Renewable Energy 99 (2016) 325e335

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Estimation of Hong Kong’s solar energy potential using GIS and

remote sensing technologies
Man Sing Wong a, *, Rui Zhu a, Zhizhao Liu a, Lin Lu b, Jinqing Peng b, Zhaoqin Tang a,
Chung Ho Lo a, Wai Ki Chan a
a
Department of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong
b
Department of Building Service Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong

a r t i c l e i n f o a b s t r a c t

Article history: This paper studies the use of Remote Sensing (RS) technologies and Geographic Information Systems
Received 2 February 2016 (GIS) for estimation of city-wide photovoltaic (PV) potential in Hong Kong. It investigates the spatial
Received in revised form distribution of cloud coverage through geostationary satellites from the Multi-functional Transport
2 June 2016
Satellite (MTSAT). The results indicate that a non-prominent spatial variation of cloud cover presides over
Accepted 3 July 2016
Available online 16 July 2016
a majority of Hong Kong territories. Appropriate locations for deploying solar PV panels, such as rooftops,
were delineated using RS, GIS, and existing ancillary data. Extraction and filtering of pixels based on a set
of criterions were used to identify optimal PV rooftops. This study shows that the summarization of PV
Keywords:
Building rooftop
potentials in Hong Kong is 2.66 TWh on building rooftops. The methodologies and findings from this
Photovoltaic potential study permits detailed spatial estimation of city-wide solar energy potential, and assists the policy-
Satellite imagery decision process on the use of renewable energy in Hong Kong.
Solar radiation © 2016 Elsevier Ltd. All rights reserved.

1. Introduction energy that are currently available in Hong Kong includes: solar
energy, wind energy, bio-gas, and bio-diesel fuel. However, the
Energy consumption in a metropolitan city such as Hong Kong is amount of renewable energy outputs only accounts for 0.1% of the
extremely high, with approximate electricity consumption of total electricity consumption. The solar energy contributes about
155,079 TJ; oil & coal products consumption of 83,275 TJ; town Gas 1.5% of total renewable energy usage in 2012 [4]. To enhance the
& LPG consumption of 49,616 TJ in 2012 [1]. Emissions of the development of solar renewable energy in Hong Kong and increase
anthropogenic greenhouse gas have increased since than ever [2] the rate of PV deployments, a study of the potential of developing
due to the rapid economic and population growth. In Hong Kong, solar photovoltaic energy in Hong Kong is urgently needed. The
about 97% of total carbon dioxide emissions resulted from elec- estimation of rooftop PV potential provides fundamental data for
tricity generation [3,4]. The greenhouse gases are the major caus- future energy policy decision-making, urban sustainable develop-
ative factors in global warming, which accelerate the insulating ment, and city planning by the Government of Hong Kong Special
effect in the atmosphere. The mitigation of climate change then Administrative Region (HKSAR). The PV potential is deemed as the
becomes predominant in public awareness. annual potential energy produced from PV technology in this study.
Using renewable energy is one of the approaches to mitigate the Solar radiation is a controlling factor in electricity generation of
greenhouse effect. Solar photovoltaic (PV) technology is a widely a PV system because of the complex energy interactions between
adaptable application and converts the solar energy into electricity the atmosphere and surface [6]. The amount of incident solar ra-
with promising efficiencies [5]. The major types of renewable diation affects the electricity produced by PV systems significantly.
The energy of solar radiation varies greatly due to the refraction and
scattering of aerosols, water vapour contents, and air particulates.
* Corresponding author. At the global scale, the geometry of the Earth and its rotation
E-mail addresses: lswong@polyu.edu.hk (M.S. Wong), r.zhu@connect.polyu.hk determine the latitudinal gradient radiation. At the local scale, the
(R. Zhu), george.liu@polyu.edu.hk (Z. Liu), vivien.lu@polyu.edu.hk (L. Lu), solar radiation map depends on the surface terrain [6]. The varia-
jallenpeng@gmail.com (J. Peng), zhaoqian.tang@connect.polyu.hk (Z. Tang), lo.
tions in elevation, slope, aspect, and shadowing effects cause sig-
chungho@connect.polyu.hk (C.H. Lo), waikiedward.chan@connect.polyu.hk
(W.K. Chan). nificant deviations of solar radiation [7]. Different methods can be

Latitude Gradient Radiation

http://dx.doi.org/10.1016/j.renene.2016.07.003 Variations in elevation ,slope,aspect and shadowing effects.
0960-1481/© 2016 Elsevier Ltd. All rights reserved.
 326 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335

