solar radiation modelling in the urban context
TRANSCRIPT
Solar Energy 77 (2004) 295–309
www.elsevier.com/locate/solener
Solar radiation modelling in the urban context
Darren Robinson *, Andrew Stone
BDSP Partnership Ltd, Summit House, 27 Sale Place, London W2 1YR, UK
Received 13 January 2004; received in revised form 6 May 2004; accepted 19 May 2004
Available online 17 June 2004
Communicated by: Associate Editor Pierre Ineichen
Abstract
This paper describes alternative methods for predicting surface irradiance in the urban context. In this the focus is on
means of accounting for the effects of nearby obstructions on reducing direct sky radiation and on contributing reflected
radiation. The first two methods involve abstracting the urban skyline into an effective canyon using isotropic and
anisotropic tilted surface irradiance models. The third predicts the irradiance contribution from two hemispheres which
are discretised into patches––given the radiance of the sky and dominant obstructions (if these exist) and associated
view factors––so that we have a new simplified radiosity algorithm (SRA). Results from the three methods (isotropic
canyon (IC), anisotropic canyon (AC) and simplified radiosity algorithm (SRA)) are compared with a ‘truth model’
under the following circumstances: (i) unobstructed sky, (ii) sky obstructed by black surfaces, (iii) sky obstructed by
grey diffusely reflecting surfaces. Results show conclusively that the SRA offers superior accuracy at comparable speed
to the canyon models. The SRA also compares well with a ray tracing program, it can handle urban scenes of arbitrary
geometric complexity and is readily amenable for inclusion into standard computer programs that require surface
irradiance as an input.
� 2004 Elsevier Ltd. All rights reserved.
Keywords: Urban; Solar radiation; Model; Theory; Validation
1. Introduction
To help to optimise the energy and environmental
performance of buildings, dynamic thermal simulation
programs are commonly deployed. These programs are
now relatively mature. Their core heat transfer mecha-
nisms have been extensively validated and they have a
growing user base in both academic and commercial
circles. Consequently, their capabilities have grown
considerably, in some cases to include simultaneous
simulation of plant systems, mass flow and electrical
power flow networks not to mention embedded com-
putational fluid dynamics (CFD) domains (Clarke,
* Corresponding author. Tel.: +44-207-298-8383; fax: +44-
207-298-8393.
E-mail address: [email protected] (D. Robinson).
0038-092X/$ - see front matter � 2004 Elsevier Ltd. All rights reserv
doi:10.1016/j.solener.2004.05.010
2001). However, these programs are presently only
weakly coupled to the urban context in which the vast
majority of the population in developed countries works
and resides. More specifically, (i) current dynamic
thermal simulation programs consider obstructions to
views of the sun but either ignore obstructions to the sky
vault or associated reflections from them, (ii) no clear
basis for translating synoptic meteorological measure-
ments to the local urban context has yet been developed
(e.g. local and background (heat island) influences on air
temperature), (iii) buildings are assumed to operate
independently, so that there is no basis for handling
centralised resource management (e.g. centralised energy
centres). There are also issues that fall beyond the scope
of conventional dynamic thermal simulation programs,
such as centralised water and waste management.
These are some of the challenges that are currently
being addressed by Project SUNtool––an EC funded
ed.
296 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
research project to develop a new integrated resource
flow modelling and associated educational tool to sup-
port the design of sustainable urban neighbourhoods
(Robinson et al., 2003). This paper is concerned with the
first of the urban coupling issues identified above: the
prediction of solar radiation in a way that is sensitive to
obstructions both to the sun and to the sky vault, with
due regard for reflection.
The motivation for this work is to improve the res-
olution of solar radiation predictions in urban contexts.
Consider a room at the base of a tall street canyon of
constant height H (say H=W ¼ 2:5). For a point in the
centre of a window plane that is say 2 m above the
ground and for a street width W of 10 m, we have an
equivalent urban horizon angle (UHA) (Baker and
Steemers, 1994) of �66�. In other words over two thirds
of the sky vault is obscured. Now if our street is in
London, UK (in which case our window is never directly
insolated) and we assume isotropic skies then our win-
dow receives less than 1/10th the direct irradiation that a
standard thermal simulation program (with an isotropic
sky model) would predict (Fig. 1)! This has clear
implications not only for energy consumption but also
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 10 20 30 40 50 60 70 80 90
Urban horizon angle, u (deg)
Idβ
/ id
h
β=0˚β=30˚
β=60˚
β=90˚
β=120˚β=150˚
Fig. 1. Ratio of incident and horizontal diffuse irradiance as a
function of urban horizon angle for surfaces at 30� increments
of slope.
for HVAC system (over)sizing and for identifying whe-
ther passive cooling options are feasible.
2. Alternative model formulations––theory
This section describes three methods to account for
sky obstructions. The first two are based on abstracting
real urban skylines into an effective canyon which re-
duces sky views but also contributes reflected radiation
to our receiving surface. This approach is used in con-
junction with both isotropic and anisotropic tilted sur-
face irradiance models. The third approach is more
physically rigorous. Using a continuous absolute radi-
ance distribution the irradiance incident on our receiving
plane is a function of its view to discrete regions of the
sky, so that the geometric and photometric integrity of
the problem is retained.
