Energy and Buildings 126 (2016) 524534

Contents lists available at ScienceDirect

Energy and Buildings

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

Multi-objective optimization of building envelope design for life cycle

environmental performance
Rahman Azari a, , Samira Garshasbi b , Pegah Amini a , Hazem Rashed-Ali a ,
Yousef Mohammadi c
a
College of Architecture, Construction and Planning, University of Texas at San Antonio, United States
b
Young Researchers Club, Central Tehran Branch, Islamic Azad University, P.O. Box 13185-768, Tehran, Iran
c
Petrochemical Research and Technology Company, National Petrochemical Company, P.O. Box 14358-84711, Tehran, Iran

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

Article history: The building envelope incorporates signicant amount of construction materials and is a key determinant
Received 12 October 2015 of the embodied energy and environmental impacts in buildings. It is also a mediator between indoor
Received in revised form 17 May 2016 and outdoor environmental conditions and has signicant impacts on operational energy use in many
Accepted 18 May 2016
types of buildings.
Available online 20 May 2016
The present article utilizes a multi-objective optimization algorithm to explore optimum building enve-
lope design with respect to energy use and life cycle contribution to the impacts on the environment in a
Keywords:
low-rise ofce building in Seattle, Washington. Design inputs of interest include insulation material, win-
Life cycle assessment (LCA)
Building envelope
dow type, window frame material, wall thermal resistance, and south and north window-to-wall ratios
Optimization (WWR). The simulation tool eQuest 3.65 is used to assess the operational energy use, while Life Cycle
Genetic algorithm Assessment (LCA) methodology and Athena IE are used to estimate the environmental impacts. Also, a
hybrid articial neural network and genetic algorithm approach is used as the optimization technique.
The environmental impact categories of interest within the LCA include: global warming, acidication,
eutrophication, smog formation, and ozone depletion. The results reveal that the optimum design sce-
nario incorporates berglass-framed triple-glazed window, about 60% south WWR, 10% north WWR, and
R-17 insulation.
2016 Elsevier B.V. All rights reserved.

1. Introduction grave, in LCA offers a more informed basis for decision-making,

The contribution of buildings to the overall environmental
impacts of human activities has been well-documented [14,11].
According to the US Energy Information Administration [11], 19%
of the worlds primary energy is consumed in the US. The major
consumers of the total energy in the US, as the EIA data reveals,
include commercial buildings (18.5%), residential buildings (22%),
industrial sector (32%), and transportation (27.5%). Buildings also
contribute 40 percent to carbon dioxide emissions in the states EIA,
2012 and about 66% to generation of non-industrial solid-waste
[14].
Life Cycle Assessment (LCA) methodology has gained increasing
popularity in recent years to assess how buildings or their compo-
nents contribute to the negative impacts on the environment over
their entire life cycle. Considering the entire stages; i.e., cradle to
Corresponding author.
E-mail address: rahman.azari@utsa.edu (R. Azari).
compared to other methodologies that focus on operation phase of
life cycle only and rely on metrics such as operational energy use.
LCA studies typically address all stages of a buildings life cycle and
cover one or more of six impact categories. LCA methodologys prin-
ciples, requirements and guidelines are prescribed by ISO 14040
[21], and ISO 14044 [22]. In 2011, the European Committee for
Standardization developed EN 15978 [12] as the standard for using
LCA in assessment of the environmental performance of buildings.
In addition, International Standards Organization (ISO) provides
metrics and requirements for determining the carbon footprint of
buildings through [23].
Many studies use LCA in assessing the environmental impacts
of buildings. For instance, Kosareo and Ries [26] compare intensive
and extensive green roofs versus conventional roofs with regard to
their impacts on ozone depletion, global warming, acidication and
eutrophication. Pulselli et al. [29] use energy analysis and emergy
evaluation methods to assess environmental costs and benets
of three different types of envelopes (an air cavity wall, a plus-
insulated wall and a ventilated wall). Azari and Kim [4] focus on

http://dx.doi.org/10.1016/j.enbuild.2016.05.054
0378-7788/ 2016 Elsevier B.V. All rights reserved.
 R. Azari et al. / Energy and Buildings 126 (2016) 524534 525

curtain walls and apply a computational process-based environ-

Acidication
mental life cycle assessment (LCA) to compare the effect of change
in mullion materials on curtain walls environmental impacts. Azari
[3] conducts a parametric LCA analysis to examine how the change

