Energy Reports 7 (2021) 981–996

Contents lists available at ScienceDirect

Energy Reports
journal homepage: www.elsevier.com/locate/egyr

A comprehensive method for optimizing the design of a regular

architectural space to improve building performance
∗
Yukai Zou a , Qiaosheng Zhan a , Ke Xiang a,b ,
a
School of Architecture, South China University of Technology, Guangzhou, Guangdong, China
b
National Key Laboratory of Subtropical Building Science, South China University of Technology, Guangzhou, Guangdong, China

article info a b s t r a c t

Article history: Improving building performance is of great significance to protecting the ecology, saving energy and
Received 4 November 2020 enhancing the living environment. This paper developed a comprehensive method for optimizing
Received in revised form 14 January 2021 the design of a regular architectural space to improve building performance. The research aims to
Accepted 31 January 2021
provide architects with robust and accurate design references when conducting design tasks. The
Available online 11 February 2021
entire optimization process is divided into three steps. The first step is to build a database by
Keywords: generating the research objects randomly and performing building simulations on them. The second
Parametric simulation step is to train artificial neural network (ANN) models as a substitute of the time-consuming building
Artificial neural networks simulation in the multi-objective optimization to predict the building performance quickly. The last
Multi-objective optimization step is to perform multi-objective optimization based on the actual design constraints. To demonstrate
Genetic algorithm the optimization process, a common type of classroom space defined by 30 design parameters was
Building performance selected as a case study. The optimization objectives were set as energy demand, thermal environment
and daylight environment. The accuracy of different ANN models was assessed. To imitate realistic
design tasks, three design scenarios with constraints are used in the optimization. All the three
optimizations finished in 350 s. Compared to the traditional optimization method based on simulation,
the optimization calculation is accelerated by approximately 2,570 times. To ensure the reliability of
the optimization results, several nondominated solutions of each case were validated by simulation,
and there were good agreements between the simulation results and the optimization results. An
integrated solution and a reference solution was defined for each case. Compared to the reference
solution, the objectives of the integrated solution of the three cases have been improved by 24.6%,
18.7% and 14.2% on average, respectively, indicating that this method is feasible and effective to
improve building design in actual tasks.
© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction 2018) – and indoor comfort—since people spend most of their

time indoors currently (Hoppe, 2002). Sometimes the goals of
Building is an essential part of lives of human beings because improving building performance conflict with each other (Wang
it is a container for various activities. However, it is undeniable et al., 2005). In the context of global warming, scarce resources,
that the impact of modern construction industry on the earth environmental degradation, and population growth, it has always
ecosystem cannot be ignored. People are committed to finding been an important issue for human society to seek a balance
ways to minimize the negative impact and maximize the benefits among many aspects of building performance.
of buildings under various restrictions, such as reducing construc-
It is well known that architectural design has a vital influence
tion waste (Liu et al., 2020a,b), producing modular buildings (Sun
on building performance, which has been confirmed in numerous
et al., 2020), etc. Among all the measures, improving building
studies on building performance optimization. Architecture is a
performance has long been of interest to the public, especially
complex system involving a wide range of fields. When architects
government personnel, researchers, and architects. Building per-
formance covers many targets, such as energy efficiency – since conduct architectural design, improving building performance is
the building sector accounts for a large part of the world’s en- not the only valuable pursuit. They must also consider issues such
ergy consumption (Cao et al., 2016; Amasyali and El-Gohary, as the spatial, spiritual, and social attributes of the architecture
and rely on design experience to find a suitable building plan to
∗ Corresponding author at: School of Architecture, South China University of meet the expectations of the users and owners of the building and
Technology, Guangzhou, Guangdong, China. even the whole of society. As far as the building performance is
E-mail address: [email protected] (K. Xiang). concerned, due to time and knowledge limitations, the task of the

https://doi.org/10.1016/j.egyr.2021.01.097
2352-4847/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

architect is to make the building comply with local regulations the energy-saving effects of building elements. The results of
and not to improve it as much as possible, especially in places this research can provide architects with valuable references
where there are a great number of construction projects with when designing tropical houses and reducing buildings’ cooling
generally short construction periods, such as China. To fully tap load (Sadeghifam et al., 2015). Fabrizio Ascionea et al. derived
the potential of building performance, it is necessary to directly a comprehensive methodology to conduct the multi-objective
tell the architect what to do instead of what standards to meet. optimization of a building’s energy design. When the method was
Indeed, the scenarios of architectural design are ever- applied to a typical office building in Milan, the primary energy
changing, and it seems to be an impossible proposition to exhaust consumption, global costs, and CO 2 emissions were significantly
every design possibility for every type of architecture under reduced (Ascione et al., 2019). Yunsong Han et al. used genetic
every weather condition in research. The issue of how to make algorithms to optimize the four performance measures of timber-
the research results on building performance optimization more glass residential buildings in severe cold regions by changing the
universal and more representative is a matter that architectural building form variables (Han et al., 2017). Wei Yu et al. combined
researchers must consider. One realistic way is to summarize and an artificial neural network and multi-objective evolutionary
extract classic architectural patterns and conduct optimization algorithm to help designers optimize building variables (Yu et al.,
research on them so that the research results have referential 2015). Similar work has been done by Gou et al. (2018). Facundo
value for similar practical cases. In addition, there are many Bree et al. combined sensitivity analysis and the genetic algorithm
constraints in actual architectural design (Machairas et al., 2014). to improve the design of a single-family house and used a simple
Architects are committed to creating a suitable rather than a weighted-sum approach to evaluate the thermal comfort and
perfect architectural solution. Therefore, another item worth pon- energy consumption of the building (Bre et al., 2016).
dering is how to incorporate the many practical constraints faced Some scholars have conducted studies on certain architectural
by architects into the scope of the research and enable architects spaces. Yu Ting Kowk et al. investigated the thermal comfort and
to quickly obtain effective and specific design guidance. energy consumption of four types of typical rental houses in Hong
Kong under typical and near-extreme climates and optimized
1.1. Overview of the recent research on architectural design opti- the design of one of them to explore its energy-saving potential
mization (Kwok et al., 2017). Anxiao Zhang et al. used the genetic algorithm
to optimize the thermal, daylight and energy performance of
New techniques promote the development of the construc- three classic types of classrooms in northern China by chang-
tion field (Wu et al., 2020; Xu et al., 2020). Building science ing several passive design parameters (Zhang et al., 2017). Xi
has experienced great improvements in the past few decades, Chen et al. reviewed the simulation-based methods for improving
especially after the introduction of computer simulation into the passive design of buildings and conducted experiments on
the field (Hensen and Lamberts, 2011). Building simulation tools a generic residential building by combining sensitivity analyses
such as EnergyPlus (https://energyplus.net/), DOE-2 and TRNSYS with the process of improving the early-stage design (Chen et al.,
enable people without relevant professional knowledge to ob- 2018). Navid Delgarm et al. combined the genetic algorithm with
tain intuitive building performance results once models are set, a building simulation tool to test the effects of some architectural
thereby greatly reducing the threshold for the building perfor- parameters of a typical room on building energy consumption in
mance optimization research. With the continuous computing different climates of Iran. The results indicated that a considerable
power enhancements and simulation tool updates, the research amount of energy reduction can be achieved when using the
on building performance optimization has entered all aspects of optimum configuration in each weather condition (Delgarm et al.,
architectural design. 2016). In a study conducted by Maria Ferrara et al. the genetic
There are many studies that take an entire building as the algorithm was applied to optimize the passive design variables
study object. Fatima Harkouss et al. took a typical residential of a classroom with the goal of saving energy, and then the
building as the research object, optimized its passive design criteria of thermal comfort and visual comfort were used to filter
through a comprehensive approach, and found the best solution the results (Ferrara et al., 2015). Rizki A. Mangkuto et al. tested
to reduce building energy consumption and the life-cycle costs in the impact of the window size, window orientation and wall
different climates (Harkouss et al., 2018). Babak Raji et al. estab- reflectance on different daylight metrics and the lighting energy
lished six natural ventilation schemes in a 21-story office building consumption of an office room in the tropics (Mangkuto et al.,
in the Netherlands to explore the possibility of adopting natural 2016).
ventilation strategies to reduce building energy consumption in In addition to the two types of building research mentioned
the summer. The results showed that the proper use of natural above, there are existing studies on building components such as
ventilation can provide comfortable conditions during most of the windows, shading devices, etc. Niccolò Aste et al. evaluated glaz-
summer occupancy time, thus reducing the working time of air ing’s technoeconomic performance for office buildings in three
conditioning and traditional mechanical ventilation and saving different European climates. They set groups of models with
a considerable amount of energy (Raji et al., 2020). Miroslav different window designs to study the impact of glazing on build-
Premrov et al. studied the influence of building shapes on the en- ing performance. The research results showed that reducing the
ergy consumption of timber-glass buildings in different climates. amount of glazing and installing external shading facilities are the
The results showed that the impact of the building shape on most reasonable measures to reduce building energy consump-
energy consumption is different under various climate conditions tion (Aste et al., 2018). Touraj Ashrafian and Nazanin Moazzen
(Premrov et al., 2016). Waleed Khalid Alhuwayil et al. conducted selected two classrooms with different orientations in a building
a study on the energy consumption of a multi-story hotel building in Turkey as cases to explore the impact of the window-to-wall
in a hot and humid area with different external shading strategies. ratio and window combinations on indoor comfort and the en-
The results indicated that compared with the base case, the ergy demand (Ashrafian and Moazzen, 2019). Mohannad Bayoumi
retrofit plan with the proposed shading effectively eliminated a discussed the relationship between the window opening grade
significant proportion of the energy use, and the investment was and energy consumption in office spaces in two hot climates.
cost-effective because of the short pay-back period (Alhuwayil The results showed that opening the windows to a certain ex-
et al., 2019). Aidin Nobahar Sadeghifam et al. used a two-story tent helps to save energy in the building, especially at certain
residential building in Kuala Lumpur as an example to examine times of the year (Bayoumi, 2017). Marina Khoroshiltseva et al.
982
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