used to estimate solar radiation according to the required scale and conclusions.
accuracy.
Variations in elevation gradient, together with shadowing effect, 2. Study area and data used
can cause significant local irradiance fluctuations [8]. In order to
Almost all
of the res improve the spatial resolution and thematic accuracy of the solar Hong Kong, located on 22
latitude with a sub-tropical climate,
studies
overlooked radiation map, high resolution data derived from airborne Light has more than 7 million population with an average population
the importance
of weather
Detection and Ranging (LiDAR) technology have also been used in density of 6544 per km2 [18], which consumed 153,362 TJ elec-
as an recent studies [9e11]. LiDAR technology is used to measure the tricity in 2013 [19]. The Hong Kong area covers Hong Kong Island,
influectial
factor for distances between sensor and objects by illuminating with a laser Lantau Island, Kowloon Peninsula, and the New Territories,
studying
solar pv beam. The point cloud processes from LiDAR data such as extrac- including 262 outlying islands. The urban landscape as known by
potential tion, segmentation, and reconstruction of building rooftops for many, comprising of high-rise buildings for both commercial and
solar photovoltaic deployment have been studied [9,10]. Research residential purposes, occupies around 8% of Hong Kong’s land area.
such as that of Nguyen et al. [11] studied the solar photovoltaic However, 90% of Hong Kong’s electricity consumption comes from
potential by considering many influential factors (e.g. terrain and these buildings.
near ground shadowing effects), however the weather conditions
have not been considered. The solar radiation model integrated 2.1. Satellite data
with Geographical Information Systems provides a means for
proper estimation [11]. The solar radiation model is a physical In order to derive a cloud cover probability map for Hong Kong, a
model, empirical equations are adopted in order to provide fast and year’s worth of Multi-functional Transport Satellite images
accurate estimation of solar radiation. It also considers the effect of (MTSAT-1R/MTSAT-2) were acquired for the year of 2012 and used
slope, aspect, and shadow from the surrounding environment. in this study. The images cover the Earth’s surface from 5
N to 55

Several GIS-based solar radiation models have been developed such N and from 75
E to 145
E with an hourly-based temporal reso-
as SolarFlux, which simulates the shadow patterns by direction lution. The acquired data are then post-processed at a 4 km spatial
insolation at specified time intervals [12], or the Solei model, which resolution for both visible and infrared channels, where the original
is a standalone model that works together with GIS software IDRISI resolutions are 1 km and 4 km respectively. All MTSAT image files
[13]. Both SolarFlux and Solei use simple empirical formulas [14] are stored in HDF-4 format, consisting of seven data layers: VIS, IR1,
and parameters represented by generalized values, which implies IR2, IR3, IR4, Latitude, and Longitude.
that they may not be suitable for an accurate estimation for a large
region. Solar Analyst is an extension module of ArcGIS [15], which 2.2. Elevation data - Hong Kong DSM and DEM
derives solar radiation map based on the input DSM and DEM data
[16,17]. Several other factors are also considered during the process The Digital Surface Model (DSM) and Digital Elevation Model
such as slope, aspect, solar angle, shadow casting by surrounding (DEM) data were generated by the airborne LiDAR point cloud data
topography, and atmospheric attenuation [9]. For the calculation of and stored as raster format. The data acquisition was carried out for
diffuse proportion and atmospheric transmissivity, the parameters the entire Hong Kong territory by the Civil Engineering and
input into the Solar Analyst can be in reference to the nearest Developing Department (CEDD) of Hong Kong SAR Government
meteorological station data or typical default values [11]. Another using airborne laser scanner observation between December 2010
solar radiation model, the SRAD model, simulates the interactions and January 2011. The average point spacing is 0.5 m (4 pt/m2) and
between longwave and shortwave solar radiation with the Earth’s the horizontal and vertical accuracies are 0.3 m and 0.1 m respec-
surface and its atmosphere. The main solar radiation factors are tively [20].
considered for the spatial variability of landscape processes based
on a simplified parameterization [6]. However, it is designed for 2.3. Hong Kong building GIS data
analyzing the topographical and meso-scale process, thus the
estimation over large areas may not be appropriate. The GIS data for building footprints in Hong Kong are in polygon
In this study, Solar Analyst was selected to estimate the solar shapefile format. Each polygon represents the corresponding
radiation of unused areas on rooftops. The input data for solar ra- building in an object space. The corresponding attribute table for
diation calculations are DSM and DEM generated from airborne each polygon contains a list of attributes including: area, geo-
LiDAR data. The incoming solar radiation received from the sun is reference number, and building names. GIS data were used to
the primary energy source, containing two main parts: direct ra- extract building outlines and building pixels, and later exported to
diation which is intercepted and unimpeded, and diffuse radiation the solar potential calculation of individual buildings.
which is scattered by atmospheric constituents such as clouds and
aerosols. 2.4. Hong Kong weather data
The objectives of this study are: (i) to analyze the spatial dis-
tribution of possible cloud covers in Hong Kong; (ii) to develop a The weather data of 2012, including the cloud cover data, can be
robust method for estimating annual solar potential in areas of retrieved from the website of Hong Kong Observatory [21].
unused building rooftops; and (iii) to further validate the results
from solar radiation modeling with ground-based observation. The 3. Derivation of cloud probability map
methodologies and findings from this study can enable detailed
spatial estimations of city-wide solar energy potential, where the This section illustrates the method for the derivation of cloud
generated potential energy can be fed back into the grid and/or be probability map using geostationary satellite images in Hong Kong.
used directly in households/offices consumption.
This paper is organized as follows: Section 2 outlines the de- 3.1. Calculation of solar and satellite angles
scriptions of the study area and data used, Section 3 presents the
derivation of cloud probability map, Section 4 describes the method For the image pre-processing, solar and satellite view angles are
for estimating solar potentials in areas of unused rooftops, and critical for determining day-time observation and in the use of
Section 5 summarizes the major findings, limitations and radiative transfer model for determining the cloud thresholds. The
 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335 327

