The Stochastic Two-State Cloud Cover Model STSCCM

Heinrich Morf

Buechraiweg 47, 5452 Oberrohrdorf, Switzerland

Received 11 May 2010; received in revised form 16 February 2011; accepted 20 February 2011
Communicated by: Associate Editor Christian Gueymard
Abstract
A stochastic model is presented that simulates cloud cover. Implementations reaching from the simple simulation of the steady state
probability distribution of cloud cover up to the generation of cloud cover as a function of time are shown. In the model, the steady state
probability distribution of cloud cover is invariant to the observation area over a wide range. Simulation results are discussed and com-
pared with real world data. Particularly, the steady state probability distributions compare well with the real world.
2011 Elsevier Ltd. All rights reserved.
Keywords: Climatology; Cloud cover; Meteorology; Solar energy; Stochastic modeling; Vertical visibility
1. Introduction
The understanding of the dynamics of solar irradiance is
important for the design and analysis of solar energy sys-
tems. Many of these systems are nonlinear. Consider for
example a solar water heater. It switches its circulation
pump on when there is collectable energy falling onto the
solar collectors, and will cut the circulation if there is no
energy to harvest. The analysis of such systems can hardly
be achieved in any other way than by simulations in the
time domain. Thus, time sequences of solar irradiance must
be available as input for simulation runs. Up to what fre-
quency such time sequences must properly reect reality
depends on the problem at hand. Solar water heaters have
time constants in the order of minutes. Photovoltaic sys-
tems react to irradiation changes within milliseconds. For
detailed systems design and reasonable load matching of
photovoltaic systems in the UK, Craggs et al. (2000) came
to the conclusion that at least 10 min average irradiance
readings are needed. For the design of command and con-
trol systems of large solar thermal plants solar irradiance
input with a sampling interval of less than one second is
used (Gall et al., 2010).
A model that takes higher frequencies of solar irradia-
tion into account, particularly those caused by the interfer-
ence of clouds, was rst proposed in Robinson (1966) by
W. Schu epp. It was then rened by Biga and Rosa
(1980). The model supposes three types of irradiation:
(1) Beam irradiation, only present when the sun is
shining.
(2) Diuse irradiation emanating from the clear part of
the sky.
(3) Diuse irradiation emanating from the covered part
of the sky.
In Morf (1998) we presented the Stochastic Two-State
Solar Irradiance Model STSIM that is based on the ideas
of Biga and Rosa. For this purpose, we implemented the
Stochastic Insolation Function SIF, a function that divides
time into insolated periods and periods when the sun is hid-
den behind clouds, into a model for average instantaneous
solar irradiance by Suehrcke and McCormick (1989). The
model, simple and fast in its execution, proved to be a
valuable tool for the dynamic simulation of solar water
heaters. However, as we mention in the conclusion of Morf
0038-092X/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.solener.2011.02.015

Tel.: +41 56 496 81 20; fax: +41 56 496 81 21.

E-mail address: heinrich.morf@bluewin.ch
www.elsevier.com/locate/solener
Available online at www.sciencedirect.com
Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
(1998): It is evident that there are still other uctuations
than those considered by the STSIM, most likely due to
varying cloud cover.
Therefore, we set out to nd a stochastic function that
simulates cloud cover as a function of time. The researches
lead to climatologists and meteorologists who were princi-
pally interested in the parameterization of the probability
distribution of cloud cover. Over the years, a considerable
selection of mathematical functions as well as methods for
storage and presentation of this distribution has been pro-
posed and used (Henderson-Sellers and McGue, 1991).
Despite an intensive literature review, the necessary sto-
chastic function could not be found. However, in Burger
(1985) the Burger model implements a probability distri-
bution for cloud cover that is the t to the results of a saw
tooth cloud simulation model (Burger and Gringorten,
1984). Burger (1985) reports rms errors
1
from 0.9% to 2.6
% between probability functions constructed with his
model and others constructed with samples collected in
the real world. Henderson-Sellers and McGue (1991)
undertook a profound analysis of the Burger model. They
conrm the excellent t of the Burger model to the real
world.
What drew our attention was the striking resemblance
of probability bar charts of relative daily sunshine duration
generated with the Stochastic Insolation Function SIF
(Morf, 1998), and the probability bar charts of cloud cover
generated with the Burger model shown in Fig. 11 on page
23 in Burger (1985).
2
This resemblance is insofar remark-
able as sunshine is a function of time whereas cloud cover
expands over an area. This gave rise to hope for nding a
stochastic model for cloud cover expanding over time and
area.
We nally succeeded to formulate such a model, which
we call the Stochastic Two-State Cloud Cover Model
STSCCM. Simulated key elements of the model are vertical
visibility vv and cloud cover cc. Vertical visibility
describes the state (covered or clear) of a single point in
the sky. Cloud cover is the fraction of the sky that is cov-
ered by clouds. To adjust the model to the real world mean
and variance of cloud cover and vertical visibility, or a time
sequence of cloud cover need to be known.
Nomenclature
Capital letters
A area (m
2
)
A generic variable
B area (m
2
)
B generic variable
N number of clusters ()
P probability ()
R autocorrelation
T time in the sense of duration (s)
W steady state probability ()
Small letters
cc cloud cover (, oktas, tenths)
d dierential element ()
f probability function ()
f probability density function
i daily relative vertical visibility ()
i daily relative sunshine duration ()
n generic counter ()
n number of grid points ()
t time (s)
v generic counter ()
vv vertical visibility ()
w probability density function of the steady state
x generic free variable
y generic dependent variable
Greek letters
D small element, dierence
l mean, expectation
r
2
variance
s time shift (s)
Special characters
[ ] transition matrix
[ [ absolute value
mean, expectation
intersection
[ condition
Indices
0 where the view is obstructed by clouds (cc)
0 part of the sky that is clear (vv)
1 part of the sky that is covered by clouds (vv)
1 where the view is clear (cc)
cc cloud cover
d daytime (the time span between sunrise and sun-
set)
D one day (= 24 h)
f when overying the cloud eld
i row number of a matrix
j column number of a matrix
obs observed
1
Given two probability functions f
a
(x
1
, x
2
, . . . x
n
) and f
b
(x
1
, x
2
, . . . x
n
),
where P
a
(x
1
) = a
1
, P
a
(x
2
) = a
2
, . . . P
a
(x
n
) = a
n
and P
b
(x
1
) = b
1
,
P
b
(x
2
) = b
2
, . . . P
b
(x
n
) = b
n
, then the rms error is given by