2.1. Isotropic canyon model
Some dynamic thermal simulation programs assume
that the sky is isotropic (the vault has the same radiance
in all directions). Under this assumption and given the
diffuse horizontal irradiance Idh, the diffuse irradiance
incident on an unobstructed plane of slope b (06 b6 p,where 0 is horizontal facing up, p=2 is vertical and p is
horizontal facing down) from the sky is:
Idb ¼ Idhð1þ cos bÞ=2 ð1aÞ
Furthermore, ignoring any obscuration of the sky vault
by our receiving plane and given the global horizontal
irradiance (Igh), the corresponding radiation received
from the ground (IqG;b) of reflectance qg is given by (1b):
IqG;b ¼ Ighqgð1� cosbÞ=2 ð1bÞ
Now, if as in the previous example (Section 1) we
combine our solar obstructions into an equivalent con-
tinuous skyline of angular height u (from the perpen-
dicular) to our point of reference, then (1a) can be
modified to give the corresponding obstruction-sensitive
diffuse irradiance due to the sky:
Idb ¼ Idh½1þ cosðb þ uÞ�=2 ð2Þ
This expression introduces a subtle complexity however,
since we need to compose a description of some paral-
lelogram which has the same characteristics, with due
regard for the angle of incidence of reflected (and indeed
obscured sky) radiation as the real, inevitably more
complex, skyline (Appendix A).
Notwithstanding the complexity of deriving u, a
convenient result of (2) is that expressed graphically
(Fig. 1) one can determine a priori the extent to which
diffuse surface irradiance (or indeed irradiation) due to
Fig. 2. Street canyon geometry, as viewed from a receiving
surface.
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 297
the sky is reduced by urban obstructions––albeit under
the assumption of sky isotropy.
Another source of complexity is the treatment of
reflected radiation in a way in which the energy received
by each reflecting surface is sensitive to other obstruc-
tions to the sky vault (Fig. 2). For example, the ground
plane’s view of the sky is obstructed by the urban skyline
that opposes our receiving plane as well as by the urban
skyline of which our receiving plane is part (views to this
ground plane are also now reduced, in proportion to the
view factor to the below horizontal vertical surface).
Each urban skyline (or wall of our equivalent street
canyon) is similarly obstructed by the opposing wall.
Given some global irradiance Igb incident on an up-
per obstructing surface of diffuse reflectance q the cor-
responding energy received by our surface of interest
(IqU;b) is given by (3a): 1
IqU;b ¼ Igbq½cos b � cosðb þ uÞ�=2 ð3aÞ
and that received due to reflection from the opposing
wall below the horizontal plane (IqL;b) is:
IqL;b ¼ Igbq½cosðb � uÞ � cos b�=2 ð3bÞ
for which tan u is simply the quotient of the mean
height of our receiving plane and the canyon width
ðz=wÞ. All that remains then is to determine the radiation
reflected onto our receiving plane from the ground
(IqG;b).
1 Subscripts q and U refer to ‘reflected’ and ‘from upper
surface’ respectively. An L would refer to lower reflecting
surfaces.
IqG;b ¼ Igbq½1� cosðb � uÞ�=2 ð4Þ
Now, in deriving the energy incident on surfaces 1 and 2
identified in Fig. 3, it is first necessary to solve for that
received directly from the sky (using (2)) and the sun.
For this latter case, the standard approach used in dy-
namic thermal simulation programs (e.g. ESP-r (Clarke,
2001)) is to discretise the surface of interest into a reg-
ular grid of cells. At hourly intervals say for the central
day of each month, a test is then performed at each cell
centroid to determine whether the solar disc can be
‘seen’. The ratio of the number of successful hits to the
number of cells is an approximation of the proportion of
the surface that is directly insolated (r). Using a pre-
processed look up table to extract the appropriate r for
time t, the surface irradiance is simply: Ibb ¼ Ibnrt cos n,for direct normal irradiance Ibn and angle of incidence n.
The energy incident on the ground is a special case, in
that knowledge of both sides of the canyon is required,
and is given by (5) below.
Id;gnd ¼ Idhðcos d1 þ cos d2Þ=2þ q1Igb;1ð1� cos d1Þ=2þ q2Igb;2ð1� cos d2Þ=2 ð5Þ
With these contributions known, reflected radiation may
be modelled by iteratively solving for expressions (3)–
(5).
Finally, there are some geometric exceptions to (2)–
(5) which limit their range of applicability. In particular:
• The maximum surface tilt is limited by the urban
horizon angle: b þ u� p26
p2(corresponding to situa-
tions where the surface has tilted beyond the vertical
and is facing towards the ground such that the top of
the equivalent obstruction is not visible).
• Care must be taken with surfaces that are tilted by
less than p2as it is assumed that the surface can see
the sky over the entire altitude range from the calcu-
lated urban horizon to p2past the surface normal (Fig.
3). This limitation could be overcome by assigning
two UHAs to each surface.
2.2. Anisotropic canyon model
Several dynamic simulation programs now use
anisotropic sky radiation models, including those of (i)
Temps and Coulson (1977) and Klucher (1979), (ii) Hay
(1979), (iii) Muneer (1997), and (iv) Perez et al. (1990).
Of these the Perez model has been found to give the best
overall performance when applied to a wide variety of
locations (Jensen, 1994);
Idb ¼ Idh½ð1� F1Þð1þ cos bÞ=2þ F1a0=a1 þ F2 sin b� ð6Þ
in which the three terms within the square brackets cor-
respond to isotropic background scattering, circumsolar
Fig. 3. Geometric limitations to applicability of the isotropic model.
298 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
brightening and horizon brightening contributions
respectively.