geostationary MTSAT satellite images consist of latitude and visible and infrared wavelengths based on monthly averaged values
longitude information of each pixel, and metadata describing the were proposed, for example the threshold test of R0.6 implemented
satellite orbital information. These data enable calculation of the using dynamic thresholds. The threshold offsets were analyzed to
instantaneous satellite viewing geometry. Solar geometry and ensure the actual threshold values have a large discrepancy be-
viewing geometry are composed of several angles: solar zenith tween cloud-contaminated and cloud-free pixels. Land and water
angle qsun (SZA); satellite zenith angle qsat (also known as viewing surfaces were applied with both proper surface RTM bidirectional
zenith angle, VZA); and the difference angle between solar azimuth reflectance distribution functions (BRDFs) and RTM simulations.
angle (asun) and satellite azimuth angle (asat), known as relative The simulation of cloud-free reflectance of each pixel was based
azimuth, represented as arel. Estimated angles in this section were on three-dimensional centroid-based interpolation method (Fig. 1).
used as inputs in the radiative transfer model. If the calculated solar The apparent reflectance of clear-sky TOA is noted as P(X, Y, Z),
zenith angle is larger than 90
, it is then defined as night-time where the X, Y, Z is the value of qsun,qsat, arel as shown in Equation
observation. (2). x1, x2 are the x values difference between P(X,Y,Z) and p1, p2.

3.2. Radiative transfer model

Radiative transfer model is used to calculate radiance of the

Earth’s atmosphere at different wavelengths under different at- 1
mospheric and surface conditions. The radiances at the top of the PðX; Y; ZÞ ¼ ðx2 y2 z2 DP1 þ x1 y2 z2 DP2 þ x1 y2 z1 DP3
DxDyDz
atmosphere (TOA) measured by satellite sensors are influenced by
atmospheric scattering and surface reflected solar radiance. In this þ x2 y2 z1 DP4 þ x2 y1 z2 DP5 þ x1 y1 z2 DP6
study, atmospheric parameters retrieved from satellite images þ x1 y1 z1 DP7 þ x2 y1 z1 DP8 Þ
together with other climatology and ancillary data were used as
(2)
inputs in radiative transfer model to simulate the cloud-free TOA
apparent reflectance for MTSAT visible wavelength. The simulation The difference between eight points in the x, y, z direction are
was conducted using the “Second Simulation of the Satellite Signal indicated as Dx, Dy, Dz, respectively, which are the step values of
in the Solar Spectrum” (6S) radiative transfer model [22]. qsun,qsat, arel (Dx ¼ x1þx2, Dy ¼ y1þy2, Dz ¼ z1þz2).
The 6S model is a fundamental radiative transfer code used for The cloud mask algorithm contains a set of threshold tests
the calculation of look-up tables in atmospheric correction. For a where several criterions are tested together. If the pixel values of
Lambertian surface, the TOA apparent reflectance rTOA can be corrected reflectance and brightness temperature are similar to
calculated using Equation (1): that of the threshold values, these pixels are considered as cloud
contaminated areas with low confidence probability. After the
rt
rTOA ðqsun ; qsat ; arel Þ ¼ ra ðqsun ; qsat ; arel Þ þ Tðqsun ÞTðqsat Þ determination of dynamic and static thresholds, threshold tests
1  rt S were implemented in order, as shown in Fig. 2. The annual average
(1) cloud cover probability over Hong Kong territory was then derived
and illustrated in Fig. 3.
where ra is path reflectance (caused by Rayleigh scattering and The cloud probability in Hong Kong shows a non-prominent
aerosol scattering) without surface contribution, rt is the surface spatial distribution varying from 0.58 to 0.62, generated from
reflectance, S is the spherical albedo of the atmosphere, and T(qsun) hourly geostationary satellite images of year 2012. The insignificant
T(qsat) is the total transmittance. For this study, land areas with variation is mainly caused by: (i) low spatial resolution of geosta-
surface properties of grass, forest, and barren land were assumed as tionary MTSAT images (e.g. 4 km), where the Hong Kong territory
land cover types. Over ocean areas, the reflectance values of lake has an area of 1104 km2 but the areal resolution of satellite images
water properties were defined in the 6S model isotropy. is 16 km2, thus the entire territory has only 69 pixels therefore
In order to reduce the computational workload and time, “Look some details or small clouds may be missed; and (ii) the geographic
Up Tables” (LUTs) were developed for various discrete atmospheric location of Hong Kong, where it is situated along a coastal region
and solar satellite viewing geometry. The mid-latitude winter/ and is at sub-tropic climate zone, which is a cloud-prone area and
summer atmospheric profiles with various surface properties ac- the homogenous spatial patterns of cloud formation are expected.
cording to different land covers were defined. The IWAVE spectral
response function (SRF) was interpolated at a 2.5 nm wavelength
interval.

3.3. Cloud detection using threshold approach

In the same solar satellite viewing geometry, the reflectance

values of clouds are often larger than the reflectance values of land
or ocean. Cloud temperatures observed by satellite are often cooler
than the underlying surface temperature. In this paper, a robust-
ness and dynamic threshold method on MTSAT images was
developed. According to the surface types and solar illuminations,
different sets of threshold tests were applied on each pixel. When
sun is above the horizon (e.g. solar elevation angle is higher than
3
), the bidirectional reflectance at 0.6 mm (R0.6) was used to detect
pixels containing cloud or snow.
The multispectral threshold identification method has been
well-researched for determining the cloud mask. In this paper, new
dynamic thresholds for visible wavelength and thresholds for both Fig. 1. Schematic representation of 3D centroid-based interpolation.
328 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335

Fig. 2. Workflow for cloud detection approach.