n
i=1
(ai bi )
2
n
_
.
2
As an orientation on the kind of probability bar charts in discussion
here, the reader may refer to Fig. 1.
2 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
2. Vertical visibility and cloud cover
Key elements of the model presented in this paper are
vertical visibility vv and cloud cover cc. With the aim
at more abstraction and elegance these two values are rig-
orously dened as described in the two sections that follow.
2.1. Denition for vertical visibility vv
Consider a partially cloud-covered sky. An observer on
the ground looking up in the zenithal direction to the sky in
a single point view would either see covered sky or clear
sky. Similarly, a geostationary satellite observing in a single
point view down to earth in the straight nadir direction
would have the view down to earth either obstructed by
clouds or not. Depending on the situation covered or
clear we would say that vertical visibility vv at the spot
of the sky under observation is either zero (0) or one (1).
3
Thus, vertical visibility alternates with time between the
two states 0 and 1.
The two descriptions of vertical visibility observation
above are only illustrative examples. In this paper, vertical
visibility is a mere abstraction; real world values are
nowhere used. Nevertheless, a way to determine the param-
eters for the simulation of vertical visibility from cloud
cover data is shown in this paper.
2.2. Denition for cloud cover cc
Cloud cover cc is the fraction of the sky where vertical
visibility is 0. For convenience, a value of zero (0) was
assigned to the clear sky area and a value of one (1) to
the covered sky area.
3
The real world values of cloud cover used in this paper
are estimates by human observers watching the sky. Such
estimates always include individual judgments of the
observers. Thus, they do not reliably match the denition
above. Appendix B enters into the problematics of the
determination and interpretation of cloud cover measure-
ments in the real world.
2.3. The relationship between cloud cover and vertical
visibility
We presume that at any location the probability for
vv = 0 corresponds to the fraction of the sky cc that is
covered by clouds
4
:
P(vv = 0[cc) = cc = P(A[B) (1)
Applying the generally valid relation
P(A B) = P(B)P(A[B); (2)
whereby setting P(B) = f
cc
(cc) dcc, P(A[B) = cc, and inte-
grating over cc yields
P(vv = 0) =
_
1
cc=0
cc f
cc
(cc) dcc; (3)
where f
cc
(cc) is the probability density function pdf of
cloud cover.
Eq. (3) reduces to
P(vv = 0) = cc; (4)
where cc is the mean of cloud cover.
Note, that there are no restrictions to the area over
which cloud cover is measured. Thus, for an area where
the model presented in this paper holds true we state:
At any location and cc taken over any area:
P(vv = 0) = cc = constant (5)
From Eq. (3) one concludes that, zooming in on a cloud
eld that meets Eq. (5) one would see a pdf f
cc
(cc) that con-
verges towards a strictly bimodal shape. And nally, reach-
ing a single point in the eld, one would nd a Stochastic
Visibility Function vv(t).
3. The stochastic vertical visibility model SVVM
In Morf (1998) we introduced the Stochastic Insolation
Model SIM, a mathematical formulation for the alternat-
ing time sequence of periods where the sun is shining and
periods where the sun is obstructed by clouds. The same
model now more appropriately called the Stochastic Ver-
tical Visibility Model SVVM ts even better for the
description of the time sequence of clear sky and covered
sky periods that comprise vertical visibility.
To arrive at the model we ask for the probability P
10
that in a clear period Dt a change from clear to covered will
occur. The totality of the clear time span T
1
consists of n
periods separated by n covered periods. Thus, there will
be a total of n changes from clear to covered. On the other
hand, T
1
could be divided into T
1
/Dt periods of length Dt.
From this one receives for T
1
/Dt n
P
10
=
n
T
1
=Dt
=
n
T
1
Dt =
Dt
t
1
(6)
where t
1
is the mean duration of a clear period.
Following similar reasoning one obtains P
01
, P
11
, and
P
00
. Introducing the hypothesis that the probability for a
change of state within a period Dt depends only on the state
in the immediately foregoing period Dt, the SVVM can
then be formulated as follows:
The sequence of clear and covered periods is generated by
a Homogeneous Recurrent Markov Process with two states
operating in a time step Dt with a transition probability
matrix
3
Note the conventions used: For vertical visibility a value of 1
corresponds to the clear sky condition, whereas for cloud cover a value
of 1 corresponds to the covered sky condition. This corresponds to the
meteorologists convention for relative sunshine duration and cloud cover.
4
The reasoning that leads to this presumption follows the classical
denition of probability as dened by P.S. de Laplace (17491827):
probability = number of positive events divided by the number of possible
events.
H. Morf / Solar Energy xxx (2011) xxxxxx 3
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
[P[
vv
=
P
00
P
01
P
10
P
11
_ _
=
1
Dt
t
0
Dt
t
0
Dt
t
1
1
Dt
t
1
_ _
; (7)
where
Dt dt
In Morf (1998) we presented a thorough analysis of the
model, and we refer the reader to that paper for more de-
tails. In Appendix A of this paper we include a summary
of the important formulae.
4. The Stochastic Cloud Cover Model SCCM
At the beginning of the development of the model pre-
sented in this paper stood the observation that probability
bar charts generated either with the Burger model (Burger,
1985), or the Stochastic Insolation Model SIM (Morf,
1998) look practically the same. Thus, the SIM (now more
appropriately called the Stochastic Vertical Visibility
Model SVVM, see Section 3) was applied to the sky area
as a mathematical formulation for the description of cloud
cover, whereby simply time t was substituted by area A.
This model is now called the Stochastic Cloud Cover
Model SCCM.
4.1. The basic denition of the model
In accordance with the SVVM (Eq. (7)), the basic for-
mulation of the SCCM is given by
[P[
cc
=
P
00
P
01
P
10
P
11
_ _
=
1
DA
A
0
DA
A
0
DA
A
1
1
DA
A
1
_ _
; (8)
where
DA dA:
A is the sky area, A
0
is the mean size of a clear area, and A
1
is the mean size of a covered area.
The basic formulation of the SCCM is not easy to com-
prehend, because it describes a phenomenon that expands
over an area with a one-dimensional construct. In particu-
lar, the visualization of the area of clear sky A
0
and of
covered sky A
1
in the real world presents some diculties
for obvious reasons, as the further development of the
model will show. Nevertheless, applications of the model
match well with the real world.
Cloud cover cc is the equivalent to relative daily visi-
bility in the SVVM. It is given by
cc =
A
1obs
A
obs
; (9)
where A
obs
is the observed sky area, typically the sky area
within a radius of about 30 km around the observer.
A
1obs
is the covered part of the observed sky area.
The implementation of the model is shown in three
steps.
(1) A one-dimensional steady state model is developed
that simulates the steady state probability distribu-
tion of cloud cover in a given area.
(2) The one-dimensional steady state model (step (1)) is
expanded to simulate cloud cover in a given area as
a function of time.
(3) A method is shown to map the one-dimensional mod-
els onto the xy plane. This method can be applied to
both, the one-dimensional steady state model (step
(1)), and also to the area-time model (step (2)).
4.2. The one-dimensional steady state model
The cloud cover model forms a random sequence of
innumerable states, each one represented by a value of
cloud cover (Eq. (9)). This sequence is a stationary process
(Morf, 1998). In such a process, every single sample con-
tributes to its steady state probability distribution. Thus,
many values of cloud cover are generated and united into
one single distribution by application of the procedure that
follows. With increasing sample size, the resulting distribu-
tion converges towards the steady state probability distri-
bution of the cloud cover sequence.
4.2.1. The generation of cloud cover
Starting parameters for the generation of cloud cover
are mean l(cc) and variance r
2
(cc) of cloud cover
derived from a sample of cloud cover collected in the real
world.
In terms of cloud cover, Eq. (A16) transforms into
r
2
(cc) = 2 W
2
0
W
2
1