One approach to producing predictions of diffuse
irradiance due to the sky which account for urban
obstructions is to adjust each term in brackets in (6), so
that (i) the first term is scaled in proportion to the angle
of elevation of the urban skyline, (ii) the second term is
multiplied by a Boolean operator B 2 to account for
whether the solar disc, expressed as a point source, can be
seen, and (iii) the final term is scaled by S according to the
proportion of the horizon band, of angular height b (say
6.5� as in the original Perez model), that is obstructed
(S ¼ max½0; ðb� uÞ=b�, where u ¼ 2p
R aw�p=2aw
u cos a da).Combining these influences, we have an obstruction-
sensitive implementation of the Perez anisotropic model
as given by (7).
Idb ¼ Idh½ð1� F1Þ ½1þ cosðb þ uÞ�=2þ B F1a0=a1 þ S F2 sinb� ð7Þ
There is a slight complication to this method due to the
fact that the horizon brightening coefficient F2 may be
negative. This is permitted because although it is not
physically realistic to consider a negative energy con-
tribution received from the horizon band, it is equivalent
to having a brighter region towards the zenith––as for
an overcast sky (Perez et al., 1987). When F2 is negativeit would not be correct to use the scaling factor defined
above as increasing u would lead to a decrease in the
magnitude of F2 and hence a reduction in the brightness
of the zenith region. Instead the CIE overcast sky can be
used to derive a different scaling based on the sky
brightness integrated over the visible range:
S ¼R p�bu ð1þ 2 sin cÞdc
p � b � u p � bR p�b
0ð1þ 2 sin cÞdc
¼ ðp � b � uÞ þ 2ðcos b þ cos uÞðp � bÞ þ 2ðcos b þ 1Þ p � b
p � b � uð8Þ
2 Or indeed some fraction, if we use a similar means for
determining views to the sun as for the direct calculation.
Note that this is only valid for overcast skies; although
the situation where F2 is negative and the sky is relatively
clear should never occur.
Contributions from reflecting surfaces may be solved
using the procedure described above, for the isotropic
canyon model ((3)–(5)). However, in using (5) it should
be noted that the contribution from horizon and cir-
cumsolar sources that are visible to the ground is ig-
nored; though this simplification is not likely to lead to
significant errors.
2.3. Simplified radiosity algorithm
Whilst the Perez anisotropic model is an undoubted
improvement over the isotropic case, it is nevertheless a
simplification of reality. Since the background contri-
bution is assumed to be isotropic and only small urban
obstructions would fully obscure the horizon band, the
sky’s anisotropy is concentrated into the circumsolar
point source, which can either be seen or not. Further-
more, whilst we may achieve a similar obstruction solid
angle in our mapping from real to continuous skylines,
the part of the sky vault that is being blocked out by a
particular obstruction may have quite a different radi-
ance to that in the isotropic case. The following is an
attempt to surmount these drawbacks.
Several models have been developed in the daylight-
ing field in recent years to represent the luminance of the
sky vault as a continuous function of the angular loca-
tion of a sky point relative to some reference, typically
the zenith or the sun. Also in the daylighting field, there
has been interest in predicting illuminances due to views
of discrete regions of the sky vault––as is typically the
case within rooms––and this has led to the development
of a now standard technique for sub-dividing the sky
vault into a series of patches. It is logical to utilise this
approach to predict the irradiance as opposed to illu-
minance at a plane. So, if we have some set of p sky
patches, each of which subtends a solid angle U (Sr) and
has radiance R (Wm�2 Sr�1) then, given the mean angle
of incidence n (rad) between the patch and our plane
together with the proportion of the patch that can be
Fig. 4. Radiance distribution for London, UK (using the Kew 1967 climate dataset in conjunction with the Perez all weather lumi-
nance distribution model).
3 Note that c, a should relate to the visible proportion of
partially obscured patches.
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 299
seen r (06 r6 1), we have the following general solution
for direct sky irradiance (Wm�2) (9):
Idb ¼Xp
i¼1
RiUiri cos ni ð9Þ
The patch view factor r in (9) may account for self-
obstructions as well as adjacent buildings, so that we
have an obstruction-sensitive method for deriving diffuse
irradiance due to the sky.
Now, means for predicting the distribution of energy
within the sky vault date back to the original CIE
overcast sky model (Moon and Spencer, 1942) and early
formulations of the CIE clear sky model (Kittler, 1967)
through more recent clear sky models (CIE, 1973, 1996)
to more general sky definitions (Perez et al., 1993; CIE,
2002). Of the latter models, which cover the range of sky
types from overcast through intermediate to clear, the
Perez model is more amenable to incorporation within a
computer program since there is a straightforward basis
for selecting different sky types and predicting the cor-
responding energy distribution. This is presently not the
case with the CIE general sky formulation.