4. Estimation of photovoltaic potential DSM, and the size of calculated solar radiation map should be the
same as the input data. The latitude and longitude are automati-
A schematic diagram for calculating photovoltaic potential in cally calculated from the input raster. Used in solar radiation cal-
Hong Kong on annual basis is presented in Fig. 4. culations, the sky-size is defined as 200 cells per side for the
In this study, a 3  3 median convolution was applied to both resolution of the viewshed, sky map and sun map raster images. A
DSM and DEM airborne LiDAR data [23]. The DSM was resized to a 3 raster sky representation is generated for both clear and obstructed
m resolution and used for calculating solar radiation in Solar Ana- views at a given location. The skymap is then estimated, with the
lyst. Several criterions were applied to filter off unwanted pixels. viewshed in eight different directions, to determine the maximum
After eliminating the ground pixels, barriers (buffer minus 1 m) on angle of sky obstruction or horizon angle. The diffuse model type is
rooftops, shadows, and steep sloping pixels using decision tree the standard overcast sky. Since the results show that a non-
classification, the optimal areas of rooftop pixels could be identi- prominent spatial variation of cloud covers presides over entire
fied. The building polygons and solar radiation map were then Hong Kong territories ranging from 0.58 to 0.62, the diffuse pro-
spatially joined. It was assumed that at least two solar panels would portion and the transmissivity were calculated using the Hong
be deployed in each candidate installation site; however polygons Kong cloud coverage data from a station at the Hong Kong Obser-
(sites) with areas smaller than 3 m2 were then removed from the vatory in year 2012. It is also assumed that a single set of equations
analysis. Finally, a territory-wide solar PV potential was derived as for estimating the diffuse proportion and the transmissivity can be
shown in Fig. 4. applied to the entire Hong Kong territories due to the insignificant
spatial variation of cloud covers. The formulae are expressed as
4.1. Solar radiation modeling Equations (3) and (4):

T ¼ 0:7 Pclear sky þ 0:5 Ppartial cloudy þ 0:3 Pcloudy (3)

A 3  3 median filter, used to remove the signal noise, was
implemented using DSM and DEM airborne LiDAR data. The pixels
were aggregated and the internal variation inside 3  3 kernels was D ¼ 0:2 Pclear sky þ 0:45 Ppartial cloudy þ 0:7 Pcloudy (4)
reduced. This filtering process worked well on non-continuous
data, such as edges of buildings which were not suitable for where T is the transmissivity, D is the diffuse proportion, Pclear is the
installing PV arrays. Since the size of entire Hong Kong DSM data is proportion of clear days, Ppartialcloudy is the proportion of partly
large and difficult to be processed in ArcGIS together at a single cloudy days and Pcloudy is the proportion of cloudy days. A clear day
time, the DSM data were divided into four parts with 6 km over- is defined as 0e30% average percentage of cloud cover, partial
lapping areas in both easting and northing coordinates. And the cloudy day is 40%e70%, and cloudy is 80%e100% [24]. The weather
resolution of the input DSM was downscaled to 3 m. data of 2012, including the cloud cover data, can be retrieved from
In the above sketch (Fig. 5), the CC part is the size of the input the Hong Kong Observatory. Table 1 shows the computation of
 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335 329

Fig. 3. Cloud probability map of Hong Kong.

diffuse proportion and transmissivity. amount of annual solar energy from solar irradiation, the most
The average transmissivity and diffuse proportion in 2012 are appropriate locations for installing solar panels were then selected.
T ¼ 0.37 and D ¼ 0.61 respectively. Compared to the default values Compared to threshold-based filtering approaches, the hierarchy of
of T0 ¼ 0.5 and D0 ¼ 0.3, the transmissivity is lower and diffuse decision tree classification was able to deliver the order of process
proportion is higher. Solar analyst can simulate the solar radiation and classification required for the different areas, which in this
at any given time instant with a much higher spatial resolution (i.e. study was used to classify the rooftops appropriate for the
several meters) compared with NASA-SSE (i.e. results are 1
lati- deployment of PV arrays.
tude by 1
longitude grid cells) and SolarGIS (i.e. original spatial
resolution is about 3e5 km), which is essential to investigate solar (1). Ground mask
radiation distribution in a micro-scale. The operation of Solar An-
alyst is time- and computer-demanding, taking around three days A filtering of ground pixels was implemented based on the
to complete the processing of one quarter of the data using a server Object Height Model, created through the subtraction of DEM from
computer. DSM to eliminate the effects of the difference caused by different
The maximum annual solar radiation from the model outputs is time frames of the airborne LiDAR data and building polygon
about 2000 kWh/m2, and the minimum value is 5 kWh/m2. The shapefiles. The pixels with an object height below 2.5 m are
mean annual value is 1497 kWh/m2. According to the histogram considered as the ground pixels.
analysis, about 40% of pixels are less than ca. 800 kWh/m2. The
value of 800 kWh/m2 was observed as the first break value of five (2). Slope
class classification using the Jenks Natural Breaks classification
method. The obstruction from surrounding high-rise buildings may The tilting of PV modules can be used to receive the maximum
greatly affect the direct radiation on rooftops. The shadowing effect solar radiation and to avoid unwanted shading [25]. The residential
then becomes predominate. Considering the cost payback time and PV systems are usually installed on sloped roofs while the com-
efficiency of solar panels, the pixels with values below the mercial systems are installed on flat or low-slope rooftops [26].
threshold (e.g. <800 kWh/m2) were excluded. However, some rooftops with steep slopes may lower the efficiency
of electricity generation. The received solar radiation from PV
modules decreases significantly, when the slopes exceed 40
[25]. It
4.2. Determination of PV potential on optimal building rooftops indicates that 40
is the maximum threshold for installing the PV
modules. Thus, a slope calculation was processed using DSM data.
There are several criteria for selecting the optimal locations of Pixels with steep slopes would be excluded.
installing PV arrays. High resolution airborne LiDAR data were used
to map the unused building rooftop areas. To acquire maximum
330 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335