A
0
A
1
A
obs
1 W
0
W
1

A
0
A
1
A
obs
_
1 exp
A
obs
W
0
W
1
(A
0
A
1
)
_ __
; (10)
and Eq. (A15) into
l(cc) = cc =
A
1
A
0
A
1
= W
1
= 1 W
0
(11)
Thus, knowing mean and variance of cloud cover, A
0
and
A
1
can be calculated. In practice one numerically solves
Eq. (10) for (A
0
A
1
)=A
obs
.
5
Then, A
0
and A
1
can be calcu-
lated with help of Eq. (11).
6
Then, one divides the area under observation into N
equal parts and determines the state of the sky (covered
or clear) at the dividing points.
7
The state of these
dividing points will then be representative for cloud
cover in the observation area, if N is suciently large.
Knowing A
0
and A
1
, the following adaptation of Eq. (A3)
to the SCCM may be used for this exercise:
5
Eq. (10) is transcendental in (A
0
A
1
)=A
obs
.
6
For convenience, it is best to use the mostly not exactly known
observation area A
obs
as the unit of measure, and consequently set
Aobs = 1.
7
Dividing an area, the division is rather a line than a point (or, as we
shall see later a cluster of points).
4 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
This equation is an expression for the probability of the
transition from state i to state j (here, there are only two
states: clear (0) or covered (1)) after an area step of A
obs
/
N. No special requirements with respect to the intermedi-
ary states need to be met.
First, a random experiment is executed to determine the
state of the starting point with an expectation of W
1
(see
Eq. (11)) for the covered state. Then, one proceeds by boot-
strapping through the dividing points by subsequently
applying Eq. (12). Finally, one receives cloud cover as the
ratio between the number of covered dividing points
and the total number of dividing points.
4.2.2. Verication of the simulation results
Cloud cover bar charts as a function of (A
0
A
1
)=A
obs
and A
1
=(A
0
A
1
) generated with the model are shown in
Fig. 1. Visibly, they match well with the cloud cover bar
charts generated with the Burger model shown in Fig. 11
on page 23 in Burger (1985).
Still, another validation of the model with more tangible
results was executed by comparing its output with a test set
of nine probability functions of cloud cover presented in
Burger (1985), as shown in the following paragraphs.
Burger (1985) validated his model by comparing its out-
put with cloud cover data obtained from the Revised
Uniform Summary of Surface Weather Observations
RUSSWOcompiled by the USAFEnvironmental Technical
Applications Center. He obtained cloud cover climatologies
for 273 stations. For every station the monthly cloud cover
observations are grouped into eight 3-h periods identied
by the local time. As an example, he shows the data sheet
of a station in Savannah, Georgia, USAin January, covering
the time span from 1948 till 1970: for every 3-h period the
probability function of cloud cover is shown by numerically
stating the probabilities of the eleven cloud cover classes
010. From his vast data collection he selected a test set of
nine probability functions with dierent combinations of
mean and variance to validate his model.
From every probability function in his test set he rst
derived the parameters for his model. Then he used his
model with these parameters to calculate the probability
function again. Finally, he compared these two probability
functions by calculating the rms error between them. Table
6, page 20 in Burger (1985) contains the two probability
functions as well as the rms errors for his test set.
To validate our model, we repeated Burgers test with
our model. The model parameters derived from the proba-
bility function of the test set are mean and variance of
cloud cover. Table 1 shows the rms errors between the
probability functions simulated with our model and those
of the test set. Also shown are Burgers test results. Ours
compare favorably with those of Burger: The mean of
the rms errors of the probability functions in Burgers test
is 1.4%, whereas it is only 1.1% in our test.
4.3. The one-dimensional time model
Cloud cover does not build up instantaneously. On the
contrary, a change of state in an element dA is at the same
time also a change of state in a time element dt. It follows
then from the basic denition for vertical visibility
(Eq. (7)), and for cloud cover (Eq. (8)) that the following
holds true:
dt
t
0
=
dA
A
1
(13)
dt
t
1
=
dA
A
0
(14)
Therefore, the combined simulation over time and area
becomes mainly a scaling exercise with the objective to
assign time marks to the simulation steps executed in the
one-dimensional steady state model (Section 4.2).
4.3.1. The generation of cloud cover as a function of time
Calculations cycle continuously over the dividing
points. For scaling purposes, one may ask for the time
span T
obs
that is needed to cycle once over the observation
area A
obs
. Taking Eqs. (13) and (14) into account, one
obtains
T
obs
=
t
0
A
1
A
obs
=
t
1
A
0
A
obs
(15)
For the simulation over time one initially proceeds as
outlined for the steady state model in Section 4.2.1. How-
ever, the states at the dividing points are now stored in
a circular buer, that contains room for all the points over
the time period T
obs
. Calculations are performed continu-
ously. Reaching the end of the buer, storage resumes at
the beginning of the buer. Consequently, the values in
the buer are continuously overwritten in a circular man-
ner. Always, after the storage of a new value (which has
overwritten the oldest value in the buer) cloud cover is
calculated as the ratio of the number of covered dividing
points in the buer and the total number of dividing
points in the buer.
[P
ij
[
A
obs
N
_ _
=
P
00
P
01
P
10
P
11
_ _
=
W
0
W
1
exp
A
obs
N