In the case of the Perez model the luminance of a
given sky point is expressed relative to some reference
point thus: ‘v ¼ f ðZ; hÞ, where Z is the zenith angle of the
considered sky point and h the angle between this point
and the sun. Conveniently, by integrating this relative
luminance throughout the sky vault we have an equiva-
lent diffuse horizontal illuminance E0 ¼R 2p0
R p=20
f ðZ; hÞ�cos c sin c dadc (for azimuth a and altitude c). The
quotient of our diffuse horizontal irradiance Idh and this
equivalent illuminance E0 defines a hemispherical nor-
malisation factor v which, when multiplied by a sky
point’s relative luminance, gives the corresponding
absolute radiance, so that the radiance of some point
within the ith patch is
Ri ¼ f ðZ; hÞiv ð10Þ
Turning now to sky vault discretisation, the technique
proposed by Tregenza and Sharples (1993) has found
widespread favour. With this method the sky vault is
split into seven azimuthal strips in which the azimuthal
range Da of the composite patches tends to increase
towards the zenith (12�, 12�, 15�, 15�, 20�, 30�, 60�), atwhich there is a single patch so that there are 145 in
total. The intention is that the patches subtend similar
solid angles, these being determined by the azimuthal
range (rad) of the ith patch and the corresponding
maximum and minimum heights of elevation, using an
expression of the form:
Ui ¼ Daiðsin ci;max � sin ci;minÞ ð11Þ
With a method of discretising the sky vault into a series
of patches of known solid angle, we can calculate an
approximate hemispherical radiance normalisation fac-
tor numerically, based on the radiance at the centroid of
each patch: v ¼ Idh=P145
i¼1 ‘viUi sin ci (Fig. 4).With the view factor known (0, 1 or some fraction for
self-obstructed patches) all that remains is to calculate
the angle of incidence from the patch centroid to our
plane of interest n:
cos n ¼ cos ci cos a0i sinb þ sin ci cosb ð12Þ
in which ci is the altitude of the patch centroid, b is the
plane tilt and a0i is the azimuth angle from the patch
centroid to the plane surface normal (a0i ¼ ai � aw).
3
One of the key attractions of this technique for pre-
dicting diffuse sky radiation is that the radiance distri-
bution of the sky vault can be pre-processed and indeed
archived in an 8760 by 145 length file, in the case of an
hourly annual simulation. Similarly, sky patch view
factors for each of j surfaces may be pre-processed––so
that one can envisage a geometric pre-processing task
taking place within some graphical user interface once
the built vicinity has been described, and subsequent
Fig. 5. Below horizontal reflections from an inverted sky dome (left) and isotropic ground plane (right).
4 Some radiosity programs merge adjacent emitting cells into
a large cell of equivalent radiance to reduce the number of
computations in predicting the energy incident on each (and
every) receiving cell for each iteration. Our approach is similar,
we just combine angularly rather than physically adjacent
(sub)surfaces.
300 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
computation may be conducted by a separate processor
which reads the radiance distribution into a two-
dimensional array and refers to this when reading in the
view factors for each j surface at the kth hour.
For the purposes of testing our model, we have
developed a simplified algorithm that discretises each
patch using a regular 10 by 10 grid. A test is then per-
formed to determine the n number of cell centroids that
can be seen by our surface (i.e. whether centroids are
beyond the limits of obstructing polygons) and a view
factor-incidence angle product is derived:
ri cos ni ¼ 10�2Xn
j¼1
cos nj ð13Þ
Now to solve for radiation from surfaces that obscure
views to the vault IqU;b we have that:
IqU;b ¼X145i¼1
ðIgbq=pÞiUið1� riÞ cos ni ð14Þ
in which the final two terms may be solved using a
corollary of (13), based on the 102 � n centroids that
lie within the bounds of obstructing polygons
(ð1� riÞ cos ni ¼ 10�2P102�n
k¼1 cos nk). All that remains
then is to determine the irradiance Igb of the obstructing
surface(s) of known reflectance q in the ith patch. For
this we make an assumption that the radiance of the
largest surface obstructing the patch is similar to that of
the other surfaces obstructing the patch. Using our
simple patch discretisation algorithm, we identify this
surface as being that which produces the largest view
factor-incidence angle. This too can be determined as a
pre-process, so that for each patch with a non-unity view
factor, a surface identifier is parsed to the solver. This
then implies that the irradiance incident on each
obstructing surface is calculated, but since much of the
computation is dealt with as a pre-process the compu-
tational overheads of this approach are reasonably
modest. In the first instance then we calculate, for all
obstructing surfaces, the beam irradiance using the
method described in Section 2.1 and the diffuse irradi-
ance due to the sky using (9). This direct (sun+ sky)
irradiance is then used for the first pass of the reflection
calculation, so that (14) in the first instance represents a
single bounce of diffusely reflected radiation (assuming
all surfaces to be Lambertian). This process may be re-
peated, but now using a contribution from (14), n times
for all obstructing surfaces until we reach some con-
verged solution. If the radiance of each surface is held in
memory and we have many surfaces to solve for (e.g. in
the case of multiple buildings as in SUNtool), then the
apparent overhead of the iterative reflection algorithm is
generously shared so that the resultant run time is rea-
sonably modest.
To complete the model, we must consider reflections
from below the horizontal plane (IqL;b). For this we can
follow a similar principal to that established for above
horizontal reflections. We do this using an upside down
vault (Fig. 5) for which expression (14) simplifies to:
IqL;b ¼X145i¼1
ðIgbq=pÞiUi cos ni ð15Þ
and the reflection algorithm described in the previous
paragraph for 1; 2; . . . ; n bounces applies. In fact, for
computational efficiency the upper and lower reflections
may be carried out as part of the same routine. Alter-
natively, for large numbers of surfaces it may be more
efficient to solve for reflection in matrix form (Appendix
B). In either case, by combining obstructions into
aggregated patch occlusions whose radiance is given by
the dominant surface and solving for reflections between
these patch occlusions––(9)–(15) collectively describe a
simplified radiosity algorithm (SRA). 4
Fig. 6. Diffuse irradiation surface plots for isotropic (upper left) and anisotropic (upper right) canyon models and corresponding
difference image (bottom).