Fig. 4. Flow chart of the estimation of PV potential.

appropriate for installing PV modules. By applying a “minus 1 m”

buffer areas, the edges of building footprints could then be filtered.
The results can be converted into raster format for further
processing.

(4). Shadow identification

In Hong Kong, shadowing effect from the nearby skyscrapers

can greatly reduce the amount of solar electricity generation. Since
the Solar Analyst accounts for the shadows cast by surrounding
topography [15], a shadow detection algorithm can then be applied.
The variation in solar radiation should then be taken into account
on the topographic features and obstructions of the surrounding
objects. If the value of solar radiation is lower than a certain
threshold, it can be assumed that it is under shadows at the ma-
jority of the time. As the insolation map already takes into account
of the aspect, slope, and viewshed of the topographic features, the
threshold of lower values can then be used to filter off the pixels
which are always covered by shadow. From the statistics of the
solar radiation distribution, thresholds can be determined using the
Jenks Natural Breaks method - those pixels with values lower than
800 kWh/m2 are defined as shadow.
Fig. 5. Sketch of overlapping areas of solar radiation estimation.

4.3. Estimating solar photovoltaic potential on rooftops

(3). Barrier of building rooftop
In Hong Kong, most of the rooftops of commercial and resi-
Most of the building rooftops contain barriers which are not
dential buildings are usually flat with bitumen surface. Therefore,
 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335 331

Table 1
Transmissivity and diffuse proportion of year 2012.

Month No. of clear days No. of partly cloudy days No. of cloudy days No. of days Transmissivity Diffuse proportion

Jan 0 9 22 31 0.358065 0.627419

Feb 0 2 27 29 0.313793 0.682759
Mar 3 6 22 31 0.377419 0.603226
Apr 0 5 25 30 0.333333 0.658333
May 0 6 25 31 0.33871 0.651613
Jun 0 6 24 30 0.34 0.65
Jul 0 14 17 31 0.390323 0.587097
Aug 0 12 19 31 0.377419 0.603226
Sep 3 14 13 30 0.433333 0.533333
Oct 2 20 9 31 0.454839 0.506452
Nov 0 10 20 30 0.366667 0.616667
Dec 1 9 21 31 0.370968 0.61129

Fig. 6. Data conversion.

deployment of commercial rooftop PV system is more appropriate where En is the nominal plant energy output (kWh), Ga is the
in Hong Kong. The PV panels are usually tilted at approximately average solar radiation intensity (kWh/m2) in a certain period of
latitude angle in order to maximize power production. In Hong time per unit area, A is the generator area (m2) of the PV plant and h
Kong, this optimal tilting is around 14
e20
[27] to produce the is the efficiency factor of the PV modules. The factor of nominal
maximum total power output and the south-facing orientation has power divided by the incident light intensity is the module effi-
a higher annual average insolation. The solar PV technologies are ciency which is the ratio of electrical output from light energy. The
suitable for large-scale deployment and could be a significant above nominal plant power output assumes that the performance
source of renewable energy in Hong Kong [27]. ratio is 100%. However, the actual output is not the same due to the
The Photovoltaic Geographic Information System (PVGIS) lumped contribution of the sources of performance loss. The actual
database was then developed to estimate the potential solar elec- energy output can be calculated using Equation (7) [29]:
tricity generation of the PV module at horizontal, vertical, and
optimal inclination [28]. The following equation was applied for Eout ¼ En  PR ¼ Ga  A  h  PR (7)
calculating the annual potential electricity generation E (unit: kWh)
using defined module configuration and orientation [28]: where Eout is the electrical energy output by the PV plant and PR is
the system performance ratio. The typical value of 0.75 for the
E ¼ Pk  PR  Gs (5) performance ratio of crystalline silicon type of PV module was
adopted in this study. The efficiency of standard crystalline silicon
where Pk is the unit nominal power or the peak power in (kWp), PR module is about 17%. The annual insolation map produced by the
is the system performance ratio and Gs is the sum of global irra- Solar Analyst could be used as the input of the annual sum of global
diation (kWh/m2) on the surface yearly. The system performance irradiation (kWh/m2). The PV potential based on each building was
ratio is a constant which describes the relationship between the then calculated in the GIS platform.
actual power output and theoretical power output. The perfor-
mance ratio may vary as the external environment changes. For 4.4. Data conversion
example, a rise in temperature will decrease the performance ratio.
A typical value of 0.75 is assumed for the roof mounted system with Each pixel in the raster image was converted into a point with
modules from mono- or poly-crystalline silicon type [28]. The size insolation value. The points falling outside the area of optimal PV
of the systems is measured in nominal power (Wp) which repre- polygon were filtered off. The output of decision tree is a binary
sents the maximum power output of PV modules at Standard Test raster image, where “1” indicates an appropriate area and “0” in-
Conditions (STC) [28]. The nominal plant energy output of a PV dicates a filtered area. The output raster was then multiplied with
plant can be calculated using Equation (6) [29]: the solar radiation output data to estimate the solar radiation of
unused areas. These solar radiation pixel values were converted
En ¼ Ga  A  h (6)
from raster into the point, such as illustrated in Fig. 6.
In Fig. 6 above, the red rectangles (a) indicates the output
332 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335

