1 s2.0 S0360132321006636 Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Building and Environment 205 (2021) 108263

Contents lists available at ScienceDirect

Building and Environment


journal homepage: www.elsevier.com/locate/buildenv

Multi-objective optimization (MOO) for high-rise residential buildings’


layout centered on daylight, visual, and outdoor thermal metrics in China
ShanShan Wang a, Yun Kyu Yi b, *, NianXiong Liu a
a
School of Architecture, Tsinghua University, Haidian District, Beijing, 10084, China
b
School of Architecture, University of Illinois at Urbana-Champaign, 117 Temple Hoyne Buell Hall, 608 Lorado Taft Dr, Champaign, IL, 61820, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Nowadays building performance optimization is extended to urban planning Multi-Objective Optimization
Artificial neural network (ANN) (MOO). Most research focuses on the optimization of energy use and daylight performance of building design.
Multi-objective optimization (MOO) Buildings optimized for performance metrics rarely consider different performances together. Without inte­
Daylight factor (DF)
grating different building performance areas, the solution found from optimization will not be a balanced or
Sunlight hours
QuVue
trade-off one. This paper proposes a method to extend the use of optimization to cover multi-discipline areas that
Universal thermal climate index (UTCI) optimize visual comfort and outdoor thermal performances on the layout of high-rise residential buildings.
Daylight, sunlight hours, the sky view, and outdoor thermal comfort were the performance objectives. A
parametric building model was built to control the buildings’ layout and simulation tools were used to find the
performance of objectives. To accelerate the simulation process, an Artificial Neural Network (ANN) was applied
to the building simulation models to calculate the performance results rapidly.
ANN model had an average accuracy of 89.9% across all outcomes. The MOO method was conducted to find
integrated solutions to the building layouts on site. By ranking the optimized solutions based on five combined
performance targets, the top 10 out of 150 building layout options were identified, indicating an almost 21%
better performance than the baseline case.
Moreover, the top 30 out of 150 optimum cases performed better than the baseline. The study demonstrates
that the proposed MOO method that combines visual comfort and outdoor thermal measurements can improve
and contribute to a sustainable building layout design.

For optimum indoor performance, sunlight hours, the daylight pos­


sibility in a room, indoor air quality, and occupant thermal comfort have
1. Introduction
been frequent topics to insure high-performance buildings [1–3]. Out­
door environmental performance, such as sunlight exposure, solar ra­
In recent years, Multi-Objective Optimization (MOO) has been
diation, wind condition, and pedestrian-level thermal comfort, have
widely used in architecture and urban planning research and practice to
been common objectives in many studies [4,5].
ensure a building’s performance. In architecture design, the building
Experts often consider the main requirements of performances while
layout design, especially in residential buildings, is becoming in high
ignoring other factors. Building design is a multi-criteria and difficult
demand to design high performance buildings in China. These designs
task for architects to accomplish in the early design stage. Due to recent
must satisfy multiple design criteria in terms of sustainability.
development in computation and analysis, this problem has become
Optimization problems in high-rise residential buildings, however,
streamlined. Computing simulation tools can be applied to integrate
often have contradictory objectives. Energy use and light planning are
different objectives into one solution.
often addressed in building layouts optimization and less care is given to
This paper proposes MOO in high-rise residential building layout
other areas [1,2]. Thermal conditions and visual comforts are also the
design which considers three different objectives including light, visual
main performance objectives imposed in sustainable design re­
comfort, and thermal conditions. The ANN method is investigated to
quirements, but it is rare to find research considering these optimization
establish an integrated objective simulation prediction. The competing
problems at the same time. The main reason for isolated optimization
objectives will result in an optimized building layout form by
can be found in experts’ special knowledge in the building domain.

* Corresponding author.
E-mail addresses: wangss17@mails.tsinghua.edu.cn (S. Wang), ykyi@illinois.edu (Y.K. Yi), phlnx@tsinghua.edu.cn (N. Liu).

https://doi.org/10.1016/j.buildenv.2021.108263
Received 11 May 2021; Received in revised form 11 August 2021; Accepted 13 August 2021
Available online 18 August 2021
0360-1323/© 2021 Elsevier Ltd. All rights reserved.
S. Wang et al. Building and Environment 205 (2021) 108263

Nomenclature MU_CD quantity of measuring points of UTCI between − 15 and


− 10
DF Daylight factor s number of UTCI measurement points of the site
WinH Sunlight hours of windows Rating the order of performance values
SiteH Sunlight hours of the site NDF weighting of DF
F(x) Multi objective function NQuVue weighting of QuVue
f1(x) DF objective function NWinH weighting of WinH
f2(x) WinH objective function NWinH weighting of SiteH
f3(x) QuVue objective function NSiteH weighting of DF
f4(x) SiteH objective function NUTCI weighting of UTCI
f5(x) UTCI objective function DFmax maximum value in optimal DF outcome
DFm Average value of target buildings of Daylight factor DFi each value in optimal DF outcome
Ddaylight factor DF of i measuring point DFmin minimum value in optimal DF outcome
n Total number of DF measurement points QuVuemax maximum value in optimal QuVue outcome
QuVuem average value of target buildings of QuVue QuVuei each value in optimal QuVue outcome
QQuVue QuVue value of i measuring point QuVuemin minimum value in optimal QuVue outcome
p Total number of QuVue measurement points WinHmax maximum value in optimal WinH outcome
WinHm average value of target buildings of window sunlight hour WinHi each value in optimal WinH outcome
Wsunlight hours WinH of each measuring point WinHmin minimum value in optimal WinH outcome
q Total number of WinH measurement points SiteHmax maximum value in optimal SiteH outcome
SiteHm average ratio value of target site of SiteH SiteHi each value in optimal SiteH outcome
MS_2H measuring points of WinH>=2 h SiteHmin minimum value in optimal SiteH outcome
r number of SiteH measurement points of the site UTCImin minimum value in optimal UTCI outcome
UTCIm ratio of average value of target site of UTCI UTCImax maximum value in optimal UTCI outcome
MU_HD quantity of measuring points of UTCI between 31 and 32 UTCIi each value in optimal UTCI outcome

maximizing the use of daylight in buildings, sunlight exposure indoors 2.2. Building optimization
and outdoors, a sky view of buildings, and minimizing the outdoor
thermal discomfort. Design variables consist of the building’s spatial Performance-based studies have been conducted on building envi­
positions and heights. The following were used for this study: a para­ ronmental optimization. Single object optimization, most commonly
metric design tool to control geometry, specific simulation tools to using a Genetic Algorithm (GA) model, has been used to solve one
analyze the different performances, ANN models to predict different performance optimization problem. In this type of optimization, only
metrics, and use of the MOO method to optimize the performance of the one objective function (e.g., energy consumption) was used to evaluate
design. the given variables [16–18]. Li and others [19] summarized the opti­
mizations used in the field of building environment [20–22]. Applying
2. Background the single object optimization approach, the optimization process can
save time and help explore sufficient optimal solutions. MOO has been
2.1. High-rise buildings of computational design more commonly used in recent studies to deal with several conflicted
performance objectives [23–30]. Yu and others focused on building
It is common to find performance-based studies on high-rise build­ energy consumption and occupants as objectives for the optimization for
ings [6]. In a high-rise building’s optimization problems such as light, a typical case building in China [31]. Building lighting, cooling, and
visual comfort [7], thermal conditions [8] as well as energy consump­ heating energy demand were considered as optimization objects in a
tion were the major criteria in recent studies. high-rise building design [32]. Fialho and others used building materials
For example, windows and shades were considered as important and orientations as variables for the objective function to optimize
factors for lighting and visual comfort in high-rise buildings [9]. Jing construction costs and energy use [33]. Building layout design is an
Zhao and Yahui Du optimized indoor thermal comfort and energy con­ integrated problem that covers many design conditions and constraints,
sumption in a high-rise office building’s façade design. Pareto optimal which must be balanced, especially when there are conflicts. Thus, the
solutions provide guidance and suggestion for designers based on pref­ MOO method can be used to solve a design problem.
erences in the early design of buildings [10]. Berk Ekici and his team
presented a multi-zone optimization methodology with ANN methods to 2.3. Performance-based design
optimize daylight performance metrics to obtain parameters of high-rise
building facades scenarios [11–13]. Sewon Lee et al. conducted an in­ Most of the publications presented indoor environment optimization
door daylight performance optimization in high-rise apartment building based on an evolution algorithms. Konis and others [34] proposed
layout in Korea [14]. Meseret T. Kahsay and others developed a pro­ daylight and energy use of optimal solutions for buildings of a given plan
cedure for simulation-based optimization of window size and configu­ ratio, a total floor area, and the number of floors. The variables were the
ration for a high-rise building to minimize its energy heating and cooling building form, window-to-wall ratio, and exterior shading devices. Yao
consumption [15]. [35] used the ‘energy performance difference between housing units’
However, those studies mostly conducted one or two aspects of (EDH) to evaluate the energy demand to manipulate the thermal ma­
building performance. In performance-based design, it is important to terials of windows and walls and internal shadings. Tronchin and others
consider and integrate different aspects of performance measures to be [36] discussed a multi-scale building optimization method in terms of
high-performance design. the building materials and components with their thermo-physical pa­
rameters to optimize the building energy demand. Echenagucia and
others [37] performed through an implementation of the NSGA-II [38]