focused on the design of energy-efficient shading devices for faced by architects must be considered in the research. This
residential building in Madrid and developed a multi-objective means that in the optimization process, it is necessary to give
approach to optimize the shape of shading devices to maintain the architects a channel for feedback. Only then can the ‘‘proper
high levels of indoor comfort on the premise of low energy con- perfect solutions’’ be found.
sumption (Khoroshiltseva et al., 2016). Kristoffer Negendahl and To compensate for these drawbacks and provide robust and
Toke Rammer Nielsen improved building performance regarding accurate guidance for architects in design tasks, the aims of this
energy consumption, the indoor environment and investment study are as follows:
by optimizing the geometry of a folded façade at early design
stages (Negendahl and Nielsen, 2015). Ellika Taveres-Cachat et al. - Take regular architectural spaces that can be described by
optimized the PV integrated shading devices with multiple goals limited design parameters as the research object.
by adjusting the number, individual tilt angle and positions along - Create a feedback channel in the optimization process to
the vertical axis of louvre blades (Taveres-Cachat et al., 2019). take the practical constraints into account.
Yuan Fang et al. employed a genetic method to optimize the - Conduct multi-objective optimization on the research object
building geometry and fenestration of a small office in three and obtain the optimal solutions in a short time.
American cities to maximize daylight environment and energy
performance (Fang and Cho, 2019). Tomás Méndez Echenagucia 2. Methodology
et al. conducted research to promote the early design stage of
a building envelope. They searched for optimal solutions for an The multi-stage optimization process consists of three steps
open office space exterior wall in the conditions of certain cities (Fig. 2). The first step is generating a database, the second step is
with and without an urban context, aiming to reduce the cooling, training the artificial neural network (ANN) model, and the third
heating and lighting energy of the building (Méndez Echenagucia step is optimization based on a genetic algorithm (GA).
et al., 2015). Zaid Romani et al. performed a simultaneous opti-
mization of building envelope of residential buildings in certain 2.1. Step 1: generation of a database
climate zones of Morocco based on desirability function concept
and simplex Nelder–Mead algorithm (Romani et al., 2015). Ming- The first step is to generate a database of research objects
Der Yang et al. took an office building in southern Taiwan as the and their performances for ANN training. This step requires para-
object to validate their multi-objective optimal building envelope metric tools and building simulation tools. Parametric tools are
energy estimation model, which was developed to minimize the used to generate a set of research objects defined by random
envelope construction cost and energy demand while maximizing design parameters. The sample size of the research object should
the window opening rate (Yang et al., 2017). meet the training requirements of the ANN. Building simulation
tools take the research objects generated by parametric tools and
1.2. Aim of the current study weather data as input, calculate the results and feed the results
back to the parametric tools for collecting.
As mentioned above, in order to provide effective and reli- There are two reasons the design parameters of the research
able guidance for architects in design, there are two issues to object are randomly generated. One is to imitate the real situa-
consider. One is the universality of the research, that is, the tion. The second is that if a certain number of values are given to
reference value to similar situations. The other is practicality, each design parameter, in addition to the potential possibility of
that is, the constraints that architects need to face should be increasing the ANN prediction error due to the limited number of
considered. However, throughout the existing studies, few studies values for each parameter, the number of cases is also too large
simultaneously incorporate these two points. in terms of the simulation running time. If there are 10 design
As is known, many design parameters are required to describe parameters with 5 values for each, then 9,765,625 simulations
an entire building completely, and the building performance is will be required. Nevertheless, the reality is that there are many
more or less affected by each design parameter. However, the more design parameters that define a space than this. The ad-
design possibilities are endless, and building performance simula- vantage of randomly generating design parameters is that in a
tion is quite time-consuming; therefore, it is unrealistic to explore limited number of simulations, each parameter has sufficiently
the impact of every design possibility on building performance. different values within the possible range. It is worth noting
Research on specific buildings or specific spaces is too character- that the setting of the design parameters and their value ranges
istic to provide detailed design references for other cases. In terms should be carefully considered and should conform to the actual
of the architectural morphology and typology, most of the spaces situation to ensure the ANN prediction accuracy.
in a common building are usually composed of regular spatial
units (Fig. 1). Examples include the guest rooms in hotels, the 2.2. Step 2: training of the artificial neural network
wards in hospitals, and the classrooms in schools. Each specific
type of regular spatial units shares similar characteristics and can The second step is to establish an ANN model as the algorithm
be described with limited design parameters, which makes them for the building performance assessment in the next step. As is
valuable research objects when considering their universality and known, to obtain the building performance simulation results, it
feasibility. is necessary to establish a simulation model and wait for the com-
In addition, improving building performance is just one of puter to run the program, which is time-consuming (Eisenhower
the valuable pursuits of architectural designs. Architects are re- et al., 2012) and inconvenient. The GA is a popular optimizer
stricted by many factors in design tasks, such as local regulations, that can be coupled with building performance simulation tools.
usage requirements, investment scale, etc. Most existing building One of the characteristics of the process is that it continuously
performance optimization studies search for the perfect solutions crosses and mutates the population to calculate optimization
to the issue. The problem is that even if the research object is results, which means multiple rounds of building simulations.
very similar to the actual design task, in most of the cases, the The purpose of developing surrogate models is to get the build-
architects still cannot get accurate and effective guidance because ing performance prediction results quickly within the acceptable
the perfect solutions to the task are unlikely to be acquired due error range when inputting the design parameters. In this case,
to the practical constraints. To address this issue, the constraints once the surrogate model of a certain kind of regular architectural
983
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 1. Architectural space composition.

Fig. 2. Optimization flowchart.