A
0
A
1
A
0
A
1
_ _
W
1
W
1
exp
A
obs
N

A
0
A
1
A
0
A
1
_ _
W
0
W
0
exp
A
obs
N

A
0
A
1
A
0
A
1
_ _
W
1
W
0
exp
A
obs
N

A
0
A
1
A
0
A
1
_ _
_

_
_

_ (12)
H. Morf / Solar Energy xxx (2011) xxxxxx 5
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
4.3.2. Discussion of the model
The basic denition of the SCCM (Eq. (8)) describes a
restricted stationary process (Chung, 1978). Such a process
allows ample exibility when sampling. Also the overlap-
ping samples described in Section 4.3.1 will contribute to
the steady state probability distribution of cloud cover.
It follows from Eqs. (13) and (14) that the maximum
absolute value of the rate of change of cloud cover over
time is limited to A
1
=t
0
=A
obs
= A
0
=t
1
=A
obs
.
4.3.3. Verication of the model
Simulations of cloud cover as a function of time have
been executed for January and July, for Perth Australia.
Simulation parameters were derived from a ten year record
of cloud cover at Perth Airport, spanning the time period
from 1998 till 2007. Cloud cover is observed every 3 h at
Perth Airport. For the determination of the parameters
A
0
and A
1
of cloud cover we proceeded as outlined in Sec-
tion 4.2.1. The parameters t
0
and t
1
of vertical visibility
were found by following the procedure shown is Section
5.3. The proper simulation was run as outlined in Section
4.3.1.
Observed and simulated cloud cover sequences are
shown in Fig. 2. Also the simulated sequences obey the
meteorologists convention to assign a value of 0 or 8 oktas
only to a totally clear, respectively a fully covered sky. The
relevant simulation parameters are given in the legend of
the gure. One must not expect a perfect match of the
observed and the simulated sequences shown in Fig. 2,
because cloud over sequences of the length of only one
Fig. 1. Simulated probability bar charts of cloud cover and daily relative vertical visibility. Probability bar charts of cloud cover cc and daily relative
vertical visibility i look the same. For cloud cover the parameters are A
0
, the mean size of a clear area; A
1
, the mean size of a covered area; and A
obs
, the
observation area. For daily relative vertical visibility the parameters are t
0
, the mean duration of a covered period; t
1
, the mean duration of a clear period;
and T
d
, the duration of the daytime. The values assigned to the x-coordinate correspond to the mean cloud cover in the observation area l(A
1obs
/A
obs
), and
the mean daily relative vertical visibility l(T
1d
/T
d
). The bar charts for x-coordinate values >0.5 are the mirror image of those of the complementary value
of x.
6 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
month may vary considerably. All the same, there is a gen-
eral agreement between observed and simulated sequences,
whereby those of January are clearly distinguishable from
those of July. The absolute maximum rate of change
matches reasonably well, although the observed sequences
tend to exceed the maximum expected by the model. There
Table 1
Errors on probability functions of cloudcover. Probability functions of cloud cover measured in the real world are compared with probability functions
simulated with the STSCCM based on mean and variance measured in the real world. Column rms error STSCCM shows the rms errors between
measured and simulated probability functions. For comparison, column rms error Burger, shows the rms errors of a similar exercise using the Burger
model.
Location Time period Mean () Variance () rms error STSCCM () rms error Burger ()
Fresno, CA, USA July, 00:0002:00 0.0545 0.0312 0.0086 0.009
Minneapolis, MN, USA January, 12:0014:00 0.6311 0.1826 0.0110 0.015
Yakutat, AK, USA July, 06:0008:00 0.8869 0.0691 0.0069 0.014
Faireld, CA, USA July, 06:0008:00 0.1334 0.0679 0.0079 0.026
Honolulu, HI, USA July, 18:0020:00 0.4952 0.0897 0.0211 0.022
Dutch Harbor, AK, USA October, 12:0014:00 0.9215 0.0222 0.0182 0.017
Chunchon, Korea October, 00:0002:00 0.2951 0.1783 0.0098 0.011
Nome, AK, USA January, 00:0002:00 0.5279 0.2211 0.0067 0.006
Finter, Germany January, 00:0002:00 0.7511 0.1567 0.0120 0.009
Mean of rms errors 0.0114 0.014
Fig. 2. Cloud cover as a function of time. Parameters are typical for Perth, Australia, with readings taken every 3 h. The top pair of images shows an
observed and a simulated sequence for January, the bottom pair for July. Simulation parameters for January: t
0
= 9:3 h; t
1
= 25:6 h,
A
0
= 2:5 A
obs
; A
1
= 0:9 A
obs
. Simulation parameters for July: t
0
= 10:9 h; t
1
= 12:1 h, A
0
= 1:4 A
obs
; A
1
= 1:2 A
obs
.
H. Morf / Solar Energy xxx (2011) xxxxxx 7
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
is a sparse observation of cloud cover readings of 8 oktas.
We have inspected the probability functions of the observa-
tions; they show an expressive bulking at the reading of
7 oktas. Thus, we conclude that observers were too reluc-
tant in the assignment of an observation of 8 oktas.
4.4. The two-dimensional model
The leading idea for the two-dimensional model is to lay
a grid over the xy plane, and then determine the state of
the sky (covered or clear) at the grid points. With a su-
ciently small mesh width, these grid points will then be rep-
resentative for the state of the sky.
The determination of the state of the sky at the grid
points of the two-dimensional model is solved by mapping
the dividing points of the one-dimensional model (Sec-
tion 4.2) onto the xy plane. This is done by considering
each of the N dividing points of the one-dimensional
model as a cluster of grid points that are randomly dis-
persed over the xy plane, see Fig. 3. A single grid point
belongs then with a probability of 1/N to a specic cluster.
Thus, one expands the one-dimensional model to the xy
plane by applying the state of a dividing point of the
one-dimensional model to all single grid points belonging
to its cluster.
In practice one proceeds by rst dening the observation
area as a square grid with N grid points.
8
Then, cluster
numbers are randomly assigned to the grid points; think
of a die with N faces that is thrown every time a cluster
number has to be assigned. Next, the result of the one-
dimensional model (Section 4.2) is mapped onto the clus-
ters. Finally, one receives cloud cover as the ratio of the
number of covered grid points to the total number of grid
points in the observation area.
This mapping technique can be applied on both, the
one-dimensional steady state model (Section 4.2), and the
area-time model (Section 4.3), because the former builds
on the latter.
4.4.1. Discussion of the model
The simulated sky area has a statistically self-similar
structure (see Mandelbrot, 1977 on this subject). Simula-
tion experiments show that over a wide range, the steady
state probability distribution of cloud cover is invariant
to the size of the observation area, see Fig. 4. An observer
focussing into a cloud eld generated by the model
always using the same resolution (i.e. the same number of
grid points) for the observed area would not be able to
determine the real extension of the observed area.
It follows from the above, that also mean and variance
of cloud cover are invariant to the size of the observation
area. Taking this into account, Eqs. (10) and (11) reveal
that A
0
as well as A
1
must always be proportional to the
observation area. Therefore, scaling of the model is easy.
For two areas A
obs
and B
obs
one may state:
A
obs
B
obs
=
A
0
B
0
=
A
1
B
1
(16)
In practice, it is best to maintain the observation area as
the unit of measure. Then, A
0
and A
1
will maintain their
numerical values, independent of the area under
observation.
One may ask for the range of validity of this invariance.
There is a lower limit set by the lack of resolution. How-
ever, this is an articial limit one can always escape from
it by increasing resolution. Nevertheless, it seems only rea-
sonable to accept that the physical properties of clouds will
nally set a real lower limit. One might also ask over what
range of area Eq. (5) is valid. This remains to be investi-
gated, but it seems that the model is valid within the range
of an observers resolution when viewing the sky dome.
The model is strictly area oriented. The steady state
probability distribution of cloud cover is invariant to the
shape of the area observed.
We have observed the cloud pattern on the xy plane in
simulation runs. The animations show how clouds (covered
grid points) grow and vanish sporadically over time. It is a