2
S. Wang et al. Building and Environment 205 (2021) 108263

algorithm to optimize building energy use, in which the building residential heating, cooling, humidifying, and dehumidifying system
window-to-wall ratio and the thickness of walls were options. Jav­ control have been proposed [60]. A surrogate-assisted evolutionary
anroodi and Giyasov attempted to create a multi-objective optimization optimization algorithm was used to model microclimate sensitive urban
framework based on a genetic algorithm. A form generation system was design problems including wind velocity ratio and gross profit, wind
proposed to find the best type of high-rise office building to meet the velocity Gini index in a residential building planning [61]. Although
minimum cooling, heating demands and thermal comfort [39,40]. Other these solutions integrated the ANN method into multi-objective genetic
studies described a larger scale such as a site design or urban design algorithms, the domain of building performance is limited and is rarely a
performance optimization based on energy use and the outdoor envi­ crossover to other domains such as lighting performance and outdoor
ronment. Early work included Andersen and Sattrup’s research [41] on thermal boundary.
urban canyons’ form and building density optimization based on energy
use, daylight, and solar radiation using simulation methods. Sarralde 2.5. Findings
et al. [42] focused on solar energy and urban morphology in London.
Gan et al. [43] developed an optimizing method using a genetic algo­ To overcome the current limitation of using optimization for a spe­
rithm to minimize the energy use of the layout of high-rise housing units. cific performance domain and expanding the application of the inte­
Yasa and Vildan [44] formulated and solved a particular case of opti­ gration simulation-based ANN into MOO to cover multidiscipline areas,
mization for a building with a courtyard to meet the climatic comfort in this paper proposes the following method using ANN-based MOO in both
different regions. Xu et al. [45] found solutions on different forms of visual comfort and outdoor thermal performance domains in the early
urban layouts to optimize the outdoor thermal performance of Universal design stage.
Thermal Climate Index (UTCI) in cold region of China. Fedorczak-Cisak A multi-objective optimization (MOO) combined with an Artificial
et al. [46] carried out a multi-criteria analysis to optimize the CO2 Neural Network (ANN) was used at the early design stage to decide a
emission, thermal comfort, energy consumption, cost, and evaluation of building layout of the residential buildings. This approach consisted of
the building layout. However, most of the research mentioned above an integrated design process that not only satisfied various building
using optimization only focused on either indoor or outdoor aspects and codes but also shared possible design options that followed local
did not include an impact of a combination of the two. building codes.
Most studies on building design and optimization procedures fol­ The main contribution of this paper was extending the use of MOO
lowed the local building codes or other assessment requirements. For unlike its current usage in a specific area. Five objectives are integrated
instance, the building energy performance regulations were discussed in from different areas to find a solution. With support from the recent
the building design application in climate zoning in Spain [47]. The new development of computational simulations in different areas, this study
energy performance codes regarding building energy, cost, and con­ was able to demonstrate that integrating MOO in different areas can be
struction in Greece were validated in various cases [48]. These studies used to guide an architect to find high-performance solutions in the early
found that some of the building codes have limitations to insure actual stage of design without a help from an expert.
building performance. This means that building regulations have room
to adopt a more realistic method to evaluate buildings closer to actual 3. Methodology
conditions [49–52].
Fig. 1 shows the overall process of the method to find design solu­
2.4. Artificial intelligence and machine learning tions that satisfy both indoor visual comfort and outdoor performance
metrics. The process included four main steps.
Some efforts were made to predict simulation results using an arti­ The first step was to prepare the Parametric Building Model (PBM)
ficial intelligence. The methods of performance prediction powered by and the Dynamic Parameter Control System (DPCS) of the buildings’
artificial intelligence are now widely used in the field of building design. layouts. Typically, in optimization when a complex shape is introduced
Artificial intelligence is a more efficient and robust method for esti­ to the optimization, the number of geometric variables increases, mak­
mating building performance metric. For instance, Vincent J.L. Gan and ing it difficult to find the optimum. To overcome this challenge, the
others proposed a natural ventilation prediction procedure based on study used a Dynamic Parameter Control System (DPCS) of the build­
simulation-based results using ANN models [8]. Based on artificial in­ ings’ layout, one that introduced a hierarchic system to reduce geo­
telligence, especially integration of machine learning and optimization metric variables. In this approach, a syntax-based system was developed
method, the relationship between input parameters in building layout to use less geometric variables to control the complex geometry.
design and output of wide domain of performance metrics can be Therefore, less points were needed to transition between the geometry
established to improve sustainable design. shifts compared to controlling each point separately.
The prediction method of artificial intelligence is fast and flexible, After the dynamic building geometry layouts were created, the sec­
which helps an user make a decision easier [53–55]. Waibel discussed ond step was to use physics equation-based models to simulate indoor,
how black-box optimization (BBO) algorithms affect building energy and outdoor environmental conditions. The five performance measure­
optimization problems [56]. Magnier and Haghighat investigated sur­ ments were used for this paper. For indoor performance evaluation, the
rogate modeling and genetic algorithms to optimize operating energy Daylight Factor (DF), sunlight hours at a window (WinH), and an open
and thermal comfort [57]. Surrogate models were selected for the sky view at the window (QuVue) were calculated with advanced simu­
development of an objective optimization process. Artificial Neural lation tools. In terms of outdoor performance measurement, the sunlight
Networks (ANN) enhanced the ability of building performance optimi­ hours (SiteH) and Universal Thermal Climate Index (UTCI) at the site
zation problems to be solved. Therefore, the ANN model was introduced ground level were calculated.
to predict the five performances in this study. The coefficient of deter­ Because simulating the multi-simulations take significant time and
mination was used as the index of accuracy performance. computing power, it is necessary to find an approach to reduce the
Some ANN-based optimization methods are applied in residential simulation time. For that reason, this paper introduced the Artificial
building design. Magnier et al. presented a simulation-based ANN Neural Network (ANN). After the selected number of performance
method combined with NSGA-II to optimize thermal comfort and energy values was gained from the simulation, the third step was modelling the
consumption optimization in a residential building [58]. Naji et al. artificial neural network (ANN). The random sample data gained from
performed ANN to assess the prediction accuracy in estimating building the previous step was linked to the separated ANN models. The DPCS’s
energy consumption [59]. building geometry control variables were set as input for ANN training,
Recently, ANN-based thermal estimation methods to optimize and the previous five simulation results were used as target data for the

3
S. Wang et al. Building and Environment 205 (2021) 108263

performance results were used to build the ANN models that can
enhance the conventional simulation approaches and were used in the
evaluation optimization process. The MOO procedure continued until
the goals were meet. The detail process of the methodology was dis­
cussed in the following sections.

3.1. Dynamic Parameter Control System (DPCS)

The test case buildings were in Beijing, China (Fig. 2) in a cold


climate area. There were twelve buildings on the site. The total site area
was 143,358 m2. Among the twelve buildings, 10 residential buildings
(No.1–6, No. 8–11) were used for the test, but 2 commercial buildings
(No. 7 and No. 12) were excluded from calculation (Fig. 2). Because the
site was located in open field, the urban context around the site was not
consider in the test.
In the parametric building geometry models, the geometry control
system such as each building’s shape, height, location, and orientation
were defined. Fig. 3 explains how the buildings could be arranged
around the site by the agent point defined by the DPSC system. Coor­
dinate x and y referred to the original position of the building, and x’, y’
referred to the position that the building could move (Fig. 3 (a)).
Vertically, z shows original height and z’ indicates the building height
change (Fig. 3 (b)). When the agent point’s (x, y, z) value was increased,
the building moved to the northeast and the building height increased.
In the same way, when the value was decreased, the building moved to
the southwest and the height reduced.
Based on the DPCS definition, the building could move along the X
and Y axis and the height of the buildings could change. To make sure
buildings were not going beyond the site boundary, the DPCS definition
was set when the building reached the edge of the site, thus it cannot
move beyond the boundary. Moreover, the buildings were setup to avoid
overlapping with the other buildings.

3.2. Performance criteria

The following measures were identified as performance indicators to


evaluate visual comforts and outdoor thermal performance of the
buildings’ layout. Based on the Chinese national building design codes
[62] and previous studies [63], five performance targets were selected
for this study, including the Daylight Factor (DF), Window Sunlight
Hours (WinH), Site Sunlight Hours (SiteH), Universal Thermal Climate
Index (UTCI), and Sky view ratio (QuVue). For indoor daylight perfor­
mances, DF and WinH were used as they are part of the Chinese building
codes. WinH is a cumulative sunlight exposure duration of the windows
for each unit on the first floor from 8 a.m. to 4 p.m. on January 21st
according to the Chinese building code.
For visual performance of the interior space, QuVue was conducted
to measure the sky view ratio from each window. For the outdoor space,
Fig. 1. Overall process. the thermal and daylight performance were two of the most important
elements to create a healthy space. For outdoor daylight performance,
ANN training. Once the ANN models were trained, they can be used to this study used the measurement from the Chinese’s building code,
predict the performance of different layouts and shapes of buildings which was the sunlight hours on the site (SiteH). To measure the thermal
without running five physic-based simulation tools. performance of the outdoor space, this study employed UTCI that
After the five ANN models were trained, the final step was optimi­ measures an integrated measurement of temperature, relative humidity,
zation. The trained ANN models from the former steps were used to solar radiation, and wind speed.
calculate the objective functions in the Multi-Objective Optimization Table 1 shows the threshold and goals of performance metrics ac­
(MOO). If the building layouts met the defined multi-objective goal, the cording to research review and building codes. The following sections
possible solutions were found. If not, the next generation of building discuss these measurements in detail. Although the study considered the
layouts was generated, and the ANN model calculated the performances indoor and outdoor environmental indicators, there were not many in-
and the MOO process to find the altered building layouts to satisfy the depth studies available on the correlation between the indoor and out­
goal. The searching process continued until the building layouts’ per­ door indicators. This requires in-depth theoretical research on how the
formance fulfills the multi-objective goal. outdoor indicator impacts the indoor indicator and vice versa, which is
In summary, based on the PBM model, the variables with DPCS were somewhat different from the study’s main purpose. For this reason, the
created to generate dynamic building layouts. Furthermore, a selected study used existing measures without considering the impact of the in­
number of simulations were conducted to calculate performance. These door and outdoor measures.

4
S. Wang et al. Building and Environment 205 (2021) 108263

Fig. 2. Overall building layouts on the test site.

Fig. 3. Explanation for the building variables.

5
S. Wang et al. Building and Environment 205 (2021) 108263

Table 1 measurements called QuVue was used in this study.


Threshold and goals of performance criteria. QuVue is the sky view ratio calculation [63] and the index for
Performance Threshold and goals measuring visible sky ratios that overcome the major hurdles of current
metrics methods that do not consider the downward view and calculates the
DF >2% actual three-dimensional open sky view area. QuVue is a plug-in for
WinH >2hr accumulatively at 8:00 to 16:00 on January 21st Grasshopper [71,72]. Its value was measured from each window’s
QuVue 0–50%, 50% means receive the most skyview which is perfect center point.
SiteH >2hr accumulatively at 8:00 to 16:00 on January 21st
UTCI >-10 and < 31
3.2.3. Window sunlight hours
Similar to DF, the sunlight provision for a special period of time is
3.2.1. Daylight factor (DF) another metric related to daylighting [65]. When direct sunlight reaches
DF is the ratio of illumination available indoors relative to the illu­ windows, it can reduce the thermal load during the winter season in the
mination outdoors at the same time under overcast skies [64]. As the DF northern part of China. Table 3 shows the requirement of the minimum
value can be calculated fast, it can be used in the early design process, sunlight hours through windows in Beijing. To this extent, Window
thus allowing architects and developers to make informed decisions. The sunlight hours (WinH) was introduced on the basis of the Chinese’s
daylight factor is influenced by shading from the adjacent buildings and design standard for energy efficiency of residential buildings in severe
the window-to-floor ratio. In our study, the building layout can change cold and cold zones (JGJ 26–2018) [73]. The cumulative sunlight
the distance between buildings, which impacts the building’s daylight exposure duration of the windows for each unit on the first floor from 8
factor (DF). a.m. to 4 p.m. on January 21st can be no more than 2 h according to the
Several building codes adopted DF to ensure that the building re­ Chinese code. The measuring points for sunlight hours were same as
ceives minimum daylight. In European Standards, EN 12464–1:2011 QuVue, which was center point of each window.
and EN 15193-1 had details of how a building should meet the daylight
requirements [65]. ISO2014a applied to calculate methods for daylight 3.2.4. Site sunlight hours
in both existing buildings and the design of new and renovated buildings Unlike indoor sunlight hours, this section focuses on the sunlight
[66]. duration on the site. Sunlight exposure not only influences indoor
Furthermore, building sustainability assessment systems made rec­ spaces, but it also has an important impact on the outdoor environment.
ommendations for daylight in their evaluation schemes. The Building In cold climate zones, having access to enough direct sunlight for pe­
Research Establishment Environmental Assessment Method (BRERAM) destrians is greatly beneficial, especially in the winter for children and
stated that at least 80% of the floor area in occupied spaces has an elders. Some countries add sunlight in open outdoor spaces into their
average DF of 2% or more [67]. Deutsche Gesellschaft für Nachhaltiges building codes. For example, in the UK at least half of the amenity area
Bauen (DGNB) showed the classification of DF in the room as very good should receive at least 2 h of sunlight on March 21st [74]. According to
(>3%), medium (>2%), slight (>1%) and none (<1%) levels for new Item 7.04 of the standard for urban residential area planning and design
construction residential buildings [68]. In China, based on the building [75], the direct sunlight hours of a site should be considered especially
code [52], the DF should be more than 2% in an annual average value. during the winter. Furthermore, the building sustainability assessment
Other advanced measurements can be used instead of DF, such as systems like BREEAM, and DGNB contained sunlight assessment in
Useful Daylight Illuminance (UDI), Daylight Autonomy (DA), or a external building environments [67,68].
measurement specifically developed for residential building [69]. The Therefore, the outdoor sunlight hour (SiteH) was included as a part
main reason DF was selected for this study is its simplicity that ensures of outdoor performance measurements. In SiteH, the measuring points
minimum daylight under an overcast condition and can co-simulate were calculated if points received more than 2 h of direct sunlight be­
with different measurements more easily. tween 8:00 a.m. to 4:00 p.m. on the coldest day (January 21st) [73]. The
Table 2 shows the materials setups for DF. The measurement of the grid size for the calculation was 2.5 m × 2.5 m and the measuring height
surface was at 0.75 m above the floor level, and the measuring grid size was 1.5 m above the ground.
was 1 × 1m. The surface reflectance for the test were set up as below and
used the same setup for whole test. 3.2.5. Universal Thermal Climate Index (UTCI)
For outdoor conditions, the wind speed, wind pressure, and heat
3.2.2. QuVue (sky view ratio) island are included in the Chinese building codes [76]. Specifically,
The view is an important measurement for the privacy, safety, wind speed is regulated from 1.5 m/s to 5 m/s at the pedestrian level.
health, and well-being of residents. In Chinese’s building regulations, However, other physical environments such as temperature and hu­
the minimum distance between each building is required in the privacy midity of the outdoor space are not included in the code, which are also
and fire control code [70]. Without the proper view towards the open important for outdoor thermal comfort [77].
space, the residents’ stress level can be increased. Although current The Universal Thermal Climate Index (UTCI), consisting of temper­
Chinese building codes do not regulate the open space view on a resi­ ature, relative humidity, solar radiation, and wind speed, is regarded as
dential building, it is important for health and well-being to consider the a holistic index. Some papers found that among different thermal
view when laying out buildings. Therefore, one of the sky view comfort indices, UTCI showed a more reliable result than PET