984
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

space in a specific climate is established, it can be reused, greatly space is limited. To give another example, the design parameters
saving the time of repeated simulation calculations. Therefore, the of glazing should be set as the attributes of a certain type of
surrogate model can be used to conduct sensitive analysis and is existing glass when setting optimization conditions in case the
a viable alternative to the simulation process in the optimization. optimization result of the glazing is not a realistic option. The
There are several techniques used to create a surrogate model setting of the design constraints is indispensable for calculating
(Li et al., 2010) for building performance prediction, such as the accurate practical optimal solutions.
kriging (Eguía et al., 2016; Yi, 2016), the support vector regression
(Dong et al., 2005; Eisenhower et al., 2012) and the Gaussian 2.4. Tools
process regression (Heo and Zavala, 2012; Yoon and Moon, 2018).
The ANN is the most popular one among them, and evidence The optimization process ran on Grasshopper (https://www.
indicates that it is an efficient way to predict building perfor- grasshopper3d.com/) and Python (https://www.python.org/)
mance (Wong et al., 2010; Castilla et al., 2013; Moon et al., 2013; (Fig. 3). The first step of the optimization was performed on
Gou et al., 2018). ANN is a non-linear mathematical model that Grasshopper. Grasshopper is a visual programming language and
imitates the structure and function of biological neural networks. environment that runs within the Rhinoceros 3D computer-aided
Typical ANN models have three kinds of layers, namely input design (CAD) application. Ladybug tools (https://www.ladybug.
layer, hidden layer and output layer, and there are numerous tools/) are the plug-ins of Grasshopper that connect Grasshopper
neurons in the layers. The more hidden layers and neurons are, and validated simulation engines such as EnergyPlus and Radi-
the more complicated the ANN is. In the field of scientific comput- ance (https://www.radiance-online.org/). These plugins convert
ing, the accuracy of the calculation is often inversely proportional parameterized information into model information as the input
to time-consuming (Huang et al., 2020). It is very important to of simulation tools. Colibri (https://core.thorntontomasetti.com/
design a suitable structure for ANN models. An ANN model that is colibri-release/) is another plug-in of Grasshopper used to gen-
too simple may not meet the accuracy requirements, and an ANN erate random design parameters and collect simulation results.
model that is too complex will increase unnecessary training time The remaining two steps were implemented on Python. Python
and the risk of overfitting. is a worldwide popular programming language, and many useful
Before training the ANN, the data must be preprocessed. First, software packages have been developed on this platform. Tensor-
we normalize the data to prevent the excessive magnitude dif- Flow (https://www.tensorflow.org) is a core open source library
ferences between the parameters from having adverse effects used to develop and train machine learning models. In this study,
on the training. Second, we divide the database into two or it used the simulation results from Colibri and trained an ANN to
three parts. One part is for training the ANN and the other part replace the simulation process. Geatpy (https://geatpy.com/) is an
is for validating and testing the model to check whether it is evolutionary algorithm (EA) toolbox for Python and was selected
underfitting or overfitting. The ANN continues updating itself by as the optimization tool in this research. When the optimization
comparing its output with the output of sampling datasets during objectives, ranges and constraints of the design parameters are
the training process until the prediction result reaches the target. set, the program is able to search for the optimal solutions using
the ANN model. The GA adopted in this research was NSGA-II
2.3. Step 3: optimization based on the genetic algorithm (Deb, 2002).

The last step is to optimize the design parameters of the 3. Description of the case study
object with multiple objectives based on the GA. GA is a kind of
population-based algorithm to deal with optimization problem. 3.1. Object and weather conditions
The invention of GA was inspired by some phenomena in evo-
lutionary biology. GA is capable to deal with nonlinear problems A classroom is an important type of daily space with regular
with discontinuities and local minima. So far, GA is widely used shapes. Studies show that the classroom environment affects
as the optimizer in building optimization research (Machairas students’ health and learning efficiency (Mishra and Ramgopal,
et al., 2014; Kheiri, 2018). In this study, the optimization process 2014; Singh et al., 2018). Additionally, owing to the massive
takes the trained ANN model in step 2 as the algorithm for the presence of classrooms, the energy saving potential of them is
fitness evaluation to search for the optimal solutions for the worthy of attention (Tan et al., 2014). Therefore, the building per-
given objectives. When the optimization process starts, the GA formance of a classroom is a valuable research area. To illustrate
will keep looking for optimal solutions until the criteria is met. the entire process and the optimization results, a typical type of
We note that the optimal solutions will be quickly calculated classroom in Guangzhou was selected as the object of the case
within a few seconds to a few minutes because the optimization study.
is based on the predictions of the ANN, thus eliminating the time- Guangzhou is in southern China and has a hot and humid
consuming simulation process. In addition, for the optimization of climate. This climate consists of hot summers with intense solar
a certain type of space under a certain climatic condition, the first radiation and warm winters (Figs. 4, 5). Due to the characteristics
two steps only need to be performed one time. Once the ANN is of the climate, the design of the local buildings focuses on ven-
established, it can be directly reused for other optimization tasks tilation and heat dissipation. A classroom with a one-sided open
of the same object, which means that for a specific design task, corridor is a very common space form in local schools. A total of
accurate and direct design guidance can be obtained in a moment. 30 design parameters were used to define a space such as that
The optimization process requires the input of design param- in this study (Fig. 6). The information of the design parameters
eters that can completely define the target space. To enhance used to generate the basic database is listed in Table 1. The char-
the practical value of the research, the design constraints should acteristics of the walls and windows were related to each other.
be set as boundary conditions of parameters in the optimization. Dozens of common types of wall and window properties were
There are two kinds of design constraints. One is the value range used as basic data, and we randomly changed the value of each
of each design parameter, which is relatively simple. The other attribute up or down by 10% as the simulation input information
is the relative relationship between design parameters. For ex- in each simulation case so that the input data were more uniform
ample, when the scope of the area of the space is given, it means within the range of values, which was also in line with reality.
that the relative relationship between the length and width of the The various properties of a wall could be altered by modifying
985
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 3. Toolset.

Fig. 5. Daily average radiation.

Fig. 4. Monthly average temperature.

The useful daylight illuminance (UDI) model was adopted as

the type of brick and the thickness of the thermal insulation layer. the index to evaluate the indoor daylight environment (Nabil and
The interface between the classroom and other indoor spaces was Mardaljevic, 2005, 2006). The UDI illustrates the percentage of
set to be adiabatic. The other important simulation model settings occupied hours of daytime that the illuminance at a certain test
were set according to the local situation, as shown in Table 2.
point in space is within the target illuminance range when only
under the influence of daylight. The upper limit of the target
3.2. Objectives
illumination range was set to 2000 lux and the lower limit was
100 lux. Multiple test points were arranged in a 2 m by 2 m grid
There were three optimization objectives in the case study,
in the classroom. The average value of all the test points was
which were the thermal comfort, the visual comfort and the
calculated to describe the daylight environment of the room.
total energy consumption. The calculation of the total energy
consumption, which is the sum of the energy consumption of
the air conditioning system and the energy consumption of the 3.3. Optimization constraints
lighting system, is relatively simple.
To save investment and energy consumption, classrooms in As previously mentioned, it is necessary to set the constraints
the hot and humid climate of southern China are equipped with of the design parameters during optimization to get results that
cooling only air conditioning systems. Therefore, the thermal meet the actual needs. Three design scenarios and the corre-
comfort of the classroom under natural ventilation in winter de- sponding constraints were assumed in this study. Building per-
serves attention. The predicted mean vote (PMV) model (Fanger, formance optimization will be separately conducted in these
1970) is widely used to evaluate the indoor thermal environment. scenarios. The ranges of the design parameters are shown in
The PMV value ranges from - 3 to 3 and is used to express people’s Table 3.
feelings from extremely cold to extremely hot. A PMV value from The orientation of the outside window of case 1 was set
−0.85 to 0.85 corresponds to an 80% acceptability of the thermal to south or north, that of case 2 was set to the east or west
comfort (ASHRAE, 2017). The number of hours where the PMV and that of case 3 faced south. For the geometry, there was no
value of the classroom was below −0.85 was used as one of the addition limit on case 1. The lower limit of the geometric design
optimization objectives. parameter’s value range of case 2 increased. In contrast, in case
986
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 6. Typical classroom in southern China.

