 2012. Sub-hour solar
data for power system modeling from static spatial variability analysis.
The solar resource model has other possible applications In: 2rd Annual Solar Integration Workshop, Lisbon, NREL/CP-
and a freely downloadable example of the model will be 6A20-56204.
provided in the hope that researchers will adopt and adapt Klucher, T., 1979. Evaluation of models to predict insolation on tilted
it for their own purposes (Bright et al., 2015). surfaces. Sol. Energy 23, 111–114.
Lave, M., Kleissl, J., 2013. Cloud speed impact on solar variability scaling
– application to the wavelet variability model. Sol. Energy 91 (0), 11–
Acknowledgements 21.
Lave, M., Kleissl, J., Arias-Castro, E., 2012. High-frequency irradiance
Thanks are given to Dave Allen for his support and fluctuations and geographic smoothing. Sol. Energy 86 (8), 2190–2199,
advice, and to the editor and anonymous reviewers for progress in Solar Energy 3.
Leahy, L.V., Wood, R., Charlson, R.J., Hostetler, C.A., Rogers, R.R.,
their thorough and helpful critiques.
Vaughan, M.A., Winker, D.M., 2012. On the nature and extent of
This work was financially supported by the Engineering optically thin marine low clouds. J. Geophys. Res.: Atmos. 117 (D22).
and Physical Sciences Research Council through the Marcos, J., Storkël, O., Marroyo, L., Garcia, M., Lorenzo, E., 2014.
University of Leeds Centre for Doctoral Training in Low Storage requirements for PV power ramp-rate control. Sol. Energy 99
Carbon Technologies (Grant No.: EP/G036608/1). (0), 28–35.
242 J.M. Bright et al. / Solar Energy 115 (2015) 229–242

Masseran, N., 2015. Markov Chain model for the stochastic behaviors of secondary organic aerosol. Atmos. Chem. Phys. 14 (1), 447–470,
wind-direction data. Energy Convers. Manage. 92 (0), 266–274. <http://www.atmos-chem-phys.net/14/447/2014/>.
Mathiesen, P., Collier, C., Kleissl, J., 2013. A high-resolution, cloud- Skartveit, A., Olseth, J., 1992. The probability density and autocorrelation
assimilating numerical weather prediction model for solar irradiance of short-term global and beam irradiance. Sol. Energy 49 (6), 477–487.
forecasting. Sol. Energy 92, 47–61. Stamnes, K., Tsay, S.-C., Wiscombe, W., Laszlo, I., 2000. DISORT, A
Matlab, 2012a. R2014a Documentation, Signal Processing, Multirate General-purpose Fortran Program for Discrete-Ordinate-Method
Signal Processing, Resample. <http://uk.mathworks.com/help/signal/ Radiative Transfer in Scattering and Emitting Layered Media:
ug/resampling.html>. Documentation of Methodology. Tech. Rep., Dept. of Physics and
Matlab, 2012b. Version 7.14.0.739 R2012a. The MathWorks Inc., Natick, Engineering Physics, Stevens Institute of Technology, Hoboken, NJ,
Massachusetts. USA.
Mayer, B., Kylling, A., 2005. Technical note: the libRadtran software Stein, J.S., Hansen, C.W., Reno, M.J., 2012. The variability index: a new
package for radiative transfer calculations – description and examples and novel metric for quantifying irradiance and PV output variability.
of use. Atmos. Chem. Phys. 5, 1855–1877. In: World Renewable Energy Forum, Denver, CO.
Morf, H., 1998. The stochastic two-state solar irradiance model STSIM. Stull, R., 1988. An Introduction to Boundary Layer Meteorology.
Sol. Energy 62 (2), 101–112. Atmospheric and Oceanographic Sciences Library. Springer.
Morf, H., 2011. The stochastic two-state cloud cover model STSCCM. Torriti, J., 2014. A review of time use models of residential electricity
Sol. Energy 85 (5), 985–999. demand. Renew Sustain Energy Rev 37 (0), 265–272.
Morf, H., 2013. A stochastic solar irradiance model adjusted on the UKMO, 1997. UK Meteorological Office – Source Book to the
Angstrom–Prescott regression. Sol. Energy 87 (0), 1–21. Forecasters Reference Book. Meteorological Office.
Müller, R., Trentmann, J., 2010. Algorithm Theoretical Baseline UKMO, 2010. UK Met Office – Observations: National Meteorological
Document Direct Irradiance at Surface. Tech. Rep., EUMETSAT Library and Archive Fact Sheet 17 – Weather Observations Over land.
Satellite Application Facility on Climate Monitoring. Tech. Rep., UK Met Office.
Ngoko, B., Sugihara, H., Funaki, T., 2014. Synthetic generation of high UKMO, 2014. UK Met Office Integrated Data Archive System (MIDAS)
temporal resolution solar radiation data using Markov models. Sol. Land and Marine Surface Stations Data (1853-current). NCAS British
Energy 103 (0), 160–170. Atmospheric Data Centre. <http://badc.nerc.ac.uk>.
Oppenheim, A.V., Schafer, R.W., Buck, J.R., 1999. Discrete-time Signal Vindel, J., Polo, J., 2014a. Intermittency and variability of daily solar
Processing, second ed. Prentice-Hall, Inc., Upper Saddle River, NJ, irradiation. Atmos. Res. 143 (0), 313–327.
USA. Vindel, J., Polo, J., 2014b. Markov processes and Zipf’s law in daily solar
Pressel, K.G., Collins, W.D., 2012. First-order structure function analysis irradiation at earth’s surface. J. Atmos. Sol.-Terrestrial Phys. 107 (0),
of statistical scale invariance in the AIRS-observed water vapor field. 42–47.
Climate 25, 5538–5555. Widen, J., Carpman, N., Castellucci, V., Lingfors, D., Olauson, J.,
Ray, J., Engelhardt, F., 1992. Produced Water: Technological, Remouit, F., Bergkvist, M., Grabbe, M., Waters, R., 2015. Variability
Environmental Issues and Solutions; [Proceedings of the 1992 assessment and forecasting of renewables: a review for solar, wind,
International Produced Water Symposium, Held February 4–7, 1992, wave and tidal resources. Renew. Sustain. Energy Rev. 44 (0), 356–375.
in SanDiego, California]. Environmental Science Research. Plenum Wood, R., Field, P.R., 2011. The distribution of cloud horizontal sizes. J.
Press. Clim. 24, 4800–4816.
Sayeef, S., Heslop, S., Cornforth, D., Moore, T., Percy, S., Ward, J.K., WRMC-BSRN, 2014. World Radiation Monitoring Center – Baseline
Berry, A., Rowe, D., 2012. Solar Intermittency: Australia’s Clean Surface Radiation NetworkPANGEA – Data Publisher for Earth and
Energy Challenge. Tech. Rep., CSIRO. Environmental Science. <http://bsrn.awi.de/>.
Scott, C.E., Rap, A., Spracklen, D.V., Forster, P.M., Carslaw, K.S., Yang, H., Li, Y., Lu, L., Qi, R., 2011. First order multivariate Markov
Mann, G.W., Pringle, K.J., Kivekäs, N., Kulmala, M., Lihavainen, H., chain model for generating annual weather data for Hong Kong.
Tunved, P., 2014. The direct and indirect radiative effects of biogenic Energy Build. 43 (9), 2371–2377.