regular pattern one would not spontaneously identify as
clouds. Being rather interested in cloud cover than in the
cloud pattern, we have paid no further attention to it.
N
A
obs
N
A
obs
Cluster 2
Cluster 3
Cluster 1
A
obs
Sky Area, A
Fig. 3. Clustering For the determination of cloud cover the sky area is
divided into N parts. The partitions are thought to be clusters. A grid
point belongs then with a probability of 1/N to a specic cluster. The
elements within a cluster are randomly distributed over the sky area;
stochastically they behave as one. This gure depicts clustering with N = 3
only.
8
The number of grid points in the observation area needs not to be N.
However, N is a convenient choice, because in this case the expected
number of elements of a specic cluster in the observation area is 1.
8 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
Cloud cover, because of the averaging eect of its extension
over the entire sky area, matches reality better.
4.4.2. Verication of the model
Probability distributions of cloud cover of the one-
dimensional and the two-dimensional model should be
equal. Thus, rst a comparison of these distributions was
made by comparing their pdfs. No statistically signicant
deviations could be detected.
According to the model, the probability distribution of
cloud cover is invariant to the observation area over a wide
range, see Fig. 4. The very high space correlation of sky
cover in the Lund data set supports this hypothesis
(Willand and Steeves, 1991).
Data made available by Henderson-Sellers and
McGue (1991) allowed a comparison that lead to numer-
ical results. They analyzed the scalability of the Burger
model with four data sets of cloud cover extracted from
photographs taken with sh eye lenses. Of the available
11,690 photographs they used 7046. (See Table 2 for the
number of images used per data set.) In every photograph
they divided the viewing area into 34 elements of equal size
arranged in three concentric sections. In every element
cloud cover was then estimated separately. For the analysis
the elements were grouped into three grids: full grid com-
prises the full sky area, small grid comprises the two inner
concentric sections and covers 18/34 of the sky area, and
centergrid comprises only the center area with 2/34 of the
sky area. Cloud cover was then totalized in each of the
three grids.
We have compared the variance of cloud cover in these
three grids, see Table 2. There is a slight growth of the var-
iance of cloud cover with decreasing area size. However,
the increase is small. Compare it with the growth of the
Fig. 4. About the invariance of cloud cover. Cloud cover bar charts for an observation area of n = 1, 2, up to 50,000 grid points. The probability
distribution of cloud cover within the observation area remains constant over a wide range of grid points. This invariance only breaks down when the
number of grid points approaches zero. The cloud eld shown contains 2000 clusters. Thus, it contains on average one grid point per cluster in an
observation area of 2000 grid points. Expected mean: 0.2, expected variance: 0.02355.
H. Morf / Solar Energy xxx (2011) xxxxxx 9
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
variance of the mean of a sample of a normal distribution
when decreasing the sample size. There, one would expect a
growth proportional to the inverse square root of the sam-
ple size reduction. We conclude that within reach of the
observable horizon normally of about 30 km the condi-
tions for the application of the model are met.
The model even suggests an explanation for this rise in
variance: lack of resolution. In brief, dividing the sky area
into 34 subelds and then estimating cloud cover in each of
the subelds in oktas, as described in Henderson-Sellers
and McGue (1991), corresponds to a division of the
sky area into 34 + 8 = 272 grid points; the center grid has
then only 16 grid points. Experiments with the two-
dimensional model (Fig. 4) showed for this resolution a
comparable increase of the variance in comparison to the
invariant case.
We have also veried the invariance of the model with
respect to the shape of the observed area in triangular as
well as square grid elds, using various shapes for the
observation area. No statistically signicant deviations
could be detected.
5. Overying the cloud eld
At the starting point of the development of the
STSCCM stood the observation that probability distribu-
tions of daily relative sunshine duration are similar to those
of cloud cover. This section closes the circle over the model
back to sunshine. It is an exploration of the model with no
comparisons with the real world. Nevertheless, it gives
valuable insights into the model. We will determine the
observed sequence of clear and covered patches when over-
ying the sky, and show how it is related to vertical visibil-
ity and sunshine.
A thought experiment will shed light on the problem at
hand. Think of a ying object that records the single point
view down to earth in straight nadir direction while overy-
ing the sky. Remaining in a geostationary position it will
record a trace that corresponds to vertical visibility. With
increasing velocity the record over a given distance will
converge to a snapshot of cloud cover.
5.1. The simulation experiment
We have conducted simulation experiments with the
STSCCM, ying over a two-dimensional cloud eld at con-
stant velocity in the time span between sunrise and sunset,
T
d
. Each experiment consisted of many simulation runs
with the same set of parameters. Hereby, we recorded a
sample of the relative duration of a clear view, T
1df
/T
d
.
Then, we extracted the characteristic parameters t
0f
and
t
1f
from the sample, and used them in the SVVM to calcu-
late another sample of T
1df
/T
d
. We generated probability
functions of the two samples by classifying them into ele-
ven classes delimited by 0.05, 0.15 . . . 0.85, and 0.95. The
rms error between these two probability functions always
matched with an rms error smaller than 1%.
The parameters when overying the sky, t
0f
and t
1f
, can
deviate signicantly from those of vertical visibility, t
0
and
t
1
. However, the expectation of the duration of a clear view
always remains the same. Fig. 5 shows t
0f
_
t
0
= t
1f
_
t
1
as
function of
t
0
T
d
_
A
1
A
obs
=
t
1
T
d
_
A
0
A
obs
. The graph is the t to many
simulation runs within the following range of A
0
and
A
1
: 1 6 (A
0
A
1
)
_
A
obs
6 4, and 0:2 6 A
1
(A
0
A
1
)
_
6
0:5; it ts simulation results with an error of less than 5%.
Table 2
On the invariance of cloud cover to the size of the observation area.
According to the STSCCM, probability distributions of cloud cover are
invariant to the size of the observed area over a wide range. This table
presents the variance of four data sets of cloud cover observed in the real
world for three sizes of observation area. Compare the dependence of
these variances on changing observation area with the change of variance
of the mean of a sample of a normal variable. There, one would expect a
growth proportional to the inverse of the square root of the sample size
reduction.
Center grid Small grid Full grid
Grid ratio 2/34 18/34 1
(Grid ratio)
0.5
4.12 1.37 1
Data set Number of observations Variance of cloud cover
JASIN 1939 0.0573 0.0444 0.0362
LUND 2319 0.2257 0.2063 0.1981
SHIPS 86 1772 0.2127 0.1862 0.1760
FIFE 1016 0.2097 0.1985 0.1826
Fig. 5. Parameter corrections for overying 1. Overying the observation
area leads to a time sequence of covered and clear periods similar to that
of the static single point view of vertical visibility. However, the time
parameters are dierent. This gure depicts the ratio t
0f
=t
0
= t
1f
=t
1
as a
function of
t
0
Td
_
A1
Aobs
=
t1
Td
_
A
0
Aobs
. It ts simulation results with an error of less
than 5%. t
0
and t
1
are the parameters for vertical visibility; t
0f
and t
1f
those for the time sequence, when the observation area is overown in the
time span between sunrise and sunset, T
d
.
10 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
For convenience, Fig. 6 shows t
0f
_
t
0
= t
1f
_
t
1
as function
of
t
0f
T
d
_
A
1
A
obs
=
t
1f
T
d
_
A
0
A
obs
. It is derived from Fig. 5 by multiplying
the free variable
t
0
T
d
_
A
1
A
obs
=
t
1
T
d
_
A
0
A
obs
with the dependent vari-
able t
0f
_
t
0
= t
1f
_
t
1
. It is useful to determine t
0
and t
1
when
t
0f
and t
1f
are known.
The prerequisite that the probability for zero vertical
visibility must always be equal to cloud cover (see Eq.