Table 1
Design parameters.
Design parameter Range Units Supplement
x1 : orientation [0, 360] deg resolution: 15
x2 : space length [8.0, 18.0] m
x3 : space width [6.0, 12.0] m
x4 : space height [3.5, 5.5] m
x5 : corridor width [1.5, 4.0] m
x6 : shading device 0, 1 – 0-no and 1-yes
x7 : shading device type 1, 2, 3 – 1-horizontal, 2-vertical, and 3-egg crate
x8 : shading device dimension [0.0, 1.2] m
x9 : outdoor sidewall brick conductivity [0.17, 1.79] W/m − K
x10 : outdoor sidewall brick density [540, 2530] kg/m3 The properties of brick should be set as properties of a certain brick.
x11 : outdoor sidewall brick specific heat [756, 1100] J/kg-K
x12 : outdoor sidewall brick thickness [0.2, 0.4] m
x13 : outdoor sidewall thermal insulation 0, 1 0-no and 1-yes
x14 : outdoor sidewall thermal insulation thickness [0.01, 0.12] m
x15 : outdoor sidewall solar absorptance [0.1, 0.9]
x16 : corridor sidewall brick conductivity [0.17, 1.79] W /m − K
x17 : corridor sidewall density [540, 2530] kg/m3 The properties of brick should be set as properties of a certain brick.
x18 : corridor sidewall specific heat [756, 1100] J/kg-K
x19 : corridor sidewall brick thickness [0.2, 0.4] m
x20 : corridor sidewall thermal insulation 0, 1 0-no and 1-yes
x21 : corridor sidewall thermal insulation thickness [0.01, 0.12] m
x22 : corridor sidewall solar absorptance [0.1, 0.9] – –
x23 : outdoor side window-to-wall ratio [0.1, 0.7] %
x24 : outdoor side window U-value [0.76, 3.79] W/m2 -K
x25 : outdoor side window SHGC (solar heat gain coefficient) [0.19, 0.83] – The properties of glazing should be set as properties of a certain glazing.
x26 : outdoor side window VT (visible transmittance) [0.37, 0.94] –
x27 : corridor side window-to-wall ratio [0.1, 0.6] %
x28 : corridor side window U-value [0.76,3.79] W/m2 -K
x29 : corridor side window SHGC [0.19, 0.83] – The properties of glazing should be set as properties of a certain glazing.
x30 : corridor side window VT [0.37, 0.94] –

Table 2
Important model settings.
Settings value
Number of people per area 0.6 people/m2
Cooling setpoint 26 ◦ C
Daylight illuminance setpoint 300 lux
Light density per area 9 W/m2
Occupancy period 8:00 to 21:00 on weekdays
Vacation period Mid-January to mid-February and mid-July to late August
Air conditioning system operating period April to October

3, the upper limit of the value was reduced. Case 2 was installed limitations were placed on the WWRs of all cases. In addition,
shading devices. There were further restrictions on the values there were six choices for the types of bricks and windows. The
of the wall solar absorption of case 2 and case 3, and special detailed information is given in Table 4 and Table 5, respectively.
987
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Table 3
Ranges of the design parameters.
Design parameter Case 1 Case 2 Case 3
x1 : orientation (outside window) (north: 0 degrees) [−30, 30] [150, 210] [75, 105] [255, 285] [−45, 45]
x2 : space length [8.0, 18.0] [12.0, 18.0] [8.0, 12.0]
x3 : space width [6.0, 12.0] [8.0, 12.0] [6.0, 10.5]
x4 : space height [3.5, 5.5] [4.2, 5.5] [3.5, 4.2]
x5 : corridor width [1.5, 4.0] [2.5, 4.0] [1.5, 2.5]
x6 : shading device 0, 1 1 0, 1
x7 : shading device type 1, 2, 3 1, 2, 3 1, 2, 3
x8 : shading device dimension [0, 1.2] [0.5, 1.2] [0, 1.2]
x9 : outdoor sidewall brick conductivity; x10 : outdoor sidewall Table 4
brick density; x11 : outdoor sidewall brick specific heat
x12 : outdoor sidewall brick thickness [0.2, 0.4] [0.2, 0.4] [0.2, 0.4]
x13 : outdoor sidewall thermal insulation 0, 1 0, 1 0, 1
x14 : outdoor sidewall thermal insulation thickness [0, 0.12] [0, 0.12] [0, 0.12]
x15 : outdoor sidewall solar absorptance [0.1, 0.9] [0.7, 0.9] [0.3, 0.75]
x16 : corridor sidewall brick conductivity; x17 : corridor sidewall Table 4
density; x18 : corridor sidewall specific heat
x19 : corridor sidewall brick thickness [0.2, 0.4] [0.2, 0.4] [0.2, 0.4]
x20 : corridor sidewall thermal insulation 0, 1 0, 1 0, 1
x21 : corridor sidewall thermal insulation thickness [0, 0.12] [0, 0.12] [0, 0.12]
x22 : corridor sidewall solar absorptance [0.1, 0.9] [0.7, 0.9] [0.3, 0.75]
x23 : outdoor side window-to-wall ratio [0.3, 0.65] [0.5, 0.65] [0.35, 0.6]
x24 : outdoor side window U-value; x25 : outdoor side window Table 5
SHGC; x26 : outdoor side window VT
x27 : corridor side window-to-wall ratio [0.2, 0.6] [0.4, 0.6] [0.25, 0.5]
x28 : corridor side window U-value; x29 : corridor side window Table 5
SHGC; x30 : corridor side window VT

Table 4
Wall brick options.
Options Brick type Conductivity (W/m-K) Density (kg/m3 ) Specific heat (J/kg-K)
Option 1 Clay brick 0.72 1920 840
Option 2 Aerated clay brick 0.3 1000 840
Option 3 Concrete brick (medium) 0.51 1400 1000
Option 4 Concrete brick (light) 0.19 600 1000
Option 5 Concrete hollow brick (heavy) 1.35 1220 840
Option 6 Concrete hollow brick (light) 0.76 780 840

In actual design tasks, in addition to the respective value of training iterations was set to 10,000. All the test ANN mod-
ranges of the parameters, there are usually some restricted re- els reached a steady state before the end of training, for the
lationships between design parameters. Therefore, the relative loss metrics were stabilized. The R-squared values of different
relationships of certain parameters, which are shown in Table 6, ANN models on the test set are shown in Table 7. Except for
were regarded as the prerequisite for searching for optimal so- the R-squared value of energy demand, the R-squared values of
lutions. The areas of case 2 and case 3 were limited to a certain thermal environment and daylight environment experience an
range. Considering the area of the façade occupied by the beam, obvious increase as the complexity of the ANN model increases.
sill, etc., the upper limit of the WWR varies with the length and The ANN models with the best predictive ability were used for
height of the space. In all cases, the relative relationship between optimization. The evaluation indicators, including the MAE (mean
the WWR and the geometry of facade was considered. In case 3, absolute error), RMSE (root mean squared error) and R-squared,
the width of the entire building was controlled below a certain of the selected ANNs corresponding to the different data sets are
value. listed in Table 8. Compared with the target data, the error of the
prediction data is within an acceptable range, which indicates
4. Results that the ANNs are feasible to be used as the surrogate models.
To visually demonstrate the generalization ability of the ANN
4.1. ANN training and validation model, 100 samples were randomly selected from the test set to
perform regression fitting on the results of the ANN prediction
More than 50,000 cases were involved in the training and and the simulation output. As illustrated in Fig. 7, there is a good
validation of the ANN. All design parameters mentioned above agreement between the prediction values and target values for
were included in the calculation of the energy consumption and all the regression coefficients, and the R-squared values are very
thermal environment, and some of them were related to the close to 1.
daylight environment. Thus, two ANNs were established. One
was established for the energy consumption and thermal comfort 4.2. Results and discussion
prediction, and the other was established for the daylight comfort
prediction. To reduce the impact of the difference in the magni- For each case, the optimization process ran for 50 generations
tude of the data on the training convergence, the input and output with 400 different solutions. The CPU used for the simulation and
of the original data were scaled. All output data are divided by optimization was an Intel Core i7-8750H @ 2.20 GHz. The three
multiples of ten to keep their values between one and two digits. optimization processes were all completed in 350 s, while the
We compared the performance of different ANN models to time required for a single simulation was approximately 45 s.
find suitable ones for this task. The upper limit of the number Compared with the traditional optimization method based on
988
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 8. Optimization results.