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 301
Alternatively, on the basis that model errors are
likely to be relatively weakly sensitive to errors in
accounting for the irradiance contribution from surfaces
below the horizontal plane, this contribution may be
represented by an equivalent isotropic ground plane
whose reflectance is a function of the below horizontal
obstructing surfaces. In order to calculate the irradiance
reflected by this ground plane, some means of deter-
mining the irradiance incident on it is required. This may
be achieved simply by passing a UHA for each surface,
which allows a sky view factor for the ground plane to
be determined (as with the isotropic model). The sim-
plification of an isotropic sky can then be used to
determine the ground plane irradiance. The error
introduced by this assumption will most likely be greater
than for the anisotropic canyon model, since the ground
reflected component will generally be greater in this case
(as the ground is consistently adjacent to the surface).
5 For all that follows, results have been produced for climate
data relating to Kew, UK for 1967.
3. Comparisons I––unobstructed sky
To verify the implementation of the sky models and
associated solar geometry results from ESP-r (Clarke,
2001) for the isotropic and anisotropic canyon models
and RADIANCE (Ward Larsen and Shakespeare,
1997)––with the additional module Gendaylit––for the
SRA were compared, using equivalent inputs. Having
achieved similar results, it is interesting to compare their
relative performance. 5 A useful approach to this is to
generate and compare irradiation surface plots (after
Robinson, 2003), based on solving for annual irradia-
tion (in this case diffuse only) incident on a plane for 5�bins of surface azimuth and altitude (entailing a total of
1387 annual calculation sets).
Taking first of all the isotropic model we see (Fig. 6)
the expected iso-contour banding, as the irradiation re-
sponds only to plane altitude. The anisotropic model on
the other hand depicts a concentration of energy about
the southern orientation. Perhaps most striking is the
irradiation difference plot which exemplifies the value in
considering sky anisotropy (in thermal simulation pro-
grams for example). Here we see that the isotropic
model considerably under-predicts the incident irradia-
tion on south facing tilted surfaces (by as much as 15%)
and over-predicts that received by north facing tilted
surfaces (by as much as 23%).
The difference between results from the (radiance
distribution based) SRA and IC models is further
exaggerated (Fig. 7), as there are now additional sources
Fig. 7. SRA irradiation surface plot (upper) and irradiation difference plots based on the isotropic (lower left) and anisotropic (lower
right) canyon models.
Table 1
Run time to produce an irradiation surface plot for an unob-
structed view of the sky vault together with associated RMS
error compared with RADIANCE
Model RMS error
(Wm�2)
Run time (s)
Isotropic canyon 32.4 7
Anisotropic canyon 11.8 34
SRA 0.6 2.11· 103SRA (archived radiance
distribution)
0.6 22
RADIANCE 0.0 4.99· 106
302 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
of anisotropy––within the treatment of the horizon band
and background scattering––so that south and north
facing surface irradiation are under-predicted and over-
predicted by as much as 26% and 18% respectively. The
contribution of this more rigorous treatment of sky
anisotropy can be appreciated by reference to the
(SRA � IC) difference plot shown lower right in Fig. 7,
as south facing vertical surfaces receiving correspond-
ingly more irradiation (ca. +5%) than those facing north
(ca. �4%), when compared to the anisotropic tilted
surface model.
A further basis for comparison that is interesting is
RMS error and how this relates to run time (Table 1), to
aid with selection of a model of greatest efficacy. The
first three rows show the expected reduction in RMS
error (as compared with RADIANCE 6) with improved
treatment of sky anisotropy, but at increasing compu-
tational cost. However, as noted earlier, it is sensible to
pre-process this radiance distribution (as well as patch
view factors) and to refer to these values, held in mem-
6 This table, perhaps unfairly, assumes that RADIANCE
should be our truth. For this unobstructed case this is not so, as
we have a direct analytical solution for the problem (ignoring
minor discretisation errors) and RADIANCE’s predictions
contain numerical errors––though these may be small, depend-
ing upon the settings used.
ory, at each time step. Consequently the run time for the
SRA improves to below that of the anisotropic tilted
surface model with an accuracy that is comparable to
RADIANCE.
4. Comparisons II––obstructed sky
The process of testing the incorporation of obstruc-
tion sensitivity into our diffuse radiation models has
been carried out in two phases. In the first instance we
have constructed representations of obstructions that
are black (non-reflective)––so that we test only the cor-
responding reduction to sky radiation. Following from
this we model grey obstructions, which are highly
Fig. 8. Obscured sky patches (left) from which a corresponding plane (centre) and an equivalent plane of constant height (right) was
produced.
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 303
reflective to exaggerate their influence on results. As in
Section 3 we have used RADIANCE as our truth model.
4.1. Black obstructions
For convenience our obstructing surface is simply a
cut out of the Tregenza sky (Fig. 8––left), so that whole
patches are obscured, for use within the SRA. These are
selected to test means for obscuring circumsolar and
horizon brightening. In the first instance the angular
coordinates of each patch vertex are translated to planar
coordinates relating to a surface at a distance of 20 m
from our viewpoint, for use within RADIANCE (see
Fig. 8––centre 7). From this we use the procedure de-
scribed in Appendix A to derive a UHA of an equivalent
continuous horizon, from which a plane of constant
height is found (Fig. 8––right), for use within the iso-
tropic and anisotropic canyon models.