Table 3
Table 2 Sunlight standard for residential buildings (JGJ26-2018).
Daylight Factor boundary condition setup. Location Beijing, China
Boundary condition settings Boundary condition Building Climate Area Cold Zone
Wall reflectance 0.6 Resident Population Size ≥500,000 <500,000
Floor reflectance 0.4 Date to measure January 21
Ceiling reflectance 0.7 Required Sunlight Hours(hrs) ≥2 ≥3
Window reflectance 0.6 Measurement Hours(local True Sun Time) 8 : 00–16 : 00
Ground floor 0.7 Measurement location Bottom of windowsill
Mesh grid size 1 × 1m
Note: the bottom windowsill refers to the external wall position 0.9 m high from
Test surface distance from floor 0.75 m
the floor of the room.

6
S. Wang et al. Building and Environment 205 (2021) 108263

(Physiological Equivalent Temperature) [78] and a different scale of Table 5


UTCI is available for many climate types [79]. Input parameters of ANN models.
UTCI is an index that reflects the comfort of the external thermal Input Variables Quantity Range
environment. The layout of residential buildings has an impact on the
x1, x2, x3, …,x12 12 [-10, 20]
temperature, humidity, and wind speed of the external space, which y1, y2, y3, …, y12 12 [-10, 20]
influences outdoor activities. Analyzing the UTCI value of the typical z1, z2, z3, …, z12 12 [-34.1, 99.2]
season at pedestrian height can obtain the building layout scheme to
improve thermal comfort. Therefore, UTCI was used to find the outdoor
thermal performance. 3.4. Multiple objective optimization (MOO)
In a previous study [80], the numerical value of UTCI was defined
into the thermal classification and it was used for this study (Table 4). In This section discussed integrating the five trained ANN models with
terms of the UTCI measurement, both the coldest and hottest days’ MOO to find solutions for the building layout design. This study used the
average values were used as a measurement. Non-Dominated Sorting Genetic Algorithm II (NSGA-II) in MATLAB [83]
The measuring grid size was set at 1.5 × 1.5 m based on the social as the optimization algorithm. NSGA-II is a controlled elitist evolu­
comfort distance at the height of 1.5 m from the ground. The study used tionary algorithm that increases the diversity of the population solutions
Eddy3d for the calculation which is a plugin for Grasshopper. It is a CFD- that is limited by a set of fitness functions. An evolutional algorithm was
based UTCI calculation, which calculates the wind factors and MRT used with surrogates for evolutionary operators such as selection, ge­
(Mean Radiant Temperature) [81]. Wind factor calculated based on the netic crossover, and genetic mutation. By the ordering of the Pareto
CFD simulation result of velocity and direction. Each hour wind factor is dominance, the population was ranked in a hierarchy of
scaled up/down based given probing the height from the ground. sub-populations. To form the front of non-dominated solutions, the
OpenFOAM was used as an engine to simulate wind conditions [82]. To similarity between members of each sub-group was evaluated. The
calculate the MRT, Eddy3d uses a method that assumes the building application of the resulting groups and similarity measures is to promote
surface temperature equals the ambient temperature. Eddy3D runs a sky a diversity of a Pareto front [84]. More information on NSGA-II algo­
view factor analysis by Radiance to account for direct solar gain [81]. rithm used for the test can be found in MATLAB.
We calculated the UTCI using probing points on the 1.5 m height of The multi-objective functions for the test were made up of five
the pedestrian level. To simplify the optimization, the average values of objective functions (DF (f1 ), WinH (f2 ), QuVue (f3 ), SiteH (f4 ), and UTCI
the total amount of spatial grid results was introduced. Because the (f5 )). Making design decisions and finding solutions is a comprehensive
special grid values were averaged, it has limitations to capture certain procedure that includes all aspects. To simplify this study, we considered
building layout spatial-characteristic problems. only five metrics based on the local code. Other impact factors such as
energy consumption and cost could be also considered for a better
3.3. Artificial neural network (ANN) design.
The solution for the optimization was to find a better solution than
As simulating all five simulations simultaneously requires significant the initial building simulation results. For an optimal building layout,
computing time and power, it is necessary to find a method to reduce DF, QuVue, WinH, and SiteH must be maximized. As the average values
simulation time. The study utilized an ANN algorithm described as of all spatial grid results were introduced, the UTCI value of the building
black–box modeling to enhance calculating efficiency. layout plan was simplified to a single value. The goal of the UTCI is bring
Based on the definition of DPCS, the 12 buildings have a total of 36 the overall thermal index value down as low as possible to minimize the
geometric controlling parameters which are independent parameters. uncomfortable outdoor special area. Therefore, the goal of UTCI was set
Each building has a set of agent values shown as x, y, and z. In ANN to be minimized.
modeling, 36 input variables were used as input variables, that is, the x, However, to maintain consistency between the objective functions,
y, and z coordinates of the 12 buildings’ location. the study converted the maximization to minimization and the objective
Each building can move in both directions along the X axis and Y function can be written as below.
axis, and it can move between negative 10 m to positive 20 m magni­ minF(x) = [ − f1 (x), − f2 (x), − f3 (x), − f4 (x), f5 (x)] (1)
tudes from the initial position. The building height range was between
34.1 m and 99.2 m. To maintain a certain range of Floor area ratio, the To build the MOO model, the objective function for the five perfor­
building’s total area was constrained within ±10% of the initial case mance measurements should be determined first. For the DF objective
(Table 5). function (f1(x)), the sum of the measurement value of each point of the
It is necessary to mention that there are many parameters that in­ DF value (Ddaylight factor) was divided by the number of measuring points
fluence daylight measures and outdoor thermal conditions. For instance, (n), as follows,
the window-to-wall ratio and floor plan are important variables in ∑n
Ddaylight factor
daylight metrics. In this study, we focused on the building layout plan DFm = i=1 (2)
n
and its environment performance; however, other parameters like the
For the QuVue, which is the same as the DF, the total measurement
window-to-wall ratio can also be considered.
points of the QuVue value (QQuVue) were divided by the number of
measuring points (p).
∑p
QQuVue
QuVuem = i=1 (3)
p
Table 4
Thermal perceptions for different UTCI classification. For window sunlight hours, the total measurement points of the
WinH value (WinHsunlight hours) were divided by the number of
Thermal perception UTCI
measuring points (q),
Cold − 27 to − 13 ∑q
Cool − 13 to 0 W sunlighthours
Slightly cool 0 to +9
WinHm = i=1 (4)
q
Comfortable +9 to +26
Warm +26 to +32 For site sunlight hours, the points that received direct sunlight for
Hot +32 to +38 more or equal to 2 h (MS_2H) in January 21st were counted and divided

7
S. Wang et al. Building and Environment 205 (2021) 108263

by the total measuring points (r). Table 7


Information on the initial window size of the buildings.
MS2H
SiteHm = (5) Window-to-Wall Ratio (%)
r
Building No. East West South North
For the UTCI in the summer period, the points where the UTCI values
were between 31 and 32 were considered, (MU_HD) and the points where 1 10 10 38 22
the UTCI value were from − 15 to − 10 in winter were also added 2 12 10 38 21
3 21 23 20 8
(MU_CD). These two values were added together and then divided by two 4–6 9 9 41 30
times the total measuring points (s). 8 20 21 28 9
9–11 9 9 41 30
MUHD + MUCD
UTCIm = (6)
2s

Table 8
4. Application of the proposed method
Simulation settings.