Table 5
Window options.
Options Window type U-value SHGC VT
Option 1 Single glazing 3.437 0.496 0.637
Option 2 Single glazing 3.201 0.361 0.482
Option 3 Double glazing 2.622 0.754 0.809
Option 4 Double glazing 1.845 0.328 0.539
Fig. 7. Comparison between the ANN predicted and simulation results.
Option 5 Triple glazing 1.403 0.305 0.601
Option 6 Triple glazing 0.930 0.268 0.456

simulation, the optimization calculation was accelerated by ap-

proximately 2570 times. Most of the regular spaces in the build- numbers of parameters using the traditional simulation-based
ing were defined by many design parameters. The greater the method was limited because of the running time (Wang et al.,
number of design parameters, the greater the number of searches 2005). The increases in computing speeds made it possible to
that are required to find the optimal solutions in the optimization. quickly solve complex optimization problems, thus enhancing the
The ability to solve optimization problems that involve large practicality of the research.
989
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Table 6
Restrictions on design parameters.
Restrictions Case 1 Case 2 Case 3
Area – x2 * x3 >100 m2 60 m2 < x2 * x3 < 120 m2
Outdoor side: [(x2 −1.2) * (x4 − 1.6)] / (x2 * x4 ) ≥ x23
Upper limit of the window-to-wall ratio
Corridor side: [(x2 −4) * (x4 − 1.6)] / (x2 * x4 ) ≥ x27
Building width – – x3 + x5 < 12 m

Table 7
R-squared values on test set of different ANN models.
Option 1 Option 2 Option 3 Option 4 Option 5 Option 6
ANN for predicting Neurons in hidden layer 1 16 32 64 128 256 512
energy demand Neurons in hidden layer 2 – – – – – –
and thermal R-squared (energy) 0.991 0.997 0.999 0.999 0.999 0.999
environment R-squared (thermal) 0.746 0.876 0.930 0.944 0.964 0.965
Option 7 Option 8 Option 9 Option 10 Option 11 Option 12
Neurons in hidden layer 1 16 32 64 128 256 512
Neurons in hidden layer 2 4 8 16 32 64 128
R-squared (energy) 0.995 0.997 0.999 0.999 0.999 0.999
R-squared (thermal) 0.721 0.872 0.936 0.954 0.962 0.967
ANN for Option 1 Option 2 Option 3 Option 4 Option 5 Option 6
predicting Neurons in hidden layer 1 16 32 64 128 256 512
daylight Neurons in hidden layer 2 – – – – – –
environment R-squared 0.942 0.973 0.986 0.988 0.993 0.993
Option 7 Option 8 Option 9 Option 10 Option 11 Option 12
Neurons in hidden layer 1 16 32 64 128 256 512
Neurons in hidden layer 2 4 8 16 32 64 128
R-squared 0.973 0.980 0.988 0.992 0.994 0.995

Table 8
Evaluation of selected ANN models.
Objective Data set MAE RMSE R-squared
Training set 46.7 72.0 1.000
Energy demand (kwh) Validation set 54.5 94.2 0.999
Test set 55.3 95.7 0.999
Training set 2.96 3.80 0.972
Thermal environment (hours of discomfort) Validation set 3.20 4.12 0.966
Test set 3.17 4.08 0.967
Training set 0.38 0.52 0.996
Daylight environment (%) Validation set 0.45 0.58 0.995
Test set 0.45 0.62 0.995

To illustrate the optimization results, an integrated solution 8.4%, respectively; and the UDI (100–2000 lux) of US increases
and a reference solution were given for each case. We assumed by 25.3%, 29.4% and 14.2%, respectively. Compared with RS, the
that the three optimization goals were equally important, we cal- three objectives of IS in case 1 are optimized by 61.1%, −3.3%,
culated a score for each optimal solution, and we picked the one and 16.1%, respectively. These values for case 2 are 35.9%, −3.8%,
with a smallest score as the integrated solution. The calculation and 24.0%, respectively; and for case 3, they are 29.6%, 1.5%, and
method is as follows: 11.5%, respectively. Therefore, the objectives of these three cases
√
∑ increase by 24.6%, 18.7% and 14.2%, respectively, on average. The
S= ((yi − yibest )/ |yi |min )2 energy consumption of ES is much smaller than that of DS in all
the cases. In all cases, the values of energy consumption of ES are
where yi represents the value of a certain objective; yibest is less than half of that of DS. There are 612.7, 572.9, and 590.0 h
the value of the optimal solution for a certain objective, which of discomfort in ES, respectively; while for DS, there are 470.3,
could be the maximum value or minimum value; and |yi |min is
515.8, and 525.8 h of discomfort, respectively. Except for RS, the
the minimum absolute value of a certain objective. The value
UDI (100–2000 lux) value of DS is the lowest in all the cases at
of each continuous parameter of the reference solution was set
81.8%, 74.6% and 84.6%, respectively; which are 0.88, 0.86 and
to the middle of the value range. The Pareto front of each case
0.91 of those in US, respectively.
is presented in Fig. 8. The solutions of the minimum energy
The building performance results differed greatly due to the
consumption (ES), minimum hours of discomfort (DS), maximum
DUI (100–2000 lux) solution (US) and integrated solution (IS) different choices of design parameters. Taking the spatial geome-
are marked on the graph, and the corresponding data are listed try as example, regarding the orientation, in case 1, ES and US face
in Table 9. To verify the effectiveness of the optimization re- north and DS faces south. The ES and US of case 2 face the east
sults, results of ten optimal solutions randomly chosen from each and DS faces the west. In terms of the other design parameters,
case were validated by simulation, and the results are presented some choices are concentrated. For instance, ES tends to have
in Fig. 9. The margin of error between the simulation result a small space area, and DS prefers a wide space. Meanwhile,
and the ANN prediction result is acceptable, indicating that the some options, such as brick and window types, are scattered. To
optimization results of this study are reliable. provide a better display and analyse the optimized results, the
In the three cases, compared with RS, the energy consumption nondominated solutions of the three cases were summarized.
of ES decreases by 63.5%, 37.2% and 31.7% of RS, respectively; the Fig. 10 presents the orientation choices. In case 1, the orienta-
number of hours of discomfort of DS drops by 10.3%, 4.8% and tion of more than 250 solutions is one of 165 degrees, 180 degrees
990
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 9. Validation of optimal solutions.

Fig. 10. Orientation of nondominated solutions.

and 195 degrees, and approximately 120 solutions face 345 de- of the space height in case 2 are all close to the lower limit.
grees, indicating that the south orientation is more popular. For The space length, width of case 3 are basically evenly distributed
case 2, there are three main options, namely 105 degrees and within the limits, and more than half of the heights are very
255 degrees with approximately 210 solutions and 150 solutions, close to 3.5 m, which is the lower limit. For the width of the
respectively, and 285 degrees with approximately 40 solutions. corridor, in all the cases, a value that is not far from the lower
The solutions for case 3 prefer a northeast orientation rather than limit is more often selected. Fig. 12 shows the shading device
a northwest orientation, for almost all the direction selections distributions. In case 1, there are 218 nondominated solutions
are distributed towards the northeast and most solutions are 330 without shading devices, and the vertical shading and egg crate
degrees or 345 degrees. shading device appears 96 times and 86 times, respectively. The
The geometric distributions are shown in Fig. 11. Case 1 and egg crate shading device and vertical shading device are selected
case 2 tend to have a small value of space length and height. The 233 times and 167 times in case 2, respectively. There are 220
maximum value of length in case 1 is below 8 m. The choices nondominated solutions of case 3 choose the egg crate shading,
991
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 11. Space geometry of nondominated solutions.