(1)) leads to another method for the generation of the
observed trace of clear and covered patches when
overying:
(1) Determine cloud cover, cc.
(2) Execute a random experiment that will generate zero
vertical visibility, vv = 0 with a probability equal to
cloud cover, cc.
Traces generated in this way are practically indistin-
guishable from those of the rst method; the rms error
between the probability functions is smaller than 1%. Thus,
one draws the following important conclusion:
The recorded trace of clear and covered patches when
overflying the sky is statistically independent of
the path followed: (17)
One also concludes from Eqs. (1) and (17) that the mean
of cloud cover over the time span T
d
always leads to an
expected value of T
0df
/T
d
that will contribute to the prob-
ability distribution of T
0df
/T
d
.
Therefore,
f
1
T
d
_
T
d
0
cc(t) dt
_ _
= f
T
0df
T
d
_ _
; (18)
where f stands for the probability density function.
5.2. Applying the model to the path of the sun
As an application of the second method consider the sun
moving over the sky. For simplicity, assume that clouds
form a horizontal plane with no vertical extension. As
the sun moves over the sky, the intercept of the line of sight
from an observer on the ground to the sun on this plane
will form a trace that evolves with varying speed. Despite
of this complex trace the second method can still be applied
for the line of sight. In reality, the vertical extent of the
clouds leads to distortions beyond the control of the model.
Still, the trace of a sunshine recorder ts the SVVM well.
See Morf (1998).
5.3. Extracting the time parameters of vertical visibility from
cloud cover data
Based on Eq. (18) one may apply the following proce-
dure for the determination of the two parameters t
0
and
t
1
of vertical visibility when a time sequence of cloud cover
is known:
(1) Divide the cloud cover sequence into equal parts.
9
(2) Calculate the mean of the cloud cover readings in
each part.
(3) Merge these means to the probability distribution of
T
0df
/T
d
and determine its expectation and variance.
(4) Calculate t
0f
and t
1f
from expectation and variance
by use of Eqs. (A15) and (A16).
10
(5) Use Fig. 6 to calculate t
0
and t
1
from t
0f
and t
1f
.
We have veried this procedure by rst generating a
time sequence of cloud cover values as outlined in Section
4.3.1, and then applying the procedure on this sequence.
We ran many experiments with dierent parameters. The
original sequence always was 300 days long and had a
recording interval of 3 h for cloud cover. Dividing the
sequence into parts of 24 h, the time parameters extracted
always matched the original parameters within an error
of 5%.
6. Conclusion
A connection between vertical visibility and cloud cover
has been demonstrated. Based on our analysis, we have
Fig. 6. Parameter corrections for overying 2. Overying the observation
area leads to a time sequence of covered and clear periods similar to that
of the static single point view of vertical visibility. However, the time
parameters are dierent. This gure depicts the ratio t
0f
=t
0
= t
1f
=t
1
as a
function of
t0f
Td
_
A1
Aobs
=
t1f
Td
_
A0
Aobs
. It ts simulation results with an error of less
than 5%. t
0
and t
1
are the parameters for vertical visibility; t
0f
and t
1f
those
for the time sequence, when the observation area is overown in the time
span between sunrise and sunset, T
d
.
9
Typically, cloud cover is observed every 3 h. Thus, one would divide
the cloud cover sequence into parts of 24 h, each containing eight
observations.
10
The procedure is outlined in Section 4.2.1 for the determination of A
0
and A
1
.
H. Morf / Solar Energy xxx (2011) xxxxxx 11
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
developed computer models that simulate cloud cover with
varying degree of complexity from the simulation of the
steady state probability distribution up to the generation
of cloud cover as a function of time.
The key to success came from three basic ideas:
(1) The Stochastic Sunshine Model SIM (Morf, 1998)
is also applicable for vertical visibility and cloud
cover.
(2) In the area of validity of the model, the probability
for zero vertical visibility must be equal to the mean
of cloud cover.
(3) Clustering to map the one-dimensional model onto
the xy plane.
It is an interesting nding of this work, that the steady
state probability distribution of cloud cover is invariant
to the size of the observed area over a wide range. Further
investigations of the limits of this invariance with respect to
resolution and area are necessary and being undertaken. So
far, it seems that the model is valid within the observation
range of about 30 km when viewing the sky.
Other mechanisms than those of the model seem to be
responsible for the formation of specic cloud types. Nev-
ertheless, cloud cover, because of the averaging eect of its
extension over the entire sky area, matches reality well.
Simulation models are often a tradeo of theoretical exact-
ness against shortcuts leading to more elegance, simplicity,
and velocity. It is the opinion of the author that a good
compromise has been found for the models of cloud cover
presented in this paper.
Acknowledgements
The author wishes to thank the Western Australian Cli-
mate Service Centre for the supply of a ten year data set of
Perth Airport, containing daily sunshine duration, daily
global irradiation, and total cloud cover readings.
Appendix A. The Stochastic Vertical Visibility Model
SVVM
This appendix summarizes the properties of the Stochas-
tic Vertical Visibility Model SVVM important in the con-
text of this paper. For a detailed understanding the reader
must refer to Morf (1998).
Morf (1998) introduced the Stochastic Insolation Model
SIM, a mathematical formulation for the alternating time
sequence of periods where the sun is shining and periods
where the sun is obstructed by clouds. For convenience, a
value of one (1) was assigned to the insolated instants
and a value of zero (0) to the clouded instants, and the
resulting digital pulse string was named the Stochastic
Insolation Function SIF. The same model now more
appropriately called the Stochastic Vertical Visibility
Model SVVM, and the corresponding Stochastic Vertical
Visibility Function SVVF, t even better as a model for
the description of the time sequence of clear sky and cov-
ered sky periods that comprise vertical visibility vv.
A.1. The basic formulation of the model
The sequence of clear and covered periods is generated
by a homogeneous recurrent Markov process with two
states operating on a time step Dt with a transition proba-
bility matrix
[P[
vv
=
P
00
P
01
P
10
P
11
_ _
=
1
Dt
t
0
Dt
t
0
Dt
t
1
1
Dt
t
1
_ _
(A1)
where
Dt dt (A2)
A.2. Transition probabilities
In a Markov process one understands as transition
probability P
ij
(s) the probability of a transition of state i
to state j in a time gap s. No special requirements with
respect to the intermediary states need to be met.
One deduces for the specic case given by Eqs. (A1) and
(A2):
[P
ij
[(s) =
P
00
P
01
P
10
P
11
_ _
=
W
0
W
1
exp s
t
0
t
1
t
0
t
1
_ _
W
1
W
1
exp s
t
0
t
1
t
0
t
1
_ _
W
0
W
0
exp s
t
0
t
1
t
0
t
1
_ _
W
1
W
0
exp s
t
0
t
1
t
0
t
1
_ _
_
_
_
_
(A3)
where W
0
is given by Eq. (A11), and W
1
by Eq. (A12).
A.2.1. Derivation
For a Markov process with discrete states the transition
probabilities are given by (Chung, 1978):
[P[(n) = [P[
n
(1) (A4)
As an instructive example the way followed to determine
P
11
(s) in Eq. (A3) is shown here:
P
11
(s Dt) = P
11
(s) 1
Dt
t
1
_ _
(1 P
11
(s))
Dt
t
0
(A5)
Regrouping the terms and making the transition to inni-
tesimal calculus, one obtains the dierential equation
dP
11
(s)
ds
P
11
(s)
t
0
t
1
t
0
t
1
=
1
t
0
(A6)
of which P
11
(s) can be calculated.
A.3. The probability density functions of the rst entry into a
state
One asks here for the probability of the rst entry from
state i into state j in a time gap s. For a Markov process
with only the two states 0 and 1 the result reduces to the
pdfs of the duration of state 0 and state 1.
12 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
The probability density function of duration of state 0:
f (t
0
) =
1
t
0
exp
t
0
t
0
_ _
(A7)
The probability density function of the duration of state 1:
f (t
1
) =
1
t
1
exp
t
1
t
1
_ _
(A8)
A.3.1. Derivation
In a Markov process with discrete states the probability
function of the rst entry into a state is given by (Chung,
1978):
P
ij
(n) =