4.1. Test case project description Metrics Simulation Constant Items Values
Tools

The test project had no residential units on the ground floor, and it DF Ladybugs Location and Beijing
was excluded from measurements. Tables 6 and 7 demonstrate the (ver 0.061) weather file
(Radiance) Grid size 1×1m
existing buildings conditions as a baseline for optimization. Among the Distance from 0.75 m
twelve buildings on the site, building no. 8 was the highest and the base surface
average number of building floors was 22.8. Window-to-wall ratios Sky Uniform CIE sky
varied depending on the building’s orientation. Most buildings showing Radiance -ps 8, -pt 0.15, -pj 0.6, -ds 0.5, -dt
parameters 0.5, -dc 0.25, -dp 64, -ab 0, -aa
a southern orientation had the highest window-to-wall ratio; however,
0.15, -ar 32, -as 32, -lr 4, -lw 0.05
buildings no. 3 and no.8 showing a western orientation had less- window width 1.2/1
window-to wall ratios. to height ratio
WinH Ladybug (ver Date and time Jan 21 8:00–16:00
0.061) Simulation 1
4.2. Existing Building’s performance times step per
hour
Grid size 3×3m
In this section, five performance criteria were simulated with the SiteH Ladybug (ver Site grid size 2.5 × 2.5 m
current building layout to understand the current conditions and to 0.061) (SiteH)
collect a dataset to build the ANN models. The details of five stimulation Date and time Jan 21 8:00–16:00
Simulation 1
tools with main settings are show in Table 8.
times step per
For the DF calculation, this study used Honeybee which uses Radi­ hour
ance and Daysim as engines to calculate the DF. The total of 2888 points UTCI EDDy3D Wind direction 0, 45, 90, 135, 180, 225, 270,
were used for the measurement. (blueCFD) 315◦
With the base test case (as it is), the average DF value of the whole Boundary type cylindrical domain
Boundary Inner 400 m
buildings’ first floor was 3.68%, which met the requirement of the rectangle
building codes. Among the 10 target buildings, the best two buildings for Boundary Outer 1000 m
the DF measure were buildings no. 1 and no. 2 as they are located at the radius
south of the site, thus receiving the most daylight. Building no.3 showed Boundary 250 m
Height
the least DF among the other buildings; this is mainly related to the
Mesh size 357,568
window area, and compared to other buildings, it had a less window-to- Mesh type OpenFOAM’s blockMesh and
wall ratio among the three orientations (Table 9). snappyHexMesh
QuVue is a plugin for Grasshopper and used to calculate the Sky view CFD Turbulence kOmegaSST
ratio. The measurement points were based on the central position of model
Pressure model SIMPLE (Semi Implicit Method
each window of every unit (Fig. 4) and a total of 2268 measuring points
for Pressure Linked Equations)
were calculated. Sky View QuVue Test surface South/East side windows
Figs. 5 and 6 show two examples of the test results from the QuVue. Ratio Measuring point Center of each window
The distribution of the QuVue of no. 5 building is displayed as a contour

Table 6 Table 9
Information on the buildings. DF result of each target building of the base case in the first floors.
Building No. Function Height(m) No. of Floors Building No. Mean (%) Minimum (%) Maximum (%)
1 Housing 65.1 20 1 5.97 1.90 18.05
2 Housing 65.1 20 2 6.15 1.41 17.91
3 Housing 68.2 21 3 2.48 0.04 12.97
4 Housing 71.3 22 4 2.88 0.34 11.76
5 Housing 55.8 17 5 4.71 0.56 36.36
6 Housing 74.4 23 6 2.95 0.47 30.35
7 Commercial 80.6 25 8 2.49 0.05 14.93
8 Housing 86.8 27 9 2.52 0.28 12.99
9 Housing 80.6 25 10 3.65 0.29 15.67
10 Housing 77.5 24 11 2.98 0.30 13.11
11 Housing 80.6 25 Average 3.68 0.56 18.41
12 Commercial 80.6 25
Total Avg. – 73.8 22.8

8
S. Wang et al. Building and Environment 205 (2021) 108263

Fig. 4. Measuring points for QuVue.

Fig. 5. QuVue result of building no. 5.

diagram of the south-facing façade in Fig. 5 (b). Each colored square more sky view on the right side. The worst QuVue existed on the left side
with a value refers to the tested windows. As the building was located on of the first floor.
the central axis of the site, the QuVue was distributed in a central The simulation result of the QuVue is shown in Table 10. Buildings
axisymmetric status, and with the higher floors, the value became no. 1 and 2 had the best sky view ratio (50%) because they were located
higher. The highest value was the one at the center of the top floor. The on the south side of the site and nothing interrupts the sky view. One
smallest value was on the edge of the first floor. It is interesting to thing to be aware of is that the measuring point was located in the
observe that two of the buildings (no. 1 and 2) that block the view were vertical surface, which meant only half of the hemisphere can be seen.
close to the edge of the test building (Fig. 5 (a)), though the measuring This means the maximum sky view it could achieve was 50%. Building
points were on the same floor, the units closer to the building side no. 3 has a 44% QuVue, which is the second top result as its east-facing
showed a lower value than the center units. windows provided more sky views without obstruction. However, the
Fig. 6 shows the QuVue of the south façade of building no.6. The QuVue of buildings no. 4 and.11 showed the worst results, that is 27%
right part of the QuVue value of the facade was higher than the left part and 26% respectively. One of the possible reasons for this was that the
because the test building was located on the right side of the site, giving building distance between no. 4 and the test buildings no. 1, 3 and 5 was

9
S. Wang et al. Building and Environment 205 (2021) 108263

Fig. 6. The QuVue result of building no. 6.

Table 10
Summary of QuVue test results.
Building No. 1 2 3 4 5 6 8 9 10 11 Total

No. of points 228 228 100 252 192 264 104 336 276 288 2268
Average (%) 50 50 44 27 35 32 35 34 29 26 36

*(Building 7 and 12 were commercial buildings that was not include in the calculation).

Fig. 7. Contour of the WinH value of the test buildings.

10
S. Wang et al. Building and Environment 205 (2021) 108263

relatively close as well as the height of those test buildings was simplicity, calculation speed, and learning from the limited data sets
comparatively high and blocked the sky view of building no. 4. [87]. Moreover, it is mostly used in the field of building energy use [88].
Window sunlight hours (WinH) used the same measuring point as Therefore, this study adapted ANN as a tool for the test.
QuVue. In Fig. 7, each window was selected. As shown in Fig. 7 and ANN models were built in MATLAB for predicting the five perfor­
Table 11, the best average of WinH was in buildings no.1 and 2 at 8 h mance results. The paper used plug-in called “nntool” to create, train,
because there were no buildings blocking the south orientation which and simulate neural network for the test. The inputs dataset for ANN
leads to enough sunlight exposure. Building no. 5 had the second-best trainings was the Dynamic Parameter Control System (DPCS) based on
average sunlight hours at 4.5 h; however, the worst result was 2.3 h in the geometry variables. The outputs were performance measurements
building no. 4, due to the shadow of buildings no.1, no. 3 and no. 5. In (e.g., DF, WinH, SiteH, QuVue and UTCI).
the case design, the minimum WinH average value (in total) of the first For training the output data set, the following simulation measures
floors was 3.3 h, and the value ranged from 2 to 8 h which met the were used for training the data set for five different ANN models. In the
requirement of the building codes. ANN model for DF (DF_ANN), the average value of measuring points in
There were 4454 points used to calculate the outdoor sunlight hours the top floor of the DF of 10 buildings were used as a training output
(SiteH). Fig. 8 shows the plot of the direct sunlight hours on January dataset. In the WinH ANN model, the results of average values in 41
21st. The result shows the value range from 0 to 8 h with the site’s values from the first floors of 10 buildings were used for the training
average value of 2.8 h. According to the local building code, it satisfied output dataset. The average value of the QuVue of each building was
the requirement. In Fig. 8, the backyard area called g and d had the used for the training (QuVue_ANN). The outdoor ANN models adopted
worst sunlight hours, and the main reason was because buildings no. 4, the value of the measurement based on the site. For the SiteH ANN
5, 6, 9, 10, and 11 were blocking the sunlight. Courtyard e and f had model (SiteH_ANN), the average value of the points was used for the
better sunlight exposure than g and d. The main reason was that sunlight training. To train the UTCI model, the average value of points on the
can be accessed from a gap between buildings no. 4 and 8 as well as hottest day and the coldest day on the site were used as the output
buildings no. 6 and 12. training dataset (UTCI_ANN).
A total of 7938 measuring points was used to calculate UTCI. The two Table 13 summarizes the setups for the ANN training for the study.
worst days in a year were selected for the calculation. These days were As discussed above, each of the five-performance metrics had its own
selected based on weather data (Table 12), which showed July 19th as ANN model. All five ANN models used the same 36 inputs variables for
the highest dry bulb temperature day (Ma_D), and February 1st as the the training, which were the x, y, and z coordinates of each building’s
lowest dry bulb temperature day (Mi_D). location. The number of output variables for the training depended on
On the coldest day (February 1st), the mean value of the UTCI was the indoor or outdoor measurements.
about − 9.79. The colder points were located at the corner of buildings Data was randomly divided into three sets: 70% was used for
no. 3, 6, 8, 10, 11 and 12 in black circle lines in Fig. 9. This coldest value Training, 15% was used for testing, and 15% was used for validation.
had about − 14.52 mean value and the higher values were from the Table 13 shows the number of datasets used for each ANN training.
central part of the site with dashed line ovals, an approximate − 8.80 For the ANN modeling and training, we tested with a different
mean value (Fig. 9(left)). Based on the classification, the overall thermal number of neurons, transfer functions, and training functions for each
condition was between a cold and cool range in the UTCI index ANN model, and the most suitable model setup was selected to predict
(Table 4). On the hottest day (July 19th) the mean value was about the outcome. The number of hidden layer neurons used were between
31.18. The difference between the measuring points was not significant 108 or 64, which generated better results. All the ANN models had two
(Fig. 9 (right)). On the hottest day, based on the UTCI index, most areas layers, except DF_ANN. All models were used with a hyperbolic tangent
were categorized in a warm condition. sigmoid transfer function (tansig) for transfer function. Levenberg-
Marquardt optimization was used as the training function, and the
Mean Squared Error (MSE) was used to check the performance.
4.3. Artificial neural network (ANN) modeling
To find the overall performance of accuracy, the study conducted 30
different trainings for each model, which allowed us to reach a more
Performance measures like DF, sunlight hours, and UTCI were mostly
stable correlation coefficient (R2). The average weight and bias of each
based on advanced simulation tools. Furthermore, some conventional
model are also shown in Table 13.
tools such as an UTCI calculation requires high computing cost and time.
Table 14 shows 30 training’s average accuracy for the five ANN
Even daylight metrics calculations only require few seconds to finish a
models’ reliability. The DF, QuVue, and WinH showed the average ac­
running calculation, but running multiple simulations simultaneously
curacy of R2 at 0.87, 0.80, 0.83, respectively. The SiteH, and UTCI
are complex as well as a heavy computation load in the early planning
models’ average R2 results were 0.80 and 0.87. The QuVue and SiteH
stage. To resolve the issue of calculating multiple performances at the
models showed less reliability; yet overall, all models showed a close
same time, data-driven (black-box) Artificial Neural Network (ANN)
relationship between the prediction and the simulation result so that a
models were used [84].
trained ANN can be further utilized for the optimization.
Based on the review, some studies have investigated several kinds of
expert knowledge-based methods for ANN, multiple-linear regression
(MLR), Gaussian process regression (GPR), a support vector machine 4.4. Multi-objective optimization (MOO) process
(SVM), and deep learning algorithms to compare their performance of
prediction on building performance [85,86]. Each method has its own To reduce different orders of magnitude between the objective
purpose and feature to accomplish a prediction assignment. Among functions, normalization on the existing base case results were applied.
different methods, the ANN method has some strong points in terms of For DF, DFm (Equation (2)) was divided by 4.40, the initial base case DF

Table 11
WinH (sunlight hours) of the first floor.
Building No, 1 2 3 4 5 6 8 9 10 11 Avg.

Mean 8.0 8.0 2.8 2.3 4.5 3.0 2.6 2.4 3.4 3.0 4.0
Max 8.0 8.0 5.0 3.0 5.0 4.0 4.0 3.0 5.0 5.0 5.0
Min 8.0 8.0 2.0 2.0 4.0 2.0 1.0 2.0 2.0 2.0 3.3

* Building 7 and 12 were commercial buildings that was not include in the calculation.