of design input values in a limited number of building envelope
Eutrophication congurations would impact the environment.
Dodoo et al. [10] is an example of studies that focus on building
structures. The study uses the consequential-based LCA to compare
three versions of timber structures; i.e., cross-laminated timber,

beam-and-column and modular structures. In an LCA study of an
ofce building, Junnila and Horvath [24] conduct a complete life-
cycle analysis on an ofce building with a service life of 50 years
Energy

and investigate its impacts with regard to acidication, eutrophica-

Use

tion, climate change, etc. They later extend this study and compare
the impacts of ofce buildings in the US and Europe [25]. In Aus-
Warming

tralian context, Treloar et al. [33] and Treloar et al. [34] study the
Global

embodied energy of construction materials (2001a) in a building

and, in another project, measure the embodied energy in several
ofce buildings and study the effect of height and number of oors
Forma-
Smog

on changes in embodied energy. In another study, Yohanis and Nor-

tion

ton Yohanis and Norton (2002) study the variations in life-cycle

embodied and operational energy as well as capital cost as a result

Impact Categories

of change in building parameters.

Toxicity
Human

Finally, Tingley et al. [32] is a study at the scale of construction

materials in which LCA is used to compare three different insula-

tion materials when applied in a typical dwelling. A snapshot of a
Demolition Other

selected number of LCA studies in the eld of built environment is

shown in Table 1.
Recovery

One limitation in many comparative LCA studies is the limited

number of variables and combinations, out of all possibilities, that
are studied. This in turn results in incomprehensive conclusions

&

with regard to optimized LCA. Because of the signicant resources

Operation

that would be needed to analyze all possible scenarios in a com-

& Main-
tenance

prehensive LCA study, computational optimization techniques are

utilized to address the challenge. To consider all possible com-

binations of design inputs and values, an optimization problem
Construction

is started with identication of design inputs and proceeds with

determining risks and constraints, nding the objective function,
setting the minimum and maximum thresholds on design inputs,

choosing an optimization algorithm, and eventually obtaining the

Manufacturing

results; i.e., the optimum solution to optimization problem [9].

The present article tries to use an intelligent optimization algo-
rithm to explore optimum building envelope design combination in
a low-rise ofce building from operational energy use and environ-

mental life-cycle impacts perspectives. The design inputs of interest

LCA stages

include wall thermal resistance (R-value), insulation material, glaz-

Material
Extrac-

ing type, window-to-wall ratio (WWR) in north and south facades,

tion

and frame material. Using these variables, we attempt to nd

the design combination that yields the lowest operational energy

Buildings

use and the least environmental impact. To achieve the objec-

tive, an integrated energy and environmental life-cycle assessment
methodology is used to quantify the environmental impacts asso-

ciated with each design combination. The results are then used
Components, and

in a hybrid articial neural network and genetic algorithm-based

approach to identify the optimum design combination.
Envelope,
ConstructionBuilding

Systems
An overview of some LCA literature.

Subject of LCA

Materials

2. Methods
Reference

A two-phase research methodology was pursued as illustrated

Table 1

[24]
[25]
[26]
[28]
[29]
[31]
[32]
[33]
[34]
[35]
[10]

in Fig. 1. The rst phase is LCA methodology followed by optimiza-

[1]
[3]
[4]

tion efforts.
526 R. Azari et al. / Energy and Buildings 126 (2016) 524534

Fig. 1. Research Methodology.

2.1. LCA methodology Table 2

Construction details of the wall.

LCA is a technique used to assess the environmental impacts Wall

associated with products, projects, processes, and programs [19]. Split-faced brick veneer10 cm (4)
The quantitative technique uses a four-step methodology based on Air gap5 cm (2)
requirements prescribed by ISO 14040 [21]: goal and scope deni- Insulation (varying materials and characteristics)
tion, inventory modeling, impact assessment, and interpretation of Vapor barrier (polyethylene)
Concrete brick10 cm (4)
results.
Gypsum plasterboard1.25 cm (0.5)