building area after decision tree classification; (b) is the estimated and diffuse solar radiation continuously between November and
solar radiation value with unit of kilowatt hour per square meter December 2012. Result shows that a value of 140 kWh/m2
(kWh/m2); (c) is the result from raster calculator, using (a) multiply compared with simulated 152 kWh/m2 suggesting about 91.4%
by (b); and (d) is the conversion from raster map (c) into point. It accuracy for the model.
was observed that there are 2 pixel values smaller than the defined
threshold, and these pixels would be excluded. There are 9 points in
total, but only 4 values are above 0. In this example, the total po- 5.3. Results
tential solar radiation on rooftop area is the sum of estimated solar
radiation in Fig. 6(c) and is equal to 589 kWh/m2 if the solar The estimated solar photovoltaic potential of rooftops in Tsim
photovoltaic arrays are only installed on the appropriate areas. Sha Tsui, Kowloon peninsula is illustrated as an example in Fig. 7.
In this study, there are a total of 29,743,281 points and among Hong Kong contains about 309,606 buildings, 239,833 buildings are
them, 10,660,310 have null value. It indicates that around 1/3 suitable for solar panel installation. The sum of PV potential is about
rooftop pixels of Hong Kong buildings are unsuitable for installa- 2.66 TWh. However, the mean potential PV per building is only
tion of solar photovoltaic modules. After converting the raster im- about 11.094 MWh. A map accounting for the percentage of optimal
age into point data, spatial join was performed to estimate the PV area divided by the total rooftop area is illustrated in Fig. 8. This
efficiency and total electricity generation in each building. map shows the feasibility of the PV system installed on individual
Spatial join is a function of joining attributes from one feature to buildings. It is also observed that large buildings have a larger
another according to the spatial relationship [30]. The polygons of percentage of PV deployment than the smaller ones. Considering
optimal PV areas and the insolation points were spatially joined certain areas on the rooftops are not appropriate for the deploy-
into building polygons. The sum of optimal PV areas and average ment of solar panels due to the all-year-round shadows caused by
solar radiation of each building were calculated if the points surrounding high-rise buildings and walls, and the necessary space
completely fall inside the building footprints. between panels, the utilization rate is calculated to indicate areas
that can be used for each rooftop. In Fig. 9, about 10% of buildings
5. Results and discussion have utilization rate below 40%, most of which are residential
buildings.
5.1. Validation of the developed methodology The total and average PV potentials in the residential regions are
larger than that of the commercial region. It indicates that the ef-
The proposed methodology was validated through the direct ficiency and total electricity generation in the residential areas are
comparison with MODIS cloud mask products MOD06 observed higher than the commercial areas. However, it is observed that the
between 2 a.m. and 3 a.m. GMT time. MODIS instrument with 36 efficiency of residential buildings in commercial areas is lower than
channels provides a higher spatial (1 km) and higher spectral res- the commercial buildings in the same areas that contain larger
olution from the shortwave visible to the longwave infrared spec- rooftops, which is mainly due to lots of skyscrapers in commercial
trum. The comparison was conducted on a pixel by pixel basis. A districts. The skyscrapers obstruct the viewshed of the low-rise
temporal matching criterion is ±5 min between the MODIS and buildings, therefore, the solar radiation incident on the PV is
MTSAT measurements. Results show that the lowest and highest greatly reduced. In the residential areas, the buildings are spatially
success ratios of validation dataset are 0.84 and 0.97, respectively. distributed and discrete. It greatly reduces the chance of obstruc-
These success ratios indicate that high degree of consistency of tion by surrounding buildings. Some buildings, especially for the
cloud detection between the MTSAT and MODIS MOD06 products. village houses, are low-rising and have better viewsheds. A better
viewshed gives a higher efficiency in generation of electricity.
5.2. Validation with ground-based solar radiation observation The residential buildings have the lowest utilization rate but
industrial buildings have the greatest utilization rate. The industrial
To validate the solar radiation simulation model, pyranometers buildings are usually located at lower building density districts and
with sensitivity of about 7 mV/(W/m2) have been set on the rooftop with larger rooftop areas. As a result, the PV systems on industrial
of the Hong Kong Polytechnic University (Fig. 7) to measure direct rooftops can generate electricity with a higher efficiency.

Fig. 7. Location and layout of the auxiliary instruments.