n
v=1
F
ij
(v) P
jj
(n v) (A9)
Introducing t = n Dt, s = v Dt, F
ij
(s) = f
ij
(s) Dt, and
making the transition to innitesimal calculus, transforms
Eq. (A9) into
P
ij
(t) =
_
s
0
f
ij
(s) P
jj
(s t) ds (A10)
f
ij
is the pdf of the rst entry from state i into state j. It can
be calculated from Eq. (A10), for example, by way of a
Fourier transformation.
A.4. The steady state probability function
We are asking here for the probability W
i
of the process
to remain in each state i. The probability function formed
by all the probabilities W
i
is called the steady state proba-
bility function of the process. In our specic case it is easy
to perceive that the following holds true:
W
0
=
t
0
t
0
t
1
= 1 W
1
(A11)
W
1
=
t
1
t
0
t
1
= 1 W
0
(A12)
A.5. The autocorrelation of the SVVF
The autocorrelation of the SVVF is given
R(s) =
t
1
2
(t
0
t
1
)
2

t
0
t
1
(t
0
t
1
)
2
exp [s[
t
0
t
1
t
0
t
1
_ _
(A13)
A.5.1. Derivation
The autocorrelation of a stationary process with discrete
states is given by (Fischer, 1969):
R
x;x
(s) =

x
1

x
2
x
1
x
2
P(x
1
; x
2
; s) (A14)
where P(x
1
, x
2
, s) is the conjugate probability that the sto-
chastic variable x(t) will have the values x
1
and x
2
at two
instances dephased by s.
For our case of the SVVF one obtains R
x,x
(s) by multi-
plication of P
11
in Eq. (A3) by Eq. (A12). Observe, that
x(t) corresponds to the SVVF, hence there exist only two
values, 0 and 1, for x
1
and x
2
.
A.6. Mean daily relative vertical visibility
Mean daily relative vertical visibility is given by
l
T
1d
T
d
_ _
=
t
1
t
0
t
1
= W
1
= 1 W
0
(A15)
A.7. The variance of daily relative vertical visibility
The variance of daily relative vertical visibility is given
by
r
2
T
1d
T
d
_ _
= 2 W
2
0
W
2
1

t
0
t
1
T
d
1 W
0
W
1

t
0
t
1
T
d
_
1 exp
T
d
W
0
W
1
(t
0
t
1
)
_ __
(A16)
A.7.1. Derivation
Eq. A16 was derived from the generally valid expression
R
T
1d
T
d
; 0
_ _
= l
2
T
1d
T
d
_ _
r
2
T
1d
T
d
_ _
(A17)
R((T
1d
/T
d
), 0) is the rst term of the autocorrelation of the
daily relative vertical visibility sequence, Eq. (A21). l
2 T
1d
T
d
_ _
is the square of Eq. (A15).
A.8. The weight of the Dirac pulses of the pdfs of daily
relative vertical visibility
The weight of the Dirac pulse at T
1d
= 0 is
w(T
1d
= 0) =
t
0
t
0
t
1
exp(T
d
=t
0
) (A18)
The weight of the Dirac pulse at T
1d
= 1 is
w(T
1d
= T
d
) =
t
1
t
0
t
1
exp(T
d
=t
1
) (A19)
A.8.1. Derivation
An instructive example of how to derive w(T
1d
= 0) fol-
lows. One may interpret w(T
1d
= 0) as the probability that
a clouded period expands over the entire period T
d
. Thus,
one obtains with the introduction of Eqs. (A11) and (A7):
w(T
1d
= 0) = W
0
1
_
T
d
t=0
1
t
0
exp
t
t
0
_ _
dt
_ _
(A20)
A.9. The autocorrelation of the sequence of daily relative
vertical visibility
R
T
1d
T
d
; n
_ _
= W
2
1
2 W
2
0
W
2
1