Fig. 12. Shading device of nondominated solutions.

Fig. 13. Wall brick of nondominated solutions.

and the rest are without shading devices. Regarding the shading appears a certain number of times. Among the brick types, option
device width, most solutions of case 1 and case 3 are beyond 4 accounts for the largest proportion. Except for the outdoor
0.8 m. sidewall of case 1, in the rest of the wall constructions, more than
Figs. 13 to 15 explain the wall construction selection of non- half of the nondominated solutions choose the brick of option 4.
dominated solutions. According to Fig. 13, each brick type option Regarding the brick thickness, except for the outdoor sidewall
992
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Fig. 14. Wall thermal insulation of nondominated solutions.

brick of case 1 and the corridor sidewall brick of case 3, the

medians of thickness of the rest wall brick are above 0.35 m.
Fig. 14 illustrates the distributions of the thermal insulation. In
case 1 and case 2, more than 300 nondominated solutions and
200 nondominated solutions do not use thermal insulation in
the construction of outdoor sidewall. For the corridor sidewall
construction, 310 solutions of case 1 and 381 solutions of case
2 are with thermal insulation. In case 3, 362 solutions suggest
installing an insulation layer in the construction of the outdoor
sidewall and 206 solutions suggest not installing that in the
construction of corridor sidewall. Regarding the distribution of
the insulation thickness, all the medians are above 0.06 m. The
corridor sidewall constructions tend to have thicker layers of
thermal insulation than the outdoor sidewall constructions, and
the medians of the thermal insulation thickness of the corridor
sidewall constructions are above 0.1 m in all the cases. Regarding
the solar absorptance (Fig. 15), in case 1, the medians of solar
absorptance of the outdoor sidewall and corridor sidewall are
approximately 0.8 and 0.5, and the lower quartiles of that are Fig. 15. Wall solar absorptance of nondominated solutions.
approximately 0.4 and 0.2, respectively. In case 2 and case 3,
the solar absorptance values of outdoor sidewall and corridor
sidewall in most nondominated solutions are close to the upper research objects should be carefully set to cover various situations
limit. that usually occur in reality. In addition, although the building
Fig. 16 shows the selections of nondominated solutions with simulation tools are not limited to EnergyPlus and Radiance, they
respect to windows. In terms of the window type of the outdoor must be verified through previous experiments to prove that
sidewall selected most frequently by nondominated solutions, the simulation results are consistent with the actual situation.
case 1 chooses option 1 at 106 times, case 2 chooses option Moreover, the accuracy of the ANN prediction results affects the
1 and option 2 for 106 times and 108 times, and option 2 is reliability of the optimization results. The predictive ability of
most popular in case 3 with a frequency of 160 times. Regarding an ANN is related to factors such as the number of training
the corridor sidewall window type, from case 1 to case 3, the samples and the ANN parameter settings. Excessive errors in
options with the most occurrences are 6, 5, and 1, respectively, the prediction results will also make the optimization results
which appear 134 times, 171 times and 197 times, respectively. inaccurate. Therefore, the prediction results of ANNs must be
Regarding the WWR, the nondominated solutions tend to have a verified before they can be used in the optimization instead of
relatively small value on both sides of the space in all the cases, building simulation.
for the maximum value of all sets of data is less than the average
of its range. 5. Conclusion
In general, this method is capable of quickly completing the
design optimization work of a regular architectural space. How- The goal of reducing building energy demand and improving
ever, there are several prerequisites for obtaining accurate design indoor comfort can be achieved by optimizing building design.
guidance using this method. As we all know, weather has a Different objectives of building performance sometimes are con-
great influence on the building simulation and optimization re- flicts. To improve the overall benefits of the building and seek
sults. Performing building simulations with weather data that a balance between the building performance objectives, it is
can represent the typical local climate to generate a database is important to provide architects with fast and accurate design
the basic condition to ensure the accuracy of the optimization guidance. This paper developed a comprehensive method for
results. In addition, when constructing a database for building optimizing the design of a regular architectural space to improve
simulations, the design parameters and parameter ranges of the building performance. The research aimed to provide architects
993
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Table 9
Comparison of the solutions and results (ES: minimum energy consumption solution, DS: minimum hours of discomfort solution, US: maximum UDI (100–2000 lux)
solution, IS: integrated solution, and RS: reference solution).
Case 1 Case 2 Case 3
ES DS US IS RS ES DS US IS RS ES DS US IS RS
x1 : orientation (outside window) 345 165 345 180 180 255 105 285 285 270 345 330 345 345 0
(north: 0 degrees)
x2 : space length 8.0 8.5 12.3 8.0 13.0 12.2 17.9 12.4 12 15.0 10.0 11.1 11.5 8.7 10.0
x3 : space width 6.0 12.0 6.9 6.0 9.0 8.2 11.8 11.8 8.4 10.0 6.0 10.4 8.4 6.9 8.2
x4 : space height 3.5 3.5 3.5 3.5 4.5 4.2 4.2 4.2 4.2 4.8 3.6 3.5 3.5 3.5 3.8
x5 : corridor width 3.5 1.5 3.5 1.5 2.7 4 2.5 3.9 2.6 3.2 1.8 1.5 2.3 1.5 2.0
x6 : shading device 1 0 1 1 1 1 1 1 1 1 1 0 1 1 1
x7 : shading device type 3 – 3 3 1 3 2 3 3 1 3 – 3 3 1
x8 : shading device dimension 1.2 – 1.2 1.15 0.50 1.2 0.55 1.2 1.2 0.50 1.15 – 1.15 1.2 0.50
Outside wall option (x9 : outdoor 1 5 1 6 2 4 4 2 3 2 6 4 6 4 2
sidewall brick conductivity, x10 :
outdoor sidewall brick density, and
x11 : outdoor sidewall brick specific
heat)
x12 : outdoor sidewall brick thickness 0.28 0.30 0.31 0.31 0.30 0.39 0.40 0.38 0.39 0.30 0.21 0.27 0.30 0.36 0.30
x13 : outdoor sidewall thermal 1 1 1 1 1 1 1 1 0 1 1 0 0 0 1
insulation
x14 : outdoor sidewall thermal 0.09 0.09 0.12 0.05 0.06 0.08 0.1 0.08 0.08 0.06 0.05 – – – 0.06
insulation thickness
x15 : outdoor sidewall solar 0.26 0.87 0.34 0.82 0.50 0.81 0.90 0.88 0.87 0.80 0.34 0.75 0.33 0.75 0.52
absorptance
Corridor wall option (x16 : corridor 5 4 4 4 2 5 4 4 4 2 5 4 5 4 2
sidewall brick conductivity, x17 :
corridor sidewall density, and x18 :
corridor sidewall specific heat)
x19 : corridor sidewall brick thickness 0.21 0.35 0.37 0.20 0.30 0.37 0.39 0.36 0.25 0.30 0.31 0.31 0.30 0.33 0.30
x20 : corridor sidewall thermal 1 0 1 0 1 1 0 1 0 1 1 1 0 0 1
insulation
x21 : corridor sidewall thermal 0.1 – 0.05 – 0.06 0.12 – 0.12 – 0.06 0.11 0.11 – – 0.06
insulation thickness
x22 : corridor sidewall solar 0.39 0.75 0.40 0.76 0.50 0.70 0.85 0.86 0.84 0.80 0.44 0.74 0.72 0.74 0.52
absorptance
x23 : outdoor side window-to-wall 0.32 0.44 0.34 0.31 0.48 0.50 0.54 0.50 0.50 0.58 0.37 0.37 0.35 0.35 0.47
ratio
Outdoor window option (x24 : outdoor 3 2 1 1 2 5 2 5 1 2 2 5 2 2 2
side window U-value, x25 : outdoor
side window SHGC, and x26 : outdoor
side window VT)
x27 : corridor side window-to-wall 0.20 0.20 0.36 0.20 0.40 0.41 0.41 0.42 0.40 0.50 0.33 0.25 0.35 0.29 0.37
ratio
Corridor window option (x28 : corridor 3 2 5 3 2 5 5 1 1 2 1 1 2 1 2
side window U-value, x29 : corridor
side window SHGC, and x30 : corridor
side window VT)
y1 : energy consumption (kwh) 4,472.8 10,832.1 8,184.9 4,765.5 12,244.5 10,257.4 22,593.4 14,979.0 10,468.3 16,337.2 5,774.8 11,874.0 9,494.0 5,950.2 8,451.7
y2 : hours of discomfort 612.7 470.3 601.8 541.6 524.5 572.9 515.8 568.4 562.6 541.9 590.0 525.8 582.3 565.3 574.0
y3 : UDI (100–2000 lux) (%) 88.7 81.8 92.5 85.7 73.8 83.9 74.6 86.8 83.2 67.1 90.1 84.6 92.5 90.3 81.0