 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335 333

Fig. 8. Estimated solar photovoltaic potential on rooftops over Kowloon peninsula and northern part of Hong Kong Island.

5.4. Limitations in data processing map and building footprints are in vector files. The different reso-
lutions between the DEM/DSM, and land use map results in posi-
(1). Sources from different datasets tional displacement. The time of acquisition between LiDAR data
and GIS data are also different.
Different resolutions of data sources have long been a challenge In addition, from the results of decision tree, it was observed
in data processing. In this study, the resolutions of DEM and DSM that the utilization rates of large buildings are higher than small
are resampled at 3 m resolution. GIS data including the land use buildings. This is mainly due to the pixels from small buildings

Fig. 9. Utilization rate of PV deployment in Kowloon peninsula and northern part of Hong Kong Island.
334 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335

being already filtered by the potential barriers or other obstacles. Funding (2013.A6.024.13A) from the Central Policy Unit, the Gov-
The inconsistencies between building polygons and LiDAR data ernment of the Hong Kong Special Administrative Region. The au-
would pose variability to the potential areas. Although the mis- thors would like to thank the Hong Kong Observatory for the
matching areas are small in this study, the total number of small MTSAT images and climate data, the Hong Kong Planning Depart-
buildings is greater than that of large buildings in Hong Kong. Thus, ment for the land use and land cover map, Hong Kong Lands
the estimated PV potentials in residential buildings, which usually Department for building GIS data, and the Hong Kong Civil Engi-
contain smaller rooftop areas, may be underestimated. neering and Development Department for airborne LiDAR data.

(2). Off-terrain points in Digital Terrain Model References

Buildings may be misclassified as the ground features from the [1] EMSD, Hong Kong Energy End-use Data 14 (2014). Retrieved July 31, 2015
from, http://www.emsd.gov.hk/filemanager/en/content_762/HKEEUD2014.
decision tree classification process, which could be due to the
pdf.
height information of some off-terrain points, e.g. buildings, which [2] Intergovernmental Panel on Climate Change (IPCC), Climate Change 2007:
may have been misclassified in both DSM and DEM data. The Synthesis Report, Summary for Policymakers, 2007. Retrieved July 31, 2015
classifier might not be able to accurately distinguish between the from https://www.ipcc.ch/pdf/assessment-report/ar4/syr/ar4_syr_spm.pdf.
[3] EMSD, Study on the Potential Applications of Renewable Energy in Hong Kong
terrain points and off-terrain points for complex topographic fea- 9 (2002). Retrieved July 11, 2016 from, http://www.emsd.gov.hk/filemanager/
tures. The off-terrain point results in a zero value of object height en/content_299/stage1_report.pdf.
model and thus the node “height above 2.5 m” may misclassify the [4] EMSD, Hong Kong Energy End-use Data 2013 53 (2013). Retrieved July 31,
2015 from http://www.emsd.gov.hk/emsd/e_download/pee/HKEEUD2013.
buildings as ground features. pdf.
[5] J.M. Pearce, Photovoltaics e a path to sustainable futures, Futures 34 (7)
(3). Timeliness of the data (2002) 663e674.


[6] J. Hofierka, M. Súri, The solar radiation model for Open source GIS: imple-
mentation and applications, in: Proceedings of Open Source GIS - GRASS Users
Another error may occur from the GIS data. As the GIS data used Conference, 2002.
in this study are not as up-to-date as the airborne LiDAR data, some [7] A. Jochem, B. Ho € fle, M. Rutzinger, N. Pfeifer, Automatic roof plane detection
and analysis in airborne LiDAR point clouds for solar potential assessment,
newer buildings were not included in the calculation whereas some Sensors 9 (2009) 5241e5262.
demolished buildings might still exist in the GIS data but not in the

[8] M. Súri, A. Thomas, T.A. Huld, E.D. Dunlop, PV-GIS: a web-based solar radiation
LiDAR. This type of error can be corrected through the ground database for the calculation of PV potential in Europe, Int. J. Sustain. Energy 24
(2) (2005) 55e67.
filtering in decision tree classification. Due to the fact that the in-



[9] N. Luka
c, S. Seme, D. Zlaus, G. Stumberger, B. Zalik, Buildings roofs photo-
formation of new buildings has not been included in this study, the voltaic potential assessment based on LiDAR (Light Detection and Ranging)
PV potential may thus be underestimated. data, Energy 66 (2014) 598e609.
[10] H.T. Nguyen, J.M. Pearce, R. Harrap, G. Barber, The application of LiDAR to
assessment of rooftop solar photovoltaic deployment potential in a municipal
6. Conclusion district unit, Sensors 12 (2012) 4534e4558.
[11] H.T. Nguyen, J.M. Pearce, Incorporating shading losses in solar photovoltaic
Greenhouse gases accelerate the insulating effect in atmosphere potential assessment at the municipal scale, Sol. Energy 86 (2012)
1245e1260.
and global warming. Adopting solar photovoltaic technology is one [12] R. Dubayah, P.M. Rich, Topographic solar radiation models for GIS, Int. J.
way to use renewable resources to mitigate the climatic change. In Geogr. Inf. Syst. 9 (1995) 405e419.
Hong Kong, the solar energy contributes only 1.5% of total renew- [13] I. Me sza
ro
s, P. Mikla nek, J. Parajka, Solar energy income modelling in
mountainous areas, in: ERB and Northern European FRIEND Project 5 Con-
able energy usage in 2012. This study estimated the annual PV ference, Dem€ anovsk adolina, Slovakia, 2002, 1998.
potential of Hong Kong considering the influences of cloud covers,