In the first instance a test was performed to ensure
that the process for determining the UHA and from this
the constant height plane was correct. For this the
varying height plane was modelled using RADIANCE
and the result (Idb) used with a re-arrangement of (2) to
give u. The numerically derived value for u (17.04�) wasfound to be within 0.3� of that predicted by RADI-
ANCE (16.75�)––well within the error tolerance dis-
cussed in Appendix A––so that we can be reasonably
confident in the validity of the comparisons of models.
Two means of accounting for visibility of the cir-
cumsolar point source in (7) have been tested. In the
first instance this is based simply on whether the
solar altitude (cs) exceeds the altitude of the constant
height obstruction at the solar azimuth (as) (i.e.
cs > u cosðaw � asÞ). In the second case we use a look up
7 So that for a perpendicular distance d to our plane a vertex
(of angular height c and azimuthal offset Da) height h in y is
given by h ¼ d tan c= cosDa and the offset distance d in x by
d ¼ d tanDa.
table of pre-processed solar views based on the real
geometry. 8
The IC model tends to under-predict the surface
irradiance (Fig. 9) as the models were run for a south
facing surface; shown on the probability plot (Fig. 10) as
the relatively large area to the left of the origin. It should
be noted that relative error for the isotropic model peaks
at around zero––due to occasions when the circumsolar
region (in RADIANCE’s representation of the Perez all
weather luminance distribution model) is completely
obscured by the obstruction. In that situation the sky
radiance distribution will be close to isotropic (as most
of the horizon band is also blocked by the obstruction).
The corresponding RMS error is 23.45%.
The AC model with simple circumsolar visibility
tends to over-predict the surface irradiance as the south
facing surface ‘sees’ the sun whenever it is at an altitude
greater than u cosDa, whereas with the real obstruction
the circumsolar region is not visible around midday,
when the circumsolar region tends to have the largest
influence. The RMS error in this case is 15.6%. This is
reduced to 9.43% with the improved treatment of cir-
cumsolar views.
Excellent agreement is seen between the results from
the ray tracer and the SRA, with an RMS error of 0.3%.
4.2. Grey obstructions
The inclusion of reflected radiation has been tested
incrementally; in the first instance considering diffuse
radiation only and in the latter for global radiation, in
which case views to the sun from each surface are solved
for. The corresponding results are described below.
8 Note that the circumsolar visibility is determined using a
sun view factor for a whole surface rather than a point. This is
because the view factor for a point would be a Boolean value
(as the sun is assumed to be a point source), which leads to large
inaccuracies when the data is interpolated.
Fig. 9. Comparison of IC (upper left) and AC models without (upper right) and with (lower left) improved circumsolar views and SRA
(lower right) with RADIANCE predictions, using the Perez all weather luminance distribution model.
Fig. 10. Model error probability distributions (relative to RADIANCE) for a black obstructing surface.
304 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
4.2.1. Diffuse radiation
In order to test the validity of the way in which
reflecting surface radiances were derived in the canyon-
based models, some initial comparisons with RADI-
ANCE using an isotropic sky model were carried out,
isolating the ground and opposing surface contributions
respectively (by adjusting their reflectances). For these
tests, the surface was split into five horizontal slices (Fig.
Fig. 11. Geometry of the test case illustrating surface discretisation (left) and ground plane flanges (right).
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 305
11). In the first instance, the IC model was found to
slightly under-predict (by 3.4%), as brightening of the
floor towards the ends of the canyon was ignored (recall
that for the IC model the surface is abstracted to a
continuous obstruction). In the latter case, the IC model
somewhat over-predicted (by 1.8%) as the dominant
reflecting surface for below horizontal obstructions
tends to be that which is perpendicular (owing to the
angle of incidence of lower surfaces), so that the dark-
Fig. 12. Grey obstruction comparison of IC (upper left), AC with
ground plane (lower left) and SRA with inverted sky dome (lower ri
weather luminance distribution model.
ening of surfaces close to the floor tends to be ignored.
Note that the above horizontal surfaces tend to have
relatively un-obscured views of the sky. These errors,
which tend towards cancelling, were assumed to be
reasonable, so that comparisons progressed to determine
the validity of the models themselves. As before, model
predictions are compared with RADIANCE using the
Perez all weather luminance distribution model (Fig.
12).
improved circumsolar views (upper right), SRA with isotropic
ght) models with RADIANCE predictions, using the Perez all
306 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
As with the black obstruction case, the IC model
tends to under-predict surface irradiance for this south
facing surface. Only results for the improved treatment
Fig. 13. Model error probability distribut
Fig. 14. Global irradiance comparisons of IC (upper left), AC with im
sky (lower) models with RADIANCE using the Perez all weather lum
of circumsolar brightening in the AC model are pre-
sented, as the trend towards improved prediction reso-
lution is similar to that observed previously. For this
ions for a grey obstructing surface.
proved circumsolar views (upper right) and SRA with inverted
inance distribution model for south facing grey obstructions.
9 Note that interpolation errors should be inversely propor-
tional to surface size.
Fig. 15. Global irradiance comparisons for North facing grey obstructions.
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 307
model, there is also a tendency to under-predict (as the
background source is isotropic), but the scatter is re-
duced due to improvements in modelling the circumso-
lar region––although its concentration into a point
source leads to an over-prediction strand.
It is clear from the lower charts in Fig. 12 that the
method of calculating below horizontal reflected irradi-
ance using an isotropic ground plane leads to a sys-
tematic under-prediction. This is principally due to
errors involved with deriving the mean radiance of the
ground plane and in simplifying the treatment of the
angle of incidence of reflected energy on our receiving
surface. The SRA implementation (with an inverted sky
dome) has no such limitations, so that the agreement is
correspondingly good (Fig. 13).