Fig. 16. Window of nondominated solutions.

with robust and accurate design references when conducting the previously generated database to replace the time-consuming
design tasks. To enhance the practicability of the research, we building simulation. The last step is to take the ANN models as
emphasize the following points: 1. taking a certain type of regular the fitness evaluation and perform multi-objective optimization
architectural space, which can be described by limited design based on the actual design constraints to obtain the optimal
parameters under certain climatic conditions, as the research solutions. To optimize a certain type of space under a certain
object; 2. considering the actual constraints in the optimization climatic condition, the first two steps only need to be performed
process; and 3. completing the multi-objective optimization and one time. Once the ANN is established, it can be directly reused
obtaining the optimal solutions in a short time. for other optimization tasks of the same object, which means that
The entire optimization process is divided into three steps. The for a specific design task, accurate and direct design guidance can
first step is to build a database of the design and the correspond- be obtained in a moment.
ing building performance of the research object by generating the Classroom is a valuable research object. On one hand, the
research objects randomly and performing building simulations classroom is closely related to the health and performance of
using them. The second step is to train ANN models based on students. On the other hand, the energy saving potential of the
994
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

classroom is great, because classroom is a type of space that Bre, F., Silva, A.S., Ghisi, E., Fachinotti, V.D., 2016. Residential building design
exists in large numbers. To demonstrate the optimization process, optimisation using sensitivity analysis and genetic algorithm. Energy Build.
133, 853–866.
a common type of classroom space in a hot and humid area
Cao, X., Dai, X., Liu, J., 2016. Building energy-consumption status worldwide and
was selected as a case study. The target space is defined by the state-of-the-art technologies for zero-energy buildings during the past
30 design parameters, including aspects such as the spatial ge- decade. Energy Build. 128, 198–213.
ometry, shading device, wall construction and window material. Castilla, M., Álvarez, J.D., Ortega, M.G., Arahal, M.R., 2013. Neural network and
The optimization objectives are the energy consumption, winter polynomial approximated thermal comfort models for HVAC systems. Build.
Environ. 59, 107–115.
thermal comfort and daylight comfort. More than 50,000 pieces Chen, X., Yang, H., Zhang, W., 2018. Simulation-based approach to optimize
of data were used to train the ANN models. Multiple ANN models passively designed buildings: A case study on a typical architectural form
were trained, and the best performing ones were used in the in hot and humid climates. Renew. Sustain. Energy Rev. 82, 1712–1725.
optimization. As a result, the errors of the selected ANNs are Deb, Kalyanmoy, 2002. A fast and elitist multiobjective genetic algorithm:
NSGA-II - evolutionary computation. IEEE Trans..
very small relative to the magnitude of the data. 100 samples
Delgarm, N., Sajadi, B., Delgarm, S., Kowsary, F., 2016. A novel approach for the
were randomly selected from the test set to perform regression simulation-based optimization of the buildings energy consumption using
fitting on the results of the ANN prediction and simulation output NSGA-II: Case study in Iran. Energy Build. 127, 552–560.
of the three objectives, and the R-squared values were 1.000, Dong, B., Cao, C., Lee, S.E., 2005. Applying support vector machines to predict
0.963 and 0.996, respectively, suggesting that the errors of the building energy consumption in tropical region. Energy Build. 37, 545–553.
Eguía, P., Granada, E., Alonso, J.M., Arce, E., Saavedra, A., 2016. Weather datasets
prediction results are within an acceptable range. To imitate re-
generated using kriging techniques to calibrate building thermal simulations
alistic design tasks, three design scenarios and the corresponding with TRNSYS. J. Build. Eng. 7, 78–91.
constraints are assumed in the optimization. For each case, the Eisenhower, B., O’Neill, Z., Narayanan, S., Fonoberov, V.A., Mezić, I., 2012. A
optimization process ran for 50 generations with 400 different methodology for meta-model based optimization in building energy models.
solutions. All the three optimizations finished in 350 s. Compared Energy Build. 47, 292–301.
Fang, Y., Cho, S., 2019. Design optimization of building geometry and fenestration
with the traditional optimization method based on simulation, for daylighting and energy performance. Sol. Energy 191, 7–18.
the optimization calculation is accelerated by approximately 2570 Fanger, P.O., 1970. Thermal Comfort. Analysis and Applications in Environmental
times. Some of the optimal solutions were validated through Engineering. Danish Technical Press, Copenhagen.
simulation. We defined an integrated solution and a reference Ferrara, M., Sirombo, E., Fabrizio, E., Filippi, M., 2015. Comfort filters in a total
energy demand optimization method for the passive design of a building.
solution for each case. After optimization, compared to the ref-
Energy Procedia 83, 418–427.
erence solutions, the objectives of the integrated solutions in the Gou, S., Nik, V.M., Scartezzini, J.-L., Zhao, Q., Li, Z., 2018. Passive design
three cases have been improved by 24.6%, 18.7% and 14.2% on optimization of newly-built residential buildings in shanghai for improving
average, respectively, indicating that this method is feasible and indoor thermal comfort while reducing building energy demand. Energy
effective to improve building design in actual tasks. Build. 169, 484–506.
Han, Y., Yu, H., Sun, C., 2017. Simulation-based multiobjective optimization of
timber-glass residential buildings in severe cold regions. Sustainability 9,
CRediT authorship contribution statement 2353.
Harkouss, F., Fardoun, F., Biwole, P.H., 2018. Passive design optimization of low
Yukai Zou: Conceptualization, Methodology, Writing - origi- energy buildings in different climates. Energy 165, 591–613.
Hensen, J., Lamberts, R., 2011. Building Performance Simulation for Design and
nal draft, Software, Formal analysis. Qiaosheng Zhan: Resources,
Operation. Spon Press, London, New York, p. 507, xxiv.
Writing - review & editing, Validation, Visualization. Ke Xiang: Heo, Y., Zavala, V.M., 2012. Gaussian process modeling for measurement and
Conceptualization, Supervision, Project administration, Funding verification of building energy savings. Energy Build. 53, 7–18.
acquisition. Hoppe, Peter, 2002. Different aspects of assessing indoor and outdoor thermal
comfort.
Huang, Z., Zheng, H., Guo, L., Mo, D., 2020. Influence of the position of artificial
Declaration of competing interest
boundary on computation accuracy of conjugated infinite element for a finite
length cylindrical shell. Acoust. Aust. 48, 287–294.
The authors declare that they have no known competing finan- Kheiri, F., 2018. A review on optimization methods applied in energy-efficient
cial interests or personal relationships that could have appeared building geometry and envelope design. Renew. Sustain. Energy Rev. 92,
to influence the work reported in this paper. 897–920.
Khoroshiltseva, M., Slanzi, D., Poli, I., 2016. A Pareto-based multi-objective
optimization algorithm to design energy-efficient shading devices. Appl.
Acknowledgements Energy 184, 1400–1410.
Kwok, Y.T., Lai, A.K.L., Lau, K.K.-L., Chan, P.W., Lavafpour, Y., Ho, J.C.K., Ng, E.Y.Y.,
This research was funded by the National Key Research and 2017. Thermal comfort and energy performance of public rental housing
under typical and near-extreme weather conditions in Hong Kong. Energy
Development Program of China (Grant No. 2017YFC0702309).
Build. 156, 390–403.
Li, Y.F., Ng, S.H., Xie, M., Goh, T.N., 2010. A systematic comparison of metamod-
References eling techniques for simulation optimization in Decision Support Systems.
Appl. Soft Comput. 10, 1257–1273.
Alhuwayil, W.K., Abdul Mujeebu, M., Algarny, A.M.M., 2019. Impact of external Liu, J., Liu, Y., Wang, X., 2020a. An environmental assessment model of con-
shading strategy on energy performance of multi-story hotel building in struction and demolition waste based on system dynamics: a case study in
hot-humid climate. Energy 169, 1166–1174. Guangzhou. Environ. Sci. Pollut. Res. Int. 27, 37237–37259.
Amasyali, K., El-Gohary, N.M., 2018. A review of data-driven building energy Liu, J., Yi, Y., Wang, X., 2020b. Exploring factors influencing construction waste
consumption prediction studies. Renew. Sustain. Energy Rev. 81, 1192–1205. reduction: A structural equation modeling approach. J. Cleaner Prod. 276,
Ascione, F., Bianco, N., Mauro, G.M., Vanoli, G.P., 2019. A new comprehensive 123185.
framework for the multi-objective optimization of building energy design: Machairas, V., Tsangrassoulis, A., Axarli, K., 2014. Algorithms for optimization of
Harlequin. Appl. Energy 241, 331–361. building design: A review. Renew. Sustain. Energy Rev. 31, 101–112.
ASHRAE, 2017. Standard 55–2017 Thermal Environmental Conditions for Human Mangkuto, R.A., Rohmah, M., Asri, A.D., 2016. Design optimisation for window
Occupancy. ASHRAE, Atlanta. size, orientation, and wall reflectance with regard to various daylight metrics
Ashrafian, T., Moazzen, N., 2019. The impact of glazing ratio and window and lighting energy demand: A case study of buildings in the tropics. Appl.
configuration on occupants’ comfort and energy demand: The case study Energy 164, 211–219.
of a school building in Eskisehir, Turkey. Sustainable Cities Soc. 47, 101483. Méndez Echenagucia, T., Capozzoli, A., Cascone, Y., Sassone, M., 2015. The
Aste, N., Buzzetti, M., Del Pero, C., Leonforte, F., 2018. Glazing’s techno-economic early design stage of a building envelope: Multi-objective search through
performance: A comparison of window features in office buildings in heating, cooling and lighting energy performance analysis. Appl. Energy 154,
different climates. Energy Build. 159, 123–135. 577–591.
Bayoumi, M., 2017. Impacts of window opening grade on improving the energy Mishra, A.K., Ramgopal, M., 2014. Thermal comfort field study in undergraduate
efficiency of a façade in hot climates. Build. Environ. 119, 31–43. laboratories – An analysis of occupant perceptions. Build. Environ. 76, 62–72.