11
S. Wang et al. Building and Environment 205 (2021) 108263

Fig. 8. Site sunlight hours (SiteH) result of the base case.

0.63, which was the base case of the UTCI value.


Table 12
To eliminate the possibility of building a geometry solution that has
The weather information for selected days.
an intersection between buildings or located outside the site boundary,
1-Feb 19-Jul the weight factor (α) was introduced to the objective functions to
Dry Bulb temperatures ◦ C − 15 37 penalize the cases. If the building layout satisfied the rule, no weight was
Relative Humidity % 45 80 introduced; however, if the building geometry did not satisfy the rule,
Solar Radiation (Direct Normal) Wh/m2 3418 2472
the weight value 999 was multiplied by the objective function that
penalized the fitness value which would be eliminated from the next
mean value. To normalize the average sky view ratio, the QuVuem iteration of optimization.
(Equation (3)) was divided by the initial case of the QuVue mean value DFm
of 32.06. With the same approach, WinHm (Equation (4)) was divided by f1 (x) = ×α (7)
4.40
6.93 for normalization. For SiteH, SiteHm (Equation (5)) was divided by
0.47, the base case of SiteH value. UTCIm (Equation (6)) was divided by

Fig. 9. Distribution of UTCI on February 1st and July 19th.

12
S. Wang et al. Building and Environment 205 (2021) 108263

Table 13
Set up for the ANN models.
ANN No. of Dataset No. of Inputs Number of Neurons No. of Outputs Layer Transfer Function Weight Bias

DF_ANN 200 36 108/60 10 3 TANSIG − 4.1995e-04 0.2708


QuVue_ANN 200 108 41 2 − 2.7289e-04 0.1247
WinH_ANN 200 108 34 0.0029 − 0.0747
SiteH_ANN 50 108 2 − 0.0011 − 0.0250
UTCI_ANN 50 108 2 − 4.1995e-04 1.3221
Training Function Levenberg-Marquardt optimization
Data Division Random
Performance measure Mean Square Error
Training parameters Epochs: 1000, Minimum performance gradient: 1e-07, Maximum validation failures: 6, Initial mu: 0.001, mu decrease factor: 0.1,
mu increase factor: 10, Maximum mu: 1e10

Table 14
Outcomes for the ANN models.
ANN R2

DF_ANN 0.87
QuVue_ANN 0.80
WinH_ANN 0.83
SiteH_ANN 0.80
UTCI_ANN 0.87

QuVuem
f2 (x) = ×α (8)
32.06

WinHm
f3 (x) = ×α (9)
6.93

SiteHm
f4 (x) = ×α (10)
0.47

UTCIm
f5 (x) = ×α (11)
0.63
The MOO ran with 150 individuals in one generation with a
constraint dependent population creation function. The maximum iter­
ation for the MOO were set at two hundred generations. For the
offspring production, the crossover fraction was 0.8 and the crossover
ratio was 1.0. The mutation step was defined as a constraint dependent
function.
Fig. 10 shows the plot of the MOO process; the optimization stopped
at 106 generations when the average spread of the generation was less
than the initial setup. The migration step was in a forward direction with
0.2 and 20 in the interval. Furthermore, the Pareto Front population
fraction was 0.35. The last generation’s distance between the in­
dividuals was around 0.03, meaning the values had higher similarities.
In the rank histogram, the fraction of individuals in the optimal Pareto
Frontier was seen around 60 individuals in the best rank. The individuals
in the latter ranks were only dominated by the most adjacent former
rank. The average Pareto spread was 0.05, which meant that last gen­
eration’s change in distance was 0.05 from the previous generation.
The last generation’s 150 individuals were shown in Fig. 11. The
vertical axes of X1-X12, Y1–Y12 and Z1-Z12 represented the values of 36
input variables (xm, ym, zm). The other axes referred to the five
objective values (D (DF), Q (QuVue), S (SiteH), W (WinH), and U
(UTCI)). The values of the range of vertical axes were from high to low. Fig. 10. MOO process plot.
Overall, the distribution of input variables had a concentrated distri­
bution at certain ranges, except for X1, Y9, Y12, Z1, and Z8. For shows the range of 36 input values. The bar charts represented the
instance, the values of X2 were located mainly between 10 and 15 and -2 average values of 36 variables. The black lines indicate the standard
to 2 in black dash ovals with some contribution at around 0. For Y6, the deviations. Each variable shows reasonable stabilized deviation around
values were between − 4 and 12 (refers to black dash ovals). According the average. In the Fig. 12 (a), the most variables show high standard
to equation (1), the baseline case’s objectives values were − 1, − 1, − 1, variances except X7 and X12. In the Fig. 12(b), of Z variables show
− 1, and 1. The distribution of optimal D, Q, S, W, and U had a main ranges from ±2.16 to ±11.04. The wide range of variables reflects on the
range around − 1.6, − 1.0, − 1.05, − 1.0, and 0.8, respectively, which design diversity from optimization.
satisfied the minimized goal. The Table 15 explains the final generation’s variance and average
To better understand the optimal solutions in the Fig. 11, the Fig. 12 values of five results. The variance of all 5 results were smaller than

13
S. Wang et al. Building and Environment 205 (2021) 108263

Fig. 11. Parallel coordinate plots of the final generation.

0.01, which indicates that each metric was converged around the
average value as shown in the Table 15.
Because the result had a great amount of information to analyze, the As seen in Fig. 13, the top 30 optimal individuals were displayed in
paper used the rankings method to reduce the complexity and to order red lines on a parallel coordinate plot. For most individuals, the values of
the solutions. To rank the result from the MOO, the following equation the parameters existed within a certain range of distribution, but the
was used to find an equally weighted rating among the five objectives. It trend was unnoticeable. To better present the range, a black dash line
was also possible to give a different weight on each objective which help was highlighted with the average value of each parameter. To compare
users to make preferred rankings: the initial case parameters with the optimal results, the blue solid line of

(DFmax − DFi ) (QuVuemax − QuVuei ) (WinHmax − WinHi ) (SiteHmax − SiteHi ) (UTCImax − UTCIi )
Rating=NDF ∗ +NQuVue ∗ + ​ NWinH ∗ +NSiteH ∗ +NUTCI *​
(DFmax − DFmin ) (QuVuemax − QuVuemin ) (WinHmax − WinHmin ) (SiteHmax − SiteHmin ) (UTCImax − UTCImin )
(12)

the initial baseline case was plotted. Between the average value and the
baseline case, x7 and x12 had almost the same value and are drawn in
red circles. Other X and Y values in the optimal average values were
higher than those of the initial case, mostly ranging from 2 to 10.
Regarding the building’s height, the average values of Z1, Z2, and Z5
were higher than the initial values. But the other average values in the
optimal cases were lower when compared with those with the baseline
case. For the objectives, D (DF), S (SiteH) and U (UTCI) had lower
average values than the baseline case value (− 1.4 vs. − 1.0, − 1.05 vs
− 1.0, 0.75 vs. 1.0, respectively), whereas the average value of W (WinH)
was higher than that of the baseline case, which was − 0.85 compared
with − 1.0.
Based on equation (12), the top 10 best ranked individuals from the
150 individuals can be identified as in Table 16. Option 1 had the most
minimum rating value among the 10 individuals. Option 10 had both a
minimum DF and UTCI at-1.79 and 0.46, respectively. Option 8 had a
minimum QuVue at − 1.18 and WinH at − 0.99, and option 4 has a
minimum SiteH at − 1.14.
Compared with the baseline case, option 1 was scored at 0.32,
around 21% lower than the baseline case (0.41), demonstrating that the
optimal cases showed better performances than the baseline case.
However, it was interesting to find some instances where the baseline
cases perform better in some objectives than the solutions.
Fig. 14 depicts the layouts of the optimal and initial baseline
schemes, where the white blocks referred to the baseline case and the
blue solid blocks represent the optimal cases. The same as the top ten
cases shown in Table 16, the building layout plans showed some com­
mon features among the optimal solutions. The optimal buildings no. 1,
3, 4, 5, 7, and 9 were located northeast of the initial baseline case.
Moving toward the north of their initial cases occurred in buildings no.

Table 15
Five performance’s average and variance of the final generation.
DF QuVue SiteH WinH UTCI

Variance 0.081586 0.007546 0.00267 0.011234 0.018298


Fig. 12. Value distribution of 36 input variables of the last generation. Average − 1.50466 − 0.98766 − 1.02992 − 0.87402 0.763389

14
S. Wang et al. Building and Environment 205 (2021) 108263

Fig. 13. Ranking of the top 30 optimized results and initial case.

towards the north orientation at between 4 and 6 m. Building 12 moved


Table 16
the farthest away from the base case, at around 10 m. Building 2 moved
Top 10 targets with even weightings.
0.05 m to the south from its origin location.
Option DF QuVue SiteH WinH UTCI Rating In terms of building height changes (z axis), building 5 shows it
1 − 0.9263 − 1.0071 − 1.0555 − 0.7224 0.8352 0.3224 raised its height the most, from 55.8 m to 72.3 m. On the contrary,
2 − 1.7301 − 0.8571 − 1.0675 − 0.7691 0.7670 0.3407 building 7 lowered it height from 80.6 m to 60.97 m, which decreased
3 − 1.7071 − 0.9199 − 1.0212 − 0.8700 0.6898 0.3547 the most. It is interesting to see that the higher buildings in the base case
4 − 1.5249 − 0.9048 − 1.1362 − 0.7105 0.8218 0.3608
5 − 1.5491 − 0.8727 − 1.0570 − 0.7465 0.8201 0.3621
building layout lost the most height values in the average top eight
6 − 1.0366 − 1.0242 − 1.0336 − 0.7977 0.8333 0.3625 building options. For instance, building 7, 8, 10, 11, 12 became shorter
7 − 1.0585 − 1.0599 − 1.0478 − 0.6509 0.8335 0.3627 from 10.06 to 19.63 than their initial height values.
8 − 1.5445 − 1.1808 − 1.0140 − 0.9882 0.8624 0.3670
9 1.7334 0.9047 1.0206 0.8378 0.5831 0.3703
5. Discussion
− − − −
10 − 1.7890 − 0.9796 − 1.0452 − 0.9577 0.4565 0.3714
Base case − 1.0000 − 1.0000 − 1.0000 − 1.0000 1.0000 0.4127
This study first modeled a parametric building layout case to inte­
grate several performance simulations models. A baseline case was
6, 11, and 12. In the optimal cases, buildings no.1, 2, 5, and 10 were evaluated by five performance areas and the results were compared with
higher than their original baseline cases. Buildings no. 3, 4, and 6 had the building code or previous studies. For indoor performance, the DF of
almost the same height between the optimal and baseline cases. the initial building layouts for the top floors was 4.40% (for the first
In detail, Table 17 shows the top eight optimal solutions’ input floors it was 3.64%). According to benchmarks regulated in related
design variables. Looking at the average movement from the X-axis, we building codes in China, the first floor of the average DF value should not
can see that most of the buildings moved to the east orientation (positive be more than 2%. The initial case satisfies this requirement. In this
values). Among the 12 buildings, building 2 moved the farthest away research, the average QuVue result of all floors was 32.06%, compared
from the initial building position, at 10.12 m towards the east. Then to the literature which was 28%. Hence, the baseline case for this study
building 10 and building 1 moved 7 m more towards the east. Building 6, had a better sky view condition [53].
7, and 12 remained close to the original location. The WinH for the first floors in the baseline case was 5.96 h, which
In terms of the y axis location movement, most buildings moved was significantly more hours than the minimum requirement of 2 h, as

Fig. 14. Layouts for the initial case and top eight optimal cases.