4.2.2. Global radiation
To complete the comparisons between models, the
effects of direct insolation on both receiving and
reflecting surfaces has been accounted for. As noted in
Section 2.1, this is based on scaling incident beam irra-
diance according to the proportion of the surface that
can ‘see’ the sun. It is desirable to reduce the number of
hours at which the sun views are calculated to limit pre-
processing computation and on-line storage demands.
Typically, hourly view factors are calculated for one day
per month between and including the solstices. To re-
duce errors between these points in time to within
reasonable limits, two-dimensional interpolation is re-
quired 9 (to account for solar to clock time conversion
as well as days between months). Two interpolation
methods have been tested––linear and three-point
Lagrangian. For these two cases the RMS error for our
south facing surface is 10 and 6.8 Wm�2 respectively,
based on the SRA, so that the modest increase in com-
plexity is warranted.
Fig. 14 compares results for the three models,
assuming Lagrangian interpolation of solar views. Due
to the magnitude of direct irradiance, the degree of
scatter due to the IC and AC models is much reduced.
Nevertheless, the relative accuracy of the SRA––for
which the visible scatter is due to remaining solar view
interpolation errors––is clearly apparent. With a
diminished solar influence Fig. 15, produced for a north
facing surface, illustrates a substantial variation in
accuracy between the models. Differences between the
models for other orientations can be expected to lie
between these extremes.
308 D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309
5. Conclusions
Three models have been described which account for
the influence of obstructions in reducing sky radiation
and contributing reflected radiation. The first two ab-
stract adjacent urban obstructions into an effective
canyon, with an opposing face that is parallel to the
receiving plane and infinitely long. This has been
implemented using both isotropic and anisotropic tilted
surface irradiance models. The third is more physically
meaningful, since the geometry of the urban skyline is
preserved. For this, two hemispheres (above and below
the horizontal) are discretised into patches for which the
radiance of the sky and dominant obstructions (if these
exist) and associated view factors are used to determine
the corresponding contribution to surface irradiance––
so that we have a simplified radiosity algorithm (SRA).
The ray tracing program RADIANCE has been adopted
as a truth with which to compare predictions from these
models. From this study we conclude that
• The current general practice of ignoring obstructions
to the sky vault leads to considerable error in predict-
ing surface irradiance.
• Differences in computation time between the three
implementations are relatively small.
• For the UK climate, the isotropic canyon model
tends to considerably under-predict for south aspects
and over-predict for northern aspects, particularly in
unobstructed situations. The anisotropic canyon
model follows a similar but dampened trend.
• In obstructed situations, the re-use of solar visibility
profiles to determine the visibility of the circumsolar
point source improves the accuracy of the anisotropic
canyon model reasonably significantly. On a related
note, it is recommended that Lagrangian interpola-
tion should be used to determine solar visibility from
look up tables based on hourly solar trajectory dis-
cretisation at monthly intervals.
• The SRA is considerably more accurate than the
alternatives tested. This approach also offers longev-
ity, since the technique may be used in conjunction
with future relative sky luminance/radiance distribu-
tion models.
• Finally, the core of the SRA may be adapted to pre-
dict illuminance, both internal and external. One such
implementation will be reported in a future paper.
The proposed SRA will be incorporated within a new
sustainable urban neighbourhood modelling tool
(SUNtool) as part of an EC-funded research project. In
principal this model is amenable for inclusion within any
software application for which obstruction-sensitive
irradiance predictions are important, such as dynamic
thermal simulation and renewable (solar) energy mod-
elling programs.
Acknowledgements
The funding for this work by the European Com-
mission’s Directorate General for Transport and Energy
is gratefully acknowledged.
Appendix A. Determination of UHA
The equivalent UHA u of some arbitrary arrange-
ment of adjacent obstructing surfaces may be found
from the following expression:
1
2cosðb þ uÞ ¼ 1
p
Z ZS
cos n dx ðA:1Þ
where n is the angle of incidence of some small
obstruction element and dx its solid angle. The accuracy
with which u can be solved for with this expression,
using some numerical procedure may be determined as
follows. Suppose it is required that j DIdb
Idhj6 0:01, where
DIdb is the error in the calculated incident diffuse irra-
diance. This error can be approximated by
DIdb ¼ dIdb
dudu ¼ � sinðb þ uÞ
2Idh du
where du is the error in the calculation of u.The worst case will therefore be when sinðb þ uÞ ¼
�1. Therefore 12du6 0:01, i.e. u must be calculated to an
accuracy of 0.02 rad, or 1.16�.In order to calculate the reflected component of
irradiance, a value for the reflectance of the equivalent
obstruction is required. This may be determined either
by averaging over all of the obstructing surfaces or by
simply taking the reflectance of the main obstructing
surface (i.e. the surface that makes the greatest contri-
bution toR RS
cos n dx).
Appendix B. Solution by matrix inversion
The equations presented in Section 2 are sufficient to
calculate the irradiance on a surface using various sky
models and given certain geometric information. The
solution of the equations however is not straightfor-
ward. The inclusion of reflections means that the irra-
diance of a particular surface potentially depends on the
irradiance of many other surfaces, particularly in the
case of the SRA, where up to 290 different surfaces could
be involved in the reflection calculation (one for each
patch).
An obvious way to solve this system of equations is
by iteration, using the component of irradiance received
directly from the sky and the sun as a starting condition.