2.1.1. Goal and scope denition

Table 3
The rst step in LCA methodology is mainly about establish- Building envelope design inputs and their corresponding values.
ing the functional unit, reference ows, system boundary, and the
Design Inputs Potential Values
impact categories to be covered. The functional unit (FU) is a mea-
sure for the function of the LCA subject and a reference to be used for South WWR (%) 10, 20, 30, 40, 50, 60
assessing the reference ows. FU in the present study is one square North WWR (%) 10, 20, 30, 40, 50, 60
Walls thermal hr.ft2 . F/Btu 11, 13, 15, 17, 19, 21
foot of a vertical building envelope (i.e., walls and windows) that resistance (R-value) m2 K/W 1.93, 2.28, 2.64, 2.99, 3.34, 3.69
encloses a hypothetical two-storey ofce building in Seattle, US, Insulation material batt, batt + berboard, batt + polystrene
with 3600.0 square feet (335 square meters) of oor area and a Window type double-glazed, triple-glazed
service life of 60 years, which represents a typical small size ofce Window frame material wood vinyl, aluminum, berglass/vinyl
building in Seattle. The envelopes thermal characteristic per FU
was considered to provide at least an R-value of 13.5 h ft2 . F/Btu
(2.36 m2 K/W), for the entire envelopes cross-section including the double-glazed window (glass U-factor = 1.65 W/m2 K; solar heat
insulation material. While the Seattles energy code for commercial gain coefcient (SHGC) of 0.42) and low-e clear triple-glazed win-
buildings prescribes a minimum R-value of 17.5, we used a lower dow (glass U-factor = 1.25 W/m2 K; SHGC of 0.31). Service life of
threshold for FU because of exploratory purpose of this research. materials are considered based on the common practice in indus-
As shown in Table 2, the wall in the building envelope consists try which is used by the LCA software, Athena Impact Estimator
of brick veneer, air gap, insulation, vapor barrier, concrete block, (IE), and is reported in Azari [3].
and gypsum plaster board. Window system consists of clear glass Several design characteristics of the building envelope of inter-
panes, frame materials, and cavity gas between the panes of glass. est are manipulated in this study in order to examine the changes
Window system consists of low-e clear glass panes, frame mate- of impacts in relation to those of a base design combination. The
rials, and 1.25-cm (0.5-inch) air gap between the panes of glass. varying design inputs include WWR in south facade, WWR in north
The window choices used for energy modeling include low-e clear facade, and external walls thermal resistance (R-value). Table 3
 R. Azari et al. / Energy and Buildings 126 (2016) 524534
shows design inputs along with their values, as examined in this
research.
The reference ow, another element in goal and scope deni-
tion, is about quantication of materials and energy needed per
design combination. Moreover, the system boundary shows what is
included in or excluded from the LCA scope. The system boundary in
the present study includes all stages of building envelope life-cycle
(from material extraction through manufacturing, to construction,
operation and maintenance, to demolition and recovery) as well
as associated transportation. The environmental impact categories
of interest included global warming, acidication, eutrophication,
smog formation, and fossil-fuel energy consumption.
2.1.2. Inventory modeling
In the next step of LCA methodology, the environmental ows
(inputs and outputs) associated with the building envelope life
cycle should be identied and quantied. Given the number of
envelope design inputs and their corresponding values assumed
in this research, the total number of possible design combinations
of building envelope was too many to be modeled considering the
limited resources available to us. Optimization techniques was a
reasonable methodological alternative in order to draw conclu-
sions based on limited number of modeling runs. We considered
only 91 random design combinations (called, the initial scenarios)
for inventory modeling and impact assessment and then used the
results in the next phase of research, to be explained in section
2.2, in a hybrid genetic algorithm (GA) and articial neural net-
works (ANN) approach to develop future populations/generations
of combinations and nd the optimum design combination.
We also made sure that diversity was present in the initial pop-
ulation of 91 design combinations used for inventory modeling
and impact assessment. Inventory modeling and impact assess-
ment were conducted using Athena Impact Estimator which uses
the US Environmental Protection Agencys (EPA) Tool for Reduction
and Assessment of Chemicals and other environmental Impacts
(TRACI) methodology for life cycle impact assessment [2]. While
Athena IE incorporates a large useful inventory database, it is not
designed to model the operational energy use associated with
design combinations. To feed it with this piece of data, we relied on
eQuest 3.65 [20], an energy performance modeling and simulation
tool which can estimate the energy consumed in the building for