15
S. Wang et al. Building and Environment 205 (2021) 108263

Table 17
Top eight optimal cases’ 36 input values.
1 2 3 4 5 6 7 8 Average Base case

X1 7.03 10.03 7.18 1.85 12.34 6.66 8.05 3.72 7.11 0


Y1 2.35 4.85 5.8 − 3.5 − 2.24 5.03 2.59 0.81 1.96 0
X2 6.63 − 1.96 1.32 2.55 2.79 11.25 1.33 10.12 4.25 0
Y2 0.03 − 1.4 − 2.19 − 0.95 1.13 1.84 1.66 − 0.51 − 0.05 0
X3 6.44 5.39 7.09 5.36 4.86 6.03 7.56 − 0.02 5.34 0
Y3 0.13 2.87 1.82 7.84 6.27 0.28 4.46 1.44 3.14 0
X4 2.7 2.11 2.61 − 1.11 1.67 5.19 1.94 6.67 2.72 0
Y4 7.38 3.72 1.51 4.95 2.69 7.36 5.65 10.75 5.5 0
X5 3.97 10.14 10.84 6.19 8.04 3.58 3.74 − 5.88 5.08 0
Y5 7.41 8.1 2.95 10.54 8.38 6.65 5.99 4.82 6.85 0
X6 − 1.16 − 0.01 2.43 − 3.51 0.35 1.05 − 1.17 0.26 − 0.22 0
Y6 8.23 1.4 3.65 0.91 2.95 11.22 8.35 13.58 6.29 0
X7 − 1.38 1.43 1.38 − 0.6 1.91 − 2.13 3.14 − 1.99 0.22 0
Y7 2.98 4.96 6.95 − 1.66 5.08 4.72 3.91 1.86 3.6 0
X8 4.47 9.98 6.03 13.27 11.49 5.46 6.37 3.42 7.56 0
Y8 7.36 9.19 10.64 1.23 8.38 2.73 1.2 5.24 5.75 0
X9 4.21 4.49 5.29 4.19 5.18 5.45 3.22 8.39 5.05 0
Y9 7.34 − 0.38 − 4.37 5.88 − 0.23 8.52 7.25 8.74 4.09 0
X10 11.71 7.3 5.46 14.53 14.33 10.51 11.21 1.72 9.6 0
Y10 4.69 5.77 7.24 10.21 4.56 0.65 − 0.15 3.99 4.62 0
X11 2.03 3.31 − 0.87 7.4 8.7 3.9 2.23 − 1.7 3.13 0
Y11 3.47 3.13 2.41 − 2.48 − 2.65 4.82 2.39 5.12 2.03 0
X12 0.15 − 1.18 − 0.45 − 2.32 − 1.38 0.64 0.87 0.16 − 0.44 0
Y12 6.37 11.36 8.52 17.93 10.8 9.94 5.16 9.68 9.97 0
Z1 74.72 89.74 80.22 91.64 86.48 76.29 70.91 73.34 80.42 65.1
Z2 69.5 67.08 68.53 70.22 70.11 67.26 71.32 65.3 68.67 65.1
Z3 59.79 70.84 63.67 76.84 67.81 59.73 51.67 64.9 64.41 68.2
Z4 66.85 59.78 64.19 60.98 58.36 65.02 60.07 69.98 63.15 71.3
Z5 83.34 72.07 67.8 71.53 74.61 76.2 86.79 46.08 72.3 55.8
Z6 69.79 71.97 72.25 72.07 71.8 63.54 65.31 66.3 69.13 74.4
Z7 58.64 65.44 71.33 50.25 52.42 73.03 51.13 65.51 60.97 80.6
Z8 73.9 78 79.31 68.92 78.78 77.37 75.66 81.96 76.74 86.8
Z9 83.46 86.98 86.33 87.18 88.2 83.53 84.55 83.32 85.44 80.6
Z10 68.37 68.37 67.97 64.71 65.61 69.86 68.1 73.03 68.25 77.5
Z11 59.33 63.59 64.67 64.71 63.59 57.5 58.32 60.52 61.53 80.6
Z12 54.4 66.71 69.19 72.69 72.51 53.53 49.96 75.91 64.36 80.6

regulated by the building code. For outdoor performance, the SiteH was with 50 simulations it would take 48–60 h. This demonstrated the ANN
2.78 h. For UTCI value on the hottest day was 31.18, and on the coldest model was able to reduce the computational time significantly.
day it was − 8.80. The SiteH satisfied the building code requirement The result from the five metric simulations and the ANN models were
(more than 2 h). In the thermal perception evaluation, the site of the used to optimize the building layouts result from the MOO shares a
UTCI value belonged to warm on the hottest day and on the coldest day significant amount of information that prevents a non-expert such as a
it was considered as cool. All five metrics showed the baseline case had designer to make a wise decision. To support this process, this paper
reasonable performance requirements, and this study tried to improve proposed a ranking system. By ranking the five-performance metrics of
them further by optimization. the MOO results, users can order the input and output variables ac­
Once the results of the basic building layout performance are cording to their preferences. Then the ranked building layouts can be
defined, this information can be used to compare with other results in demonstrated to support a variety of design schemes.
optimization process. However, optimizing the layout-related perfor­ The MOO result indicated that the top 10 solutions showed a better
mance metrics of the five building requires a tremendous number of rating score than the baseline case (Table 16). In Fig. 14, the optimal
simulations. To improve the computational time and load, the study solutions of the building layouts were compared with the initial baseline
introduced an ANN to predict the performance metrics. layout. In the optimal building layouts, the height of buildings no. 1, 2,
The first step was setting the training dataset. Five ANN models were and 5 were higher than the baseline case, whereas the height of build­
trained and tested separately. In indoor performance of the ANN models, ings no. 4, 8, 9, 11, and 12 was lower than the initial base case. However,
a total of 200 cases of simulation results of daylight factor, window the distance between buildings in the optimal solutions were varied and
sunlight hours, and the sky view ratio from equation-based models were did not show an obvious trend except building no. 5, which was located
used for training. A total of 50 cases were conducted in the outdoor to the north of the initial building in all the optimal cases.
performance (site sunlight hour and UTCI) of ANN models. After the In summary, it was decided that the height of the buildings located in
networks were trained, 40 cases for indoor ANN models and 10 cases for the south of site ground could be higher and the buildings positioned
outdoor ANN models were used for the validation. along the site boundary line could be lower than the baseline case. The
Results indicated a close relationship between the ANN prediction average value of the top 30 cases ranked better than baseline case value
and simulation outcomes. Then these trained ANN models were applied shown in Fig. 13. The X, Y values were positive, which indicated that the
to predict five performance metrics. The study used a computer with optimal buildings should be mainly located to the north and east of the
eight processors of Intel (R) Core i7-7500U CPU @ 2.70 GHz and a GPU baseline case.
of NVIDIA® GeForce® GTX 1660 Ti 6 GB GDDR6. For instance, to run
the UTCI with the ANN model, in average it took about 30 min for 6. Conclusion
training and less than minutes to run 50 simulations, for a total of 35 min
to run 50 UTCI simulation with training. With a typical UTCI calculation The study integrated various performance measures to find the
based on CFD simulation, one simulation took approximately 1 h, and optimization solutions for a high-rise residential buildings’ layout by