995
Y. Zou, Q. Zhan and K. Xiang Energy Reports 7 (2021) 981–996

Moon, J.W., Yoon, S.-H., Kim, S., 2013. Development of an artificial neural Taveres-Cachat, E., Lobaccaro, G., Goia, F., Chaudhary, G., 2019. A methodol-
network model based thermal control logic for double skin envelopes in ogy to improve the performance of PV integrated shading devices using
winter. Build. Environ. 61, 149–159. multi-objective optimization. Appl. Energy 247, 731–744.
Nabil, A., Mardaljevic, J., 2005. Useful daylight illuminance: a new paradigm for Wang, W., Zmeureanu, R., Rivard, H., 2005. Applying multi-objective ge-
assessing daylight in buildings. Light. Res. Technol. 37, 41–57. netic algorithms in green building design optimization. Build. Environ. 40,
Nabil, A., Mardaljevic, J., 2006. Useful daylight illuminances: A replacement for 1512–1525.
daylight factors. Energy Build. 38, 905–913. Wong, S.L., Wan, K.K.W., Lam, T.N.T., 2010. Artificial neural networks for energy
Negendahl, K., Nielsen, T.R., 2015. Building energy optimization in the early analysis of office buildings with daylighting. Appl. Energy 87, 551–557.
design stages: A simplified method. Energy Build. 105, 88–99. Wu, C., Wu, P., Wang, J., Jiang, R., Chen, M., Wang, X., 2020. Critical review of
Premrov, M., Žegarac Leskovar, V., Mihalič, K., 2016. Influence of the building data-driven decision-making in bridge operation and maintenance. Struct.
shape on the energy performance of timber-glass buildings in different Infrastruct. Eng. 8, 1–24.
climatic conditions. Energy 108, 201–211. Xu, S., Wang, J., Shou, W., Ngo, T., Sadick, A.-M., Wang, X., 2020. Computer vision
Raji, B., Tenpierik, M.J., Bokel, R., van den Dobbelsteen, A., 2020. Natural summer techniques in construction: A critical review. Arch. Comput. Methods Eng. 29
ventilation strategies for energy-saving in high-rise buildings: a case study (211).
in the Netherlands. Int. J. Vent. 19, 25–48. Yang, M.-D., Lin, M.-D., Lin, Y.-H., Tsai, K.-T., 2017. Multiobjective optimization
Romani, Z., Draoui, A., Allard, F., 2015. Metamodeling the heating and cooling design of green building envelope material using a non-dominated sorting
energy needs and simultaneous building envelope optimization for low genetic algorithm. Appl. Therm. Eng. 111, 1255–1264.
energy building design in Morocco. Energy Build. 102, 139–148. Yi, Y.K., 2016. Adaptation of kriging in daylight modeling for energy simulation.
Sadeghifam, A.N., Zahraee, S.M., Meynagh, M.M., Kiani, I., 2015. Combined use Energy Build. 111, 479–496.
of design of experiment and dynamic building simulation in assessment of Yoon, Y.R., Moon, H.J., 2018. Energy consumption model with energy use factors
energy efficiency in tropical residential buildings. Energy Build. 86, 525–533. of tenants in commercial buildings using Gaussian process regression. Energy
Singh, M.K., Kumar, S., Ooka, R., Rijal, H.B., Gupta, G., Kumar, A., 2018. Status Build. 168, 215–224.
of thermal comfort in naturally ventilated classrooms during the summer Yu, W., Li, B., Jia, H., Zhang, M., Wang, Di, 2015. Application of multi-objective
season in the composite climate of India. Build. Environ. 128, 287–304. genetic algorithm to optimize energy efficiency and thermal comfort in
Sun, Y., Wang, J., Wu, J., Shi, W., Ji, D., Wang, X., Zhao, X., 2020. Constraints building design. Energy Build. 88, 135–143.
hindering the development of high-rise modular buildings. Appl. Sci. 10, Zhang, A., Bokel, R., van den Dobbelsteen, A., Sun, Y., Huang, Q., Zhang, Q., 2017.
7159. Optimization of thermal and daylight performance of school buildings based
Tan, H., Chen, S., Shi, Q., Wang, L., 2014. Development of green campus in China. on a multi-objective genetic algorithm in the cold climate of China. Energy
J. Cleaner Prod. 64, 646–653. Build. 139, 371–384.

996