heating, cooling, lighting, etc. We modeled the building in eQuest
3.65 and simulated its energy use for all design combinations. For
activity-, schedule- and system-related assumptions, we relied on
the softwares built-in default assumptions which represent com-
pliance with the energy code requirements. The energy simulation
results were then used in the LCA tool, Athena IE, along with the
other building design and construction information. Athena IE used
this information to generate the inventory results, i.e. the envi-
ronmental inputs and outputs/emissions, associated with design
combinations.
It is important to note that the LCAs computational approach
to inventory modeling introduced by Heijungs and Suh [19] con-
structs the inventory vector based on the processes in the system
boundaries and the environmental inputs and outputs to those pro-
cesses. The approach uses the equation below for this purpose:
[g] = [B] [A]1 [f ] (1)
where g is inventory vector, B intervention matrix, A1
inverse
matrix of technology matrix, and f is nal demand vector.
The Athena Impact Estimator (IE) uses LCA methodology and
a regionalized life-cycle inventory database in compliance with
ISO 14040/14044 [2]. The software allows for specication of
cities based on which the inventory data are regionalized, i.e.
regional differences with regard to transportation modes and dis-
tances, electricity grid, and manufacturing technology are taken
527
into account [2]. While the developers of the software do not pro-
vide access to its inventory database or details about the specic
electricity or transportation assumptions, it is expected that hydro-
electric power is considered to be the source of Seattles electricity
generation in the softwares inventory database, as suggested by
the US Energy Information Administration (2015).
2.1.3. Impact assessment
Impact assessment is the LCA activity that tries to classify the
environmental inputs and outputs for each design combination
based on their impacts on the environment. The type of impacts
we were interested in this research included global warming, acid-
ication, eutrophication, smog formation, and ozone depletion as
some of the most widely considered impact categories in the liter-
ature. Sometimes, an environmental emission could fall into more
than one impact category because of its multiple effects on the
environment.
Contribution of each environmental emission (methane, for
instance) to each impact category (e.g., global warming) is
quantied by multiplying its quantity into an impact-specic
science-based factor characterization factor that represents
relative importance of the emission of interest in that category
of impacts. In other words, characterization factors convert the
quantity of environmental emissions into measures that repre-
sent impact categories. The measure for global warming potential
(GWP), as an example, is equivalent kilogram of CO2 . This step of
impact assessment is called characterization and its outcome is
quantied contribution of the LCA subject, FU of building envelope,
to the impact categories of interest; i.e., global warming poten-
tial (GWP), acidication potential (AP), eutrophication potential
(EP), smog formation potential (SFP), and ozone depletion poten-
tial (ODP). Table 4 includes the information about design inputs and
environmental outputs per FU for 91 design combinations. The top
and bottom 5 percent values for environmental outputs in Table 4
are highlighted red and green, respectively.
2.1.4. Interpretation of results
The last step in LCA methodology, based on ISO 14040 [21], is
interpretation of results through which the results, limitations and
recommendations are presented. We present the results following
the second phase of research, optimization efforts.
2.2. Optimization
The next step in research was optimization. Several attempts
have been made to classify different optimization techniques. One
popular classication proposed by Cohon and Marks [8] classies
the optimization techniques to a priori (such as global criterion
method, etc.), a posteriori (such as -constraint, etc.) and pro-
gressive preference articulation techniques based on the way that
search and decision-making problems are handled by each method.
The evolutionary algorithm method (such as genetic algorithm,
etc.) that is used in our research is a different kind of approach
based on the theory that non-dominated solutions are chosen to
remain in a population as it evolves [6].
The LCA results for 91 design combinations were entered into
the optimization phase with the objective of identifying the opti-
mum values for design inputs. We pursued a two-step optimization
algorithm, as explained below.
2.2.1. Step 1: Modeling by Articial Neural Networks (ANNs)
ANNs are effective computational frameworks that imitate the
complex relationships of biological networks to solve intricate non-
linear problems, especially where classical mathematical modeling
processes fail to succeed [5,16,18]. In the present study, separate
ANNs were used to model and optimize operational energy and
528 R. Azari et al. / Energy and Buildings 126 (2016) 524534