16
S. Wang et al. Building and Environment 205 (2021) 108263

utilizing indoor- and outdoor-related built environment measurements. Declaration of competing interest
This framework can enrich the research area by considering both visual
comforts and outdoor thermal performance conditions. The authors declare that they have no known competing financial
To do multiple simulations for different performances, the study interests or personal relationships that could have appeared to influence
introduced an ANN model. With its support, the proposed MOO models the work reported in this paper.
were able to be applied at the early stages of the site plan design process.
This study illustrates how an integrated multiple building perfor­ Acknowledgements
mance optimization framework can work, by combining a parametric
geometry model, performance-based simulation tools, machine-learning This research was supported by The Program of National Natural
algorithm, and a multi-objective optimization process. This contribution Science Foundation of China (Grant No. 52078262) and the National
of an integrated design process with an evaluation process allows the Key R&D Program of China (No. 2016YFC0700200).
user to find a better design solution.
The study also included both visual comfort and outdoor building References
thermal performance, highlighting the importance of integrating both
areas that, in turn, makes the design scheme more practical and realistic. [1] Soheil Fathi, Ravi Srinivasan, Andriel Fenner, Sahand Fathi, Machine learning
This study was the author’s first attempt to tackle this problem and it applications in urban building energy performance forecasting: a systematic
review, Renew. Sustain. Energy Rev. 133 (2020), 110287, https://doi.org/
came with certain limitations and challenges that need further attention 10.1016/j.rser.2020.110287. ISSN 1364-0321.
and research. It is important to note that this study only tested restricted [2] Sicheng Zhan, Adrian Chong, Data requirements and performance evaluation of
elements, and other conditions may require further research. For model predictive control in buildings: a modeling perspective, Renew. Sustain.
Energy Rev. 142 (2021), 110835, https://doi.org/10.1016/j.rser.2021.110835.
instance, the study only considered the building layout positions as ISSN 1364-0321.
research variables, but future research is needed to explore more design [3] Nadia D. Roman, Facundo Bre, Victor D. Fachinotti, Roberto Lamberts, Application
factors such as building orientation, the resident unit floor plan, and and characterization of metamodels based on artificial neural networks for
building performance simulation: a systematic review, Energy Build. 217 (2020),
window size. Moreover, additional test case studies should be investi­
109972.
gated to avoid the limitation and assumption of one specific case. [4] Raffaele Pelorosso, Modeling and urban planning: a systematic review of
Additionally, the study considered limiting the number of perfor­ performance-based approaches, Sustainable Cities and Society 52 (2020), 101867,
https://doi.org/10.1016/j.scs.2019.101867. ISSN 2210-6707.
mance measures for optimization and further research is required to
[5] Sunghee Lee, Youngchul Kim, A framework of biophilic urbanism for improving
explore other performance areas such as building energy demand, ma­ climate change adaptability in urban environments, Urban For. Urban Green. 61
terial cost, life cycle costs, etc. (2021), 127104, https://doi.org/10.1016/j.ufug.2021.127104. ISSN 1618-8667.
We calculated the UTCI by using probing points on the 1.5 m height [6] Evangelia Despoina Giouri, Tenpierik Martin, Michela Turrin, Zero energy
potential of a high-rise office building in a Mediterranean climate: using multi-
of the pedestrian level. To simplify the optimization, the average values objective optimization to understand the impact of design decisions towards zero-
of the total amount of spatial grid results was introduced. However, it energy high-rise buildings, Energy Build. 209 (2020) 109666, https://doi.org/
may have some limitations to capture certain building layout spatial 10.1016/j.enbuild.2019.109666.
[7] Kavan Javanroodi, Vahid M. Nik, Mohammadjavad Mahdavinejad, A novel design-
problems. One problem that emerged in the average spatial grid value of based optimization framework for enhancing the energy efficiency of high-rise
hottest and coldest days of UTCI was simplified to calculate but reduced office buildings in urban areas, Sustainable Cities and Society 49 (2019) 101597,
the value sensitivity and variations. A more reasonable function of UTCI https://doi.org/10.1016/j.scs.2019.101597.
[8] V.J.L. Gan, B. Wang, C.M. Chan, et al., Physics-based, data-driven approach for
should be discussed in the future. predicting natural ventilation of residential high-rise buildings, Build. Simul.
The constraints for the building layout only consider the limited (2021), https://doi.org/10.1007/s12273-021-0784-9.
number of building codes and it requires more specific building regu­ [9] Alejandra Susa-Páez, Piderit-Moreno, María Beatriz, 16, Geometric Optimization of
Atriums with Natural Lighting Potential for Detached High-Rise Buildings, vol. 12,
lations and standards to be included.
2020, https://doi.org/10.3390/su12166651. Sustainability.
Further study is needed to investigate a better way to model the [10] Jing Zhao, Yahui Du, Multi-objective optimization design for windows and shading
ANN. The study used a separated ANN model for different measures; configuration considering energy consumption and thermal comfort: a case study
for office building in different climatic regions of China, Sol. Energy 206 (2020)
however, it would be worthwhile to investigate different ANN models
997–1017, https://doi.org/10.1016/j.solener.2020.05.090.
and the grouping of different performance areas to reduce the number of [11] Berk Ekici, Z. Tuğçe Kazanasmaz, Michela Turrin, M. Fatih Taşgetiren, I. Sevil
ANN models. Also, research on the accuracy and a different decision- Sariyildiz, Multi-zone optimisation of high-rise buildings using artificial
making model should be tested to improve accuracy and the time. For intelligence for sustainable metropolises. Part 1: background, methodology, setup,
and machine learning results, Sol. Energy 224 (2021) 373–389, https://doi.org/
example, the number of samples in the ANN models could be improved 10.1016/j.solener.2021.05.083.
to increase its accuracy. The artificial neural network type, as well as the [12] Berk Ekici, Z. Tuğçe Kazanasmaz, Michela Turrin, M. Fatih Taşgetiren, I. Sevil
training and learning function methods, should be further discussed, and Sariyildiz, Multi-zone optimisation of high-rise buildings using artificial
intelligence for sustainable metropolises. Part 2: optimisation problems,
compared in order to select a more efficient ANN model. algorithms, results, and method validation, Sol. Energy (224) (2021) 309–326,
The ANN can predict performance with comparatively less time than https://doi.org/10.1016/j.solener.2021.05.082.
a physic-based simulation tool. However, because the ANN model is a [13] Berk Ekici, Tuğçe Kazanasmaz2, Michela Turrin1, M. Fatih Tasgetiren, I. Sevil
Sariyildiz, A Methodology for daylight optimisation of high-rise buildings in the
black box, when the input and output conditions are changed, the ANN dense urban district using overhang length and glazing type variables with
model is no longer validated and requires new training. surrogate modelling, J. Phys. Conf. 1343 (2019), 012133, https://doi.org/
In the optimization, the study used a penalty function for controlling 10.1088/1742-6596/1343/1/012133.
[14] Sewon Lee, Kyung Sun Lee, Optimization of apartment-complex layout planning
the geometry; however, it is a basic method and more advanced methods
for daylight accessibility in a high-density city with a temperate climate, Energies
can be used for penalization, for example a near-feasibility threshold 13 (16) (2020) 4172, https://doi.org/10.3390/en13164172.
(NFT). Also, the NSGA-II can be replaced with the NSGA-III or MOPSO [15] Meseret T. Kahsay, Girma T. Bitsuamlak, Fitsum Tariku, Thermal zoning and
window optimization framework for high-rise buildings, Appl. Energy 292 (2021)
(Multiple Objective Particle Swarm Optimization). NSGA-II can
116894, https://doi.org/10.1016/j.apenergy.2021.116894.
encounter difficulty when handling more than three objectives, and [16] R. Ooka, K. Komamura, Optimal design method for building energy systems using
NSGA-III or MOPSO could be a better algorithm to solve multi-objective. genetic algorithms, Build. Environ. 44 (7) (2009) 1538–1544, https://doi.org/
Instead of using Matlab for optimization other tools (Opossum [89], 10.1016/j.buildenv.2008.07.006.
[17] E. Znouda, N. Ghrab-Morcos, A. Hadj-Alouane, Optimization of Mediterranean
SilverEye [90], and Optimus [91]) can be more proper MOO model to building design using genetic algorithms, Energy Build. 39 (2) (2007) 148–153,
improve optimization. https://doi.org/10.1016/j.enbuild.2005.11.015.
[18] A. Alajmi, J. Wright, Selecting the most efficient genetic algorithm sets in solving
unconstrained building optimization problem, International Journal of Sustainable
Built Environment 3 (1) (2014) 18–26, https://doi.org/10.1016/j.
ijsbe.2014.07.003.