t
0
t
1
T
d

_
1 W
0
W
1

t
0
t
1
T
d
1 exp
T
d
W
0
W
1
(t
0
t
1
)
_ __
exp
n T
D
W
0
W
1
(t
0
t
1
)
(A21)
H. Morf / Solar Energy xxx (2011) xxxxxx 13
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
T
D
represents the 24 h of the day. T
d
stands for the dura-
tion of the daytime.
A.9.1. Derivation
The autocorrelation of the time sequence of daily rela-
tive vertical visibility is given by
R(T
1d
; n) = 2
_
nT
D
T
d
s=nT
D
(nT
D
T
d
s) R(s) ds: (A22)
Eq. (A22) integrates all the possible local autocorrelations
within the time frame T
d
. This is a valid proceeding for a
stationary process. Introducing R(s) with Eq. (A14) and
integrating leads to the result.
Appendix B. The problematics of cloud cover and vertical
visibility measurements
For this paper cloud cover cc and vertical visibility vv
are rigorously dened. The two routinely collected meteo-
rological values that match these denitions best are cloud
cover estimated by human observers and sunshine derived
from the traces of sunshine recorders. However, they are
not the same. One must be aware of the many diculties
with the determination and interpretation of these values
in the real world.
B.1. Systematical errors
Records of sunshine and cloud cover match the condi-
tion that the probability for zero vertical visibility must
be equal to the mean of cloud cover (Eq. (5)) very poorly
(Hoyt, 1977; Biga and Rosa, 1980; Gueymard et al., 1995).
There are many reports onsystematical errors. Hoyt (1977)
reports an overestimate of cloud cover by ground-based
observers. He attributes the errors mostly to projection prob-
lems where sides of clouds are viewed and added to the total
estimate of cloud cover. Sunshine recorders sometimes detect
sunshine, while an observer would not. Based on satellite
measurements he concludes that sunshine recordings are bet-
ter estimates of cloud cover than those of a ground-based
observer. Moriarty (1991) comes to similar results.
A systematical error also stems from the sometimes lax
denition of sky cover (Moriarty, 1991). Some people inter-
pret both the fraction of the sky dome that is hidden by all
the visible clouds and the fraction of the surface region
directly above which there is a cloud as equivalent deni-
tions for cloud cover. We only accept the latter. The former
leads to higher cloud cover readings.
Vertical visibility is a single point measurement, whereas
sunshine is measured while the sun is moving over the sky.
Section 5 of this paper shows that the eect of overying
the sky cannot be disregarded.
B.2. Matching cloud cover and vertical visibility
This paper presumes that the probability for zero verti-
cal visibility must be equal to the mean of cloud cover (Eq.
(5)). Where necessary, measurements of cloud cover and
daily vertical visibility in the real world were adjusted to
meet this requirement by the proceedings outlined in the
following paragraphs.
When mainly interested in cloud cover, we accept the
measured pdf of cloud cover as correct. The values of
cc; A
0
, and A
1
are then derived directly from this pdf.
For vertical visibility we set vv = 1 cc. The mean dura-
tion of a covered period t
0
and of a clear period t
1
are then
determined by use of either the weight of the Dirac pulse at
T
1d
/T
d
= 0, or at T
1d
/T
d
= 1 by use of Eqs. (A18) or (A19).
The weight of the Dirac pulse at T
1d
/T
d
= 0 is the more
reliable value, because it does not include an estimate of
the duration of the daytime T
d
. Other promising methods
are to preserve the variance of vertical visibility or t
0
t
1
instead of recurring to a Dirac pulse.
Likewise, when the interest is principally directed
toward vertical visibility we accept the measured pdf of ver-
tical visibility as correct. The values of vv; t
0
, and t
1
are
then derived directly from this pdf. For cloud cover we
set cc = 1 vv and calculate A
0
, and A
1
from he weight
of the Dirac pulse at either A
1obs
/A
obs
= 0, or at A
1obs
/
A
obs
= 1. The weight of the Dirac pulse at A
1obs
/A
obs
= 0
seems to be the value that can be observed more reliably.
Other promising methods are to preserve the variance of
cloud cover or A
0
A
1
instead of recurring to a Dirac
pulse.
B.3. The classication of cloud cover into oktas (tenths)
A problem of particular interest is the classication of
the probability function of cloud cover into oktas (tenths)
(Jones, 1992; Henderson-Sellers and McGue, 1991).
v Meteorologists are instructed to report only a com-
pletely clear, or a fully covered sky as 0 oktas (0 tenths),
respective 8 oktas (10 tenths). They then argue that the
limits of the classication should be at 0, 1.5, 2.5 . . .
5.5, 6.5, 8 oktas (0, 1.5, 2.5 . . . 7.5, 8.5, 10 tenths) for
this kind of reporting.
v Mathematically inclined persons prefer a classication
with the limits at 0.5, 1.5 . . . 6.5, 7.5 oktas (0.5, 1.5 . . .
8.5, 9.5 tenths). This is the preferred division when pre-
senting the result as a pdf.
The consequence of these diering types of limit setting
is a bulking of the probability for the second and the sec-
ond last bin in the meteorologists view when compared
to the mathematically inclined view at the cost of a
decrease of the probability for the rst and the last bin.
Be aware, that even other limit settings are in use, or have
been in the past (Jones, 1992).
A much more visible dierence in the bar charts than
that imposed by diering classication limits is related to
what the data represent:
14 H. Morf / Solar Energy xxx (2011) xxxxxx
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015
v Many meteorologists present the data in form of bar
charts, where the length of the bars is proportional to
the probability of the bins.
v A more mathematical inclination would be to present
the data in form of a pdf, where the area of the bars
in the chart would correspond to the probability.
The dierence of appearance of these two ways of pre-
sentation can be dramatic. Keep in mind that the readings
at zero and full cloud cover are Dirac pulses of diminishing
width and height reaching innity.
References
Biga, A.J., Rosa, R., 1980. Estimating solar irradiation sums from
sunshine and cloudiness observations. Solar Energy 25/3, 265272.
Burger, C.F., 1985. World Atlas of Total Sky Cover. Environmental
Research Papers, No. 927, AFGL-TR-85-0198. Air Force Geophysics
Laboratory, Hanscom Airbase, MA. pp. 112.
Burger, C.F., Gringorten, I.I., 1984. Two-Dimensional Modeling of
Lineal and Areal Probabilities of Weather Conditions. Environmental
Research Papers, No. 875, AFGL-TR-84-0126. Air Force Geophysics
Laboratory, Hanscom Airbase, MA. pp. 58.
Chung, K.L., 1978. Elementary Probability Theory with Stochastic
Processes. Springer, Berlin, pp. 240304.
Craggs, C., Conway, E.M., Pearsall, N.M., 2000. Statistical investigation
of the optimal averaging time for solar irradiance on horizontal and
vertical surfaces in the UK. Solar Energy 68/2, 179187.
Fischer, F.A., 1969. Einfu hrung in die statistische U

bertragungstheorie.
Bibliographisches Institut, Mannheim/Zu rich.
Gall, J., Abel, D., Ahlbrink, N., Pitz-Paal, R., Anderson, J., Diehl, M.,
Teixeira Boura, M., Schmitz, M., Hoschmidt, B., 2010. Simulation
and control of thermal power plants. In: Proceedings of the Interna-
tional Conference on Renewable Energies and Power Quality
(ICREPQ10), Granada (Spain), March 2325, 2010, pp. 294298.
Gueymard, C., Jindra, P., Estrada-Cajigal, V., 1995. A critical look at
recent interpretations of the A

ngstro m approach and its future in

global solar radiation prediction. Solar Energy 54/5, 357363.
Henderson-Sellers, A., McGue, K., 1991. An investigation of the burger
distribution to characterize cloudiness. Journal of Climate 4,
11811209.
Hoyt, D.V., 1977. Percent of possible sunshine and the total cloud cover.
Monthly Weather Review 105, 648652.
Jones, P.A., 1992. Cloud-cover distributions and correlations. Journal of
Applied Meteorology 31, 732741.
Mandelbrot, B.B., 1977. The Fractal Geometry of Nature. Freeman, New
York, pp. 349350.
Morf, H., 1998. The stochastic two state irradiance model (STSIM). Solar
Energy 62/2, 101112.
Moriarty, W.W., 1991. Cloud cover as derived from surface observations,
sunshine duration, and satellite observations. Solar Energy 47/3,
219222.
Robinson, N. (Ed.), 1966. Solar Radiation. Elsevier, Amsterdam, pp.
111160 (Chapter 4 by W. Schu epp).
Suehrcke, H., McCormick, P.G., 1989. The distribution of average
instantaneous terrestrial solar radiation over the day. Solar Energy 42/
4, 303309.
Willand, J.H., Steeves, J., 1991. Sky-cover correlation within a sky dome.
Journal of Applied Meteorology 30, 10371039.
Glossary
pdf: probability density function (continuous distributions)
pf: probability function (discrete distributions)
rms error: root mean square error
RUSSWO: Revised Uniform Summary of Surface Weather Observations
compiled by the USAF Environmental Technical Applications Center
SCCF: Stochastic Cloud Cover Function
SCCM: Stochastic Cloud Cover Model
SIF: Stochastic Insolation Function
SIM: Stochastic Insolation Model
STSIM: Stochastic Two-State Solar Irradiance Model
STSCCM: Stochastic Two-State Cloud Cover Model. (The model de-
scribed in this paper.)
SVVF: Stochastic Vertical Visibility Function
SVVM: Stochastic Vertical Visibility Model
USAF: United States Air Force
H. Morf / Solar Energy xxx (2011) xxxxxx 15
Please cite this article in press as: Morf, H. The Stochastic Two-State Cloud Cover Model STSCCM. Sol. Energy (2011), doi:10.1016/
j.solener.2011.02.015