This approach is reasonably physically realistic, simu-
lating the effect of successive reflections (albeit in a crude
D. Robinson, A. Stone / Solar Energy 77 (2004) 295–309 309
way). An alternative approach would be to formulate
the problem as a matrix equation and solve by inversion.
For the SRA this would give Id ¼ AIg þ BR; where
Ig ¼ Id þ Ib, is a vector listing the global irradiance on
each surface, and R a vector giving the radiance of each
sky patch.
Rearranging:
Id ¼ ðI � AÞ�1ðAIb þ BRÞ ðB:1Þ
The matrix A is square and describes how the direct
component of irradiance falling on each surface is
eventually distributed around the n surfaces in the world
(entry ði; jÞ in the array describes the proportion of di-
rect insolation on surface j that is reflected to surface i):
A ¼
q1k1;1p
q2k1;2p qnk1;n
p
q1k2;1p
. .. ..
.
..
. . .. ..
.
q1kn;1p
q2kn;2p qnkn;n
p
2666664
3777775
ðB:2Þ
where qi is the reflectance of surface i, and ki;j describes ascaling factor for the effect of the energy reflected from
surface j to surface i. If surface j obstructs m sky patches
when viewed from surface i, denoted by x1; x2; . . . ; xm,then
ki;j ¼Xmk¼1
Ui;xk ð1� ri;xk � rself ;xk Þ cos ni;xk
(after (13)) (ri;xk is the view factor from surface i to sky
patch xk and Ui;xk is the solid angle of sky patch xk fromsurface i).
Matrix B describes the contribution from each sky
patch (of unit radiance) to the irradiance received by
each surface within the world:
B¼
U1;1r1;1 cosn1;1 U1;2r1;2 cosn1;2 U1;pr1;p cosn1;p
U2;1r2;1 cosn2;1. .. ..
.
..
. . .. ..
.
U1;nr1;n cosn1;n U2;nr2;n cosn2;n Un;prn;p cosnn;p
266664
377775
ðB:3Þ
The matrices ðI � AÞ�1A and ðI � AÞ�1B need only be
computed once for any given geometry. Therefore at each
time step once the direct component of irradiance has
been solved for, only two matrix multiplications and an
addition are required to solve for the diffuse component.
References
Baker, N.V., Steemers, K.A., 1994. LT Method Version 2.0.
Cambridge Architectural Research Ltd, Cambridge, UK.
CIE, 1973. Standardisation of luminous distribution on clear
skies. CIE publication no. 022.
CIE, 1996. Spatial distribution of daylight––CIE Standard
Overcast Sky and Clear Sky. CIE S 003/E-1996.
CIE, 2002. Spatial distribution of daylight––CIE Standard
General Sky. CIE DS 011.2/E-2002.
Clarke, J.A., 2001. Energy Simulation in Building Design,
second ed. Butterworth Heinemann, Oxford.
Hay, J., 1979. Calculation of monthly mean solar radiation for
horizontal and inclined surfaces. Solar Energy 23, 301.
Jensen, S.O., 1994. Validation of building energy simulation
programs. Final Report, PASSYS Model Validation Sub-
group, EUR 15115 EN (European Commission).
Kittler, R., 1967. Standardisation of the outdoor conditions for
the calculation of the Daylight Factor with clear skies. In:
Proc. Conference on Sunlight in Buildings, Bouwcentrum
Rotterdam, pp. 273–286.
Klucher, T.M., 1979. Evaluation of models to predict insolation
on tilted surfaces. Solar Energy 23 (2), 111–114.
Moon, P., Spencer, D.E., 1942. Illumination form a non-
uniform sky. Illuminating Engineering 37 (10), 707–726.
Muneer, T., 1997. Solar Radiation and Daylight Models for the
Energy Efficient Design of Buildings. Architectural Press,
London.
Perez, R., Seals, R., Ineichen, P., Stewart, R., Menicucci, D.,
1987. A new simplified version of the perez diffuse irradiance
model for tilted surfaces. Solar Energy 39 (3), 221–231.
Perez, R., Ineichen, P., Seals, R., Michalsky, J., Stewart, R.,
1990. Modelling daylight availability and irradiance com-
ponents from direct and global irradiance. Solar Energy 44
(5), 271–289.
Perez, R., Seals, R., Michalsky, J., 1993. All-weather model for
sky luminance distribution––preliminary configuration and
validation. Solar Energy 50 (3), 235–243.
Robinson, D., 2003. Climate as a pre-design tool. In: Proceed-
ings of the Eighth International IBPSA Conference on
‘Building Simulation 2003’, Eindhoven, Netherlands, 11–
14th August 2003, pp. 1109–1116.
Robinson, D., Stankovic, S., Morel, N., Deque, F., Rylatt, M.,
Kabele, K., Manolakaki, E., Nieminen, J., 2003. Integrated
resource flow modelling or urban neighbourhoods: project
SUNtool. In: Proceedings of the Eighth International
IBPSA Conference on ‘Building Simulation 2003’, Eindho-
ven, Netherlands, 11–14th August 2003, pp. 1117–1122.
Temps, R.C., Coulson, K.L., 1977. Solar radiation incident on
slopes of different orientation. Solar Energy 19 (2), 179–184.
Tregenza, P., Sharples, S., 1993. Daylighting algorithms. ETSU
S 1350-1993, UK.
Ward Larsen, G., Shakespeare, R., 1997. Rendering with
Radiance––The Art and Science of Lighting Visualisation.
Morgan Kauffmann, San Francisco.