17
S. Wang et al. Building and Environment 205 (2021) 108263

[19] S. Li, L. Liu, C. Peng, A review of performance-oriented architectural design and Energy Build. 73 (2014) 192–199, https://doi.org/10.1016/j.
optimization in the context of sustainability: dividends and challenges, enbuild.2013.12.042. ISSN 0378-7788.
Sustainability 12 (2020) 1427, https://doi.org/10.3390/su12041427. [46] Xiaodong Xu, Yupeng Liu, Wei Wang, Ning Xu, Ke Liu, Gang Yu, Urban layout
[20] Y.K. Yi, A.M. Malkawi, Optimizing building form for energy performance based on optimization based on genetic algorithm for microclimate performance in the cold
hierarchical geometry relation, Automotation in Construction 18 (2009) 825–833, region of China, Appl. Sci. 9 (2019) 4747.
https://doi.org/10.1016/j.autcon.2009.03.006. [47] Małgorzata Fedorczak-Cisak, Kotowicz Anna, El˙zbieta Radziszewska-Zielina,
[21] A. Agirbas, Performance-based design optimization for minimal surface based Bartłomiej Sroka, Tadeusz Tatara, Krzysztof Barnas, Multi-criteria optimisation of
form, Architecture Science Review 61 (2018) 384–399, https://doi.org/10.1080/ an experimental complex of single-family nearly zero-energy buildings, Energies
00038628.2018.1523130. 13 (2020) 1541.
[22] V. Granadeiro, J.P. Duarte, J.R. Correia, V.M.S. Leal, Building envelope shape [48] José Sánchez de la Flor, Servando Álvarez Dominguez, Jose Luis Molina Felix,
design in early stages of the design process: integrating architectural design Rocío González- Falcón, Climatic zoning and its application to Spanish building
systems and energy simulation, Automotation in Construction 32 (2013) 196–209, energy performance regulations Francisco, Energy Build. 40 (2008), https://doi.
https://doi.org/10.1016/j.autcon.2012.12.003. org/10.1016/j.enbuild.2008.05.006. ISSN 0378-7788,1984–1990.
[23] W. Marks, Multicriteria optimisation energy-saving buildings, Build. Environ. 32 [49] Athina G. Gaglia, Aikaterini G. Tsikaloudaki, Costantinos M. Laskos, Evangelos
(4) (1997) 331–339, https://doi.org/10.1016/S0360-1323(96)00065-0. N. Dialynas, Athanassios Argiriou, The impact of the energy performance
[24] S. Khajehpour, D.E. Grierson, Method for conceptual design applied to office regulations’ updated on the construction technology, economics and energy
buildings, J. Comput. Civ. Eng. 6 (2) (2002) 83–103, https://doi.org/10.1061/ aspects of new residential buildings: the case of Greece, Energy Build. (15) (2017)
(ASCE)0887-3801(2002)16:2(83). 225–237.
[25] K. Suga, S. Kato, K. Hiyama, Structural analysis of Pareto-optimal solution sets for [50] D. Pilzer, Performance Based Building Regulations, Technical Research Centre of
multi-objective optimization: an application to outer window design problems Finland, 2005.
using Multiple Objective Genetic Algorithms, Build. Environ. 45 (5) (2010) [51] Yi Yun Kyu, Hyoungsub Kim, Agent-based geometry optimization with Genetic
1144–1152, https://doi.org/10.1016/j.buildenv.2009.10.021. Algorithm (GA) for tall apartment’s solar right, Sol. Energy 113 (2015) 236–250.
[26] D. Tuhus-Dubrow, M. Krarti, Genetic-algorithm based approach to optimize [52] Edward Ng, Studies on daylight design and regulation of high-density residential
building envelope design for residential buildings, Build. Environ. 45 (7) (2010) housing in Hong Kong, Light.Res. Lighting Research and Technology Technol 352
1574–1581, https://doi.org/10.1016/j.buildenv.2010.01.005. (2003) 127–139.
[27] C. Diakaki, E. Grigoroudis, D. Kolokotsa, Towards a multi-objective optimization [53] Hussain H. Alzoubi, Shadan Dwairi, Re-assessment of national energy codes in
approach for improving energy efficiency in buildings, Energy Build. 40 (2008) Jordan in terms of energy consumption and solar right in residential buildings,
1747–1754, https://doi.org/10.1016/j.enbuild.2008.03.002. Sustainable Cities and Society 15 (2015) 161–165, https://doi.org/10.1016/j.
[28] W. Wang, H. Rivard, R. Zmeureanu, An object-oriented framework for simulation- scs.2014.10.006. ISSN 2210-6707.
based green building design optimization with genetic algorithms, Adv. Eng. Inf. [54] Aurélie Foucquier, Sylvain Robert, Frédéric Suard, Louis Stéphan, Arnaud Jay,
19 (1) (2005) 5–23, https://doi.org/10.1016/j.aei.2005.03.002. State of the art in building modelling and energy performances prediction: a
[29] O.T. Karaguzel, R. Zhang, K.P. Lam, Coupling of whole-building energy simulation review, Renew. Sustain. Energy Rev. 23 (2013) 272–288, https://doi.org/10.1016/
and multidimensional numerical optimization for minimizing the life cycle costs of j.rser.2013.03.004. ISSN 1364-0321.
office buildings, Building Simulation 7 (2) (2014) 111–121, https://doi.org/ [55] Zeyu Wang, Ravi S. Srinivasan, A review of artificial intelligence based building
10.1007/s12273-013-0128-5. energy use prediction: contrasting the capabilities of single and ensemble
[30] W. Wang, R. Zmeureanu, H. Rivard, Applying multi-objective genetic algorithms in prediction models, Renew. Sustain. Energy Rev. 75 (2017) 796–808, https://doi.
green building design optimization, Build. Environ. 40 (11) (2005) 1512–1525, org/10.1016/j.rser.2016.10.079. ISSN 1364-0321.
https://doi.org/10.1016/j.buildenv.2004.11.017. [56] Danica Lea Larcombe, Eddie van Etten, Alan Logan, L. Susan, Prescott and pierre
[31] W. Yu, B. Li, H. Jia, M. Zhang, D. Wang, Application of multi-objective genetic horwitz. High-rise apartments and urban mental health—historical and
algorithm to optimize energy efficiency and thermal comfort in building design, contemporary views, Challenges 10 (34) (2019).
Energy Build. 88 (2015) 135–143, https://doi.org/10.1016/j. [57] Christoph Waibel, Thomas Wortmann, Ralph Evins, Carmeliet Jan, Building energy
enbuild.2014.11.063. optimization: an extensive benchmark of global search algorithms, Energy Build.
[32] X. Chen, H. Yang, A multi-stage optimization of passively designed high-rise 187 (2019) 218–240, https://doi.org/10.1016/j.enbuild.2019.01.048. ISSN 0378-
residential buildings in multiple building operation scenarios, Appl. Energy 206 7788.
(2017) 541–557, https://doi.org/10.1016/j.apenergy.2017.08.204. [58] Laurent Magnier, Fariborz Haghighat, Multiobjective optimization of building
[33] Fialho, Y. Hamadi, M. Schoenauer, Optimizing architectural and structural aspects design using TRNSYS simulations, genetic algorithm, and Artificial Neural
of buildings towards higher energy efficiency, in: Proceedings of the 13thAnnual Network, Build. Environ. 45 (3) (2010) 739–746, https://doi.org/10.1016/j.
Conference Companion on Genetic and Evolutionary Computation, 2011, buildenv.2009.08.016.
pp. 727–732. http://dl.acm.org/citation.cfm?id=2002077. [59] Sareh Naji, Afram Keivani, Shahaboddin Shamshirband, U. Johnson Alengaram,
[34] L. Magnier, F. Haghighat, Multiobjective optimization of building design using Mohd Zamin Jumaat, Zulkefli Mansor, Malrey Lee, Estimating building energy
TRNSYS simulations,genetic algorithm, and Artificial Neural Network, Build. consumption using extreme learning machine method, Energy 97 (2016) 506–516,
Environ. 45 (3) (2010) 739–746, https://doi.org/10.1016/j.buildenv.2009.08.016. https://doi.org/10.1016/j.energy.2015.11.037.
[35] K. Konis, A. Gamas, K. Kensek, Passive performance and building form: an [60] Jin Woo Moon, Jong-Jin Kim, ANN-based thermal control models for residential
optimization framework for early-stage design support, Sol. Energy 125 (2016) buildings, Build. Environ. 45 (7) (2010) 1612–1625, https://doi.org/10.1016/j.
161–179. buildenv.2010.01.009.
[36] J. Yao, Energy optimization of building design for different housing units in [61] Yihan Wu, Qingming Zhan, Steven Jige Quan, Yuli Fan, Yu Yang, A surrogate-
apartment buildings, Appl. Energy 94 (2012) 330–337. assisted optimization framework for microclimate-sensitive urban design practice,
[37] L. Tronchin, M. Manfren, L.C. Tagliabue, Optimization of building energy Build. Environ. 195 (2021) 107661, https://doi.org/10.1016/j.
performance by means of multi-scale analysis – lessons learned from case studies, buildenv.2021.107661.
Sustainable Cities and Society 27 (2016) 296–306. [62] MOHURD of China, Standard for Daylighting Design of Buildings (GB50033-2013),
[38] Kalyanmoy Deb, Associate Member, , IEEE, Amrit Pratap, Sameer Agarwal, China Architecture and Building Press, Beijing, 2013 (in Chinese).
T. Meyarivan, A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE [63] Hyoungsub Kim, Yi Yun Kyu, Quvue implementation for decisions related to high-
Trans. Evol. Comput. 6 (2) (2002) 182–197. rise residential building layouts, Build. Environ. 148 (2019) 116–127, https://doi.
[39] T.M. Echenagucia, A. Capozzoli, Y. Cascone, M. Sassone, The early design stage of a org/10.1016/j.buildenv.2018.10.055. ISSN 0360-1323.
building envelope: multi-objective search through heating, cooling and lighting [64] http://patternguide.advancedbuildings.net/using-this-guide/analysis-methods/da
energy performance analysis, Appl. Energy 154 (2015) 577–591. ylight-factor.
[40] Kavan Javanroodi, Vahid M. Nik, Mohammadjavad Mahdavinejad, A novel design- [65] Boubekri Mohamed, An overview of the current state of daylight legislation,
based optimization framework for enhancing the energy efficiency of high-rise Journal of the Human Environmental System 7 (2) (2004) 57–63.
office buildings in urban areas, Sustainable Cities and Society 49 (2019) 101597, [66] ISO 10916:2014 Calculation of the impact of daylight utilization on the net and
https://doi.org/10.1016/j.scs.2019.101597. ISSN 2210-6707. final energy demand for lighting. https://www.iso.org/standard/46394.html.
[41] Botir Giyasov, Vyacheslav Semenov, Irina Giyasova, Optimization of thermal [67] https://www.bregroup.com/products/breeam/.
stability of structural elements of high-rise buildings to increase their energy [68] https://www.dgnb-system.de/en/buildings/.
efficiency, MATEC Web of Conferences 196 (2018), 04084. [69] Timur Dogan, Ye Chan Park, A new framework for residential daylight
[42] J. Strømann-Andersen, P.A. Sattrup, The urban canyon and building energy use: performance evaluation, in: Proceedings of the International IBPSA Building
urban density versus daylight and passive solar gains, Energy Build. 43 (8) (2011) Simulation Conference, 2017. San Francisco, USA.
2011–2020, https://doi.org/10.1016/j.enbuild.2011.04.007. ISSN 0378-7788. [70] MOHURD of China, Code for Fire Protection Design of Buildings, GB 50016-2014,
[43] Juan José Sarralde, David James Quinn, Daniel Wiesmann, Koen Steemers, Solar China Architecture and Building Press, Beijing, 2018, 2015 (in Chinese).
energy and urban morphology: scenarios for increasing the renewable energy [71] https://publish.illinois.edu/quvuetool/.
potential of neighbourhoods in London, Renew. Energy 73 (2015) 10–17, https:// [72] David Rutten, RhinoScript for Rhinoceros 4.0, Robert McNeel & Associates, Seattle,
doi.org/10.1016/j.renene.2014.06.028. ISSN 0960-1481. 2007.
[44] Vincent J.L. Gan, H.K. Wong, K.T. Tse, Jack C.P. Cheng, M.C. Lo Irene, C.M. Chan, [73] MOHURD of China, Design Standard for Energy Efficiency of Residential Buildings
Simulation-based evolutionary optimization for energy-efficient layout plan design in Severe Cold and Cold Zones (JGJ 26-2018), China Architecture and Building
of high-rise residential buildings, J. Clean. Prod. 231 (2019) 1375–1388, https:// Press, Beijing, 2018 (in Chinese).
doi.org/10.1016/j.jclepro.2019.05.324. ISSN 0959-6526. [74] Sookuk Park, E. Stanton, Tuller, myunghee jo, application of universal thermal
[45] Enes Yaşa, Ok Vildan, Evaluation of the effects of courtyard building shapes on climate index (UTCI) for microclimatic analysis in urban thermal environments,
solar heat gains and energy efficiency according to different climatic regions,

18
S. Wang et al. Building and Environment 205 (2021) 108263

landscape and urban planning, Landsc. Urban Plann. 125 (2014) 146–155, https:// [84] http://www.cleveralgorithms.com/.
doi.org/10.1016/j.landurbplan.2014.02.014. ISSN 0169-2046. [85] Liping Wang, Robert Kubichek, Xiaohui Zhou, Adaptive learning based data-driven
[75] Littlefair Paul, Site Layout Planning for Daylight and Sunlight: a Guide to Good models for predicting hourly building energy use, Energy Build. 159 (2018)
Practice, second ed., IHS BRE Press, 2011. 454–461, https://doi.org/10.1016/j.enbuild.2017.10.054. ISSN 0378-7788.
[76] MOHURD of China, Standard for Urban Residential Area Planning and Design (GB [86] S.L. Wong, Kevin K.W. Wan, Tony N.T. Lam, Artificial neural networks for energy
50180-2018), China Architecture and Building Press, Beijing, 2018 (in Chinese). analysis of office buildings with daylighting, Appl. Energy 87 (Issue 2) (2010)
[77] MOHURD of China, Assessment Standard for Green Building GB/T 50378-2019, 551–557, https://doi.org/10.1016/j.apenergy.2009.06.028. ISSN 0306-2619.
China Architecture and Building Press, Beijing, 2019 (in Chinese). [87] Zhaoyang Ye, Moon Keun Kim, Predicting electricity consumption in a building
[78] Bröde Peter, Dusan Fiala, Krzysztof Błażejczyk, Ingvar Holmér, Gerd Jendritzky, using an optimized back-propagation and Levenberg–Marquardt back-propagation
Bernhard Kampmann, Birger Tinz, Havenith George, Deriving the operational neural network: case study of a shopping mall in China, Sustainable Cities and
procedure for the universal thermal climate index (UTCI), Int. J. Biometeorol. 56 Society 42 (2018) 176–183, https://doi.org/10.1016/j.scs.2018.05.050. ISSN
(2012) 481–494, https://doi.org/10.1007/s00484-011-0454-1. 2210-6707.
[79] Andreas Matzarakis, Stefan Muthers, Rutz Frank, Application and comparison of [88] Cihan Turhan, Tugce Kazanasmaz, Ilknur Erlalelitepe Uygun, Kenan Evren Ekmen,
UTCI and PET in temperature climate conditions, Finisterra 49 (98) (2014) 11–21, Gulden Gokcen Akkurt, Comparative study of a building energy performance
https://doi.org/10.18055/Finis6453. Universidade de Lisboa. software (KEP-IYTE-ESS) and ANN-based building heat load estimation, Energy
[80] Sajad Zare, Naser Hasheminejad, Hossein Elahi Shirvan, Rasoul Hemmatjo, Build. 85 (2014) 115–125, https://doi.org/10.1016/j.enbuild.2014.09.026. ISSN
Keyvan Sarebanzadeh, Saeid Ahmadi, Comparing Universal Thermal Climate Index 0378-7788.
(UTCI) with selected thermal indices/environmental parameters during 12 months [89] Thomas Wortmann, Opossum-introducing and Evaluating a Model-Based
of the year, Weather and Climate Extremes 19 (2018) 49–57, https://doi.org/ Optimization Tool for Grasshopper, 2017.
10.1016/j.wace.2018.01.004. ISSN 2212-0947. [90] Judyta M. Cichocka, Agata Migalska, Will N. Browne, Edgar Rodriguez,
[81] Patrick Kastner, Timur Dogan, A cylindrical meshing methodology for annual SILVEREYE-the implementation of Particle Swarm Optimization algorithm in a
urban computational fluid dynamics simulations, Journal of Building Performance design optimization tool, in: International Conference on Computer-Aided
Simulation 13 (1) (2020) 59–68, https://doi.org/10.1080/ Architectural Design Futures, Springer, Singapore, 2017.
19401493.2019.1692906. [91] Cemre çubukçuoğlu, Berk Ekici, Mehmet Fatih Tasgetiren, Sevil Sariyildiz,
[82] Hrvoje Jasak, Open source CFD in research and industry, International Journal of OPTIMUS: self-adaptive differential evolution with ensemble of mutation strategies
Naval Architecture and Ocean Engineering 1 (2) (2009) 89–94. for grasshopper algorithmic modeling, Algorithms 12 (7) (2019) 141.
[83] J. Higham Desmond, J. Higham Nicholas, MATLAB Guide, Society for Industrial
and Applied Mathematics, 2016.

19

You might also like