[14] M. Súri, J. Hofierka, A new GIS-based solar radiation model and its application
and natural and artificial geomorphology in the solar radiation to photovoltaic assessments, Trans. GIS 8 (2) (2004) 175e190.
[15] P. Fu, P.M. Rich, Sol. Analyst 1.0 User Man. (2000). Retrieved July 31, 2015
estimation. The optimal areas to deploy PV panels were identified
from https://www.sciencebase.gov/catalog/item/5140acafe4b089809dbf56f5.
on unused building rooftops based on a set of decision tree criteria. [16] P.M. Rich, A Manual for Analysis of Hemispherical Canopy Photography, Los
The summarization of PV potential for the optimal rooftops is 2.66 Alamos National Laboratory Report, 1989. LAe11733eM.
[17] P.M. Rich, Characterizing plant canopies with hemispherical photography,
TWh.
Remote Sens. Rev. 5 (1990) 13e29.
This study was the first ever to derive a cloud cover probability [18] Census and Statistics Department, Summary results of population census,
map over the entire Hong Kong territories using geostationary 2011. Retrieved August 8, 2013 from http://www.census2011.gov.hk/pdf/
satellite images, which are vital for cloud detection and cloud summary-results.pdf.
[19] EMSD, Hong Kong Energy End-use Data 2015 13 (2015). Retrieved July 11,
movement tracking. Although the spatial resolution at 4 km may 2016 from, http://www.emsd.gov.hk/filemanager/en/content_762/
not be sufficient for detailed depictions over territories, it is still a HKEEUD2015.pdf.
promising measure with its high temporal resolution (e.g. 1 h res- [20] C.S. Lai, C.T. So, K.C. Ng, D. Jonas, The Territory e Wide Airborne Light
Detection and Ranging Survey for the Hong Kong Special Administrative Re-
olution). A non-prominent spatial distribution of cloud probability gion, 2012. Retrieved Apr 27 2016 from http://www.aamgroup.com/_
in Hong Kong was observed, thus only a single set of equations for literature_112830/LiDAR_Survey_for_the_Hong_Kong_Special_
estimating the diffuse proportion and the transmissivity in the Administrative_Region.
[21] Hong Kong Observatory, The Year’s Weather e 2012, 2012. Retrieved Apr 27
solar radiation model can be applied to the entire territory due to 2016 from http://www.hko.gov.hk/wxinfo/pastwx/ywx2012.htm.
the insignificant spatial variation of cloud covers. [22] E.F. Vertmote, D. Tanre , J.L. Deuze, M. Herman, J. Morcrette, Second simulation
Results of solar potentials in Hong Kong are promising. This of the satellite signal in the solar Spectrum, 6S: an overview, Trans. Geo-
science Remote Sens. 35 (3) (1997) 675e686.
study has developed a robust and accurate method for estimating [23] R.A.A. Nobrega, J.A. Quintanilha, C.G. O’Hara, A Noise-removal Approach for
the optimal PV potential on building rooftops in the entire city of LiDAR Intensity Images Using Anisotropic Diffusion Filtering to Preserve Ob-
Hong Kong. The developed methodology can be applied to other ject Shape Characteristics, 2007. Retrieved Apr 27 2016 from http://www.
asprs.org/wp-content/uploads/2011/01/0049.pdf.
places to study the application of PV modules in order to promote
[24] ESRI, An overview of the Solar Radiation tools (2016). Retrieved July 11, 2016
and encourage the use of solar energy. from, http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial-analyst-
toolbox/an-overview-of-the-solar-radiation-tools.htm.
Acknowledgments [25] H.X. Yang, L. Lu, The optimum tilt angles and orientations of PV claddings for
building-integrated photovoltaic (bipv) applications, J. Sol. Energy Eng. 129
(2007) 253e255.
This research was sponsored by the Public Policy Research [26] A. Goodrich, T. James, M. Woodhouse, Residential, Commercial, and Utility-
 M.S. Wong et al. / Renewable Energy 99 (2016) 325e335 335

scale Photovoltaic (PV) System Prices in the United States: Current Drivers Energy 24 (2) (2005) 55e67.
and Cost-reduction Opportunities, 2012. [29] SMA Solar Technology AG, Performance Ratio Quality Factor for the PV Plant,
[27] EMSD, Study on the Potential Applications of Renewable Energy in Hong Kong 2010. Retrieved July 31, 2015 from http://files.sma.de/dl/7680/Perfratio-
Stage 1 Study Report, Electrical and Mechanical Services Department, 2002. UEN100810.pdf.
Retrieved July 31, 2015 from http://www.emsd.gov.hk/emsd/e_download/ [30] ESRI, ArcGIS Help 10.1 Spat. Join. Anal. (2013). Retrieved July 31, 2015 from
wnew/stage1_report.pdf. http://resources.arcgis.com/en/help/main/10.1/index.html#//


[28] M. Súri, A. Thomas, T.A. Huld, E.D. Dunlop, PV-GIS: a web-based solar radia- 00080000000q000000.
tion database for the calculation of PV potential in Europe, Int. J. Sustain.