Table 4
Impact assessment results. (For interpretation of the references to colour in the text, the reader is referred to the web version of this article.)
 R. Azari et al. / Energy and Buildings 126 (2016) 524534
environmental impacts. The process, which is the same for all out-
puts, is explained below using Eutrophication Potential (EP), as a
random example.
To model EP by an ANN, a four-stage methodology was followed.
In the rst stage, the normalization process was conducted on both
design combination data (input data) and EP data (targets). The
data normalization process minimizes the inuence of magnitude
and range of variations of the input variables throughout the ANN
training process. Generally, a linear transformation is applied to
normalize training and test data sets.
To nd the optimum parameters of an ANN including weights
and biases, the initial data were divided into two subsets, i.e. train-
ing and test data sets. Data division, the second stage of ANN, was
conducted through random selection. In this stage, 70% of the nor-
malized data were utilized to train the ANN and optimize weights
and biases. 30% of the data that was not considered in the training
process was used for validation and testing process.
The third stage of ANN methodology is denition of ANN archi-
tecture, including the number of layers and the number of neurons
in each layer. A ve-layer ANN was dened with 5-3-4-2-1 neurons
in hidden and output layers. The number of neurons in the input
layer is equal with the number of input variables, i.e. 6 neurons. The
architecture of proposed ANN is illustrated in Fig. 2.
Logistic sigmoid, hyperbolic tangent sigmoid, and linear transfer
functions are the most commonly used activation functions in the
optimization of ANNs. Here, the hyperbolic tangent sigmoid was
utilized as the activation function for the hidden and output layers
to scale outputs of each neuron within the range of 1 to +1.
The last stage in modeling the data by an ANN is to dene
and apply an appropriate training algorithm. In other words, ANN
parameters (including weights and biases) should be adjusted and
optimized to minimize the generated errors for both training and
test data sets. Indeed, the performance of an ANN is evaluated via
the calculation of the mean squared errors (MSEs) for both training
and test data sets. These values are unique signature of the net-
works performance and dened as the average of the sum of the
differences between the targets and the ANN predictions:
1 N   structured string, is made up of tightly connected genes and carries
MSE =  y yi,target (2)
N i=1 i
where N, yi , and yi,target are number of data for training or test pro-
cesses, the ANN output and target value for the ith experiment,
respectively.
In the present study, an evolutionary optimization algorithm
was used to train the dened ANNs and precisely adjust and opti-
mize their weights and biases. A computer code was developed
capable of handling a predened ANN and stochastically com-
puting and readjusting its weights and biases applying genetic
algorithm evolutionary method. For this purpose, the networks
unknown/adjustable parameters including 67 weights and 15
biases were codied into a chromosome composed of 82 genes.
Each gene within the dened chromosome can take a value
between 1 and +1. So, the developed computer code was trained
to scan and explore chromosomes by means of genetic algorithm
and nd the best chromosome results with minimum MSEs for both
training and test data sets. Hence, an initial population of these
chromosomes was randomly generated. For each chromosome, e.g.
chromosome number j, the genes were fed into the predened ANN
structure. Then, the training data set was used to evaluate the per-
formance of the fed ANN. The calculated MSE was considered as the
tness of chromosome number j. This training process is repeated
for all chromosomes separately. Then, the chromosomes are sorted
based on the calculated training MSEs. The best tness obviously
belongs to a chromosome that results in a minimum MSE. Apply-
ing selection, mating, crossover, and mutation operators, the new
529
Table 5
Genetic algorithm parameters used in ANN training.
Optimization Parameter Value
Initial population size 100
Selection operator Killing off the worst 50% of chromosomes
Mating operator Tournament selection
Crossover operator Single-point crossover
Mutation rate 20% of new generation
Training and test errorsa 1% and 2% respectively
a
In case of Operational Energy (OE) the training and test errors are set to be 3%
and 5%, respectively.
generation of chromosomes is produced and sent to the evaluation
unit. This optimization process is repeated till a preset training MSE
is established. Then, the trained ANN (the ANN lled with the genes
of the best chromosome) is evaluated via presenting test data set
and calculating test MSE. Final solution occurs when a test MSE of
less then or equal to a preset test MSE is achieved. Otherwise, the
algorithm returns to the third stage and is repeated.
The genetic algorithm parameters used in training process of
ANNs are presented in Table 5. Moreover, Fig. 3 illustrates the ow
chart applied for training and test processes of dened ANNs.
Based on the ve-stage computational algorithm and the ow
chart presented in Fig. 3, a computer code was developed in Pas-
cal programming language (Lazarus 1.2.4 IDE) and compiled into
64-bits executable using FPC 2.6.2. A Mersenne Twister-based sub-
routine was used to produce the required random numbers for the
modeling processes [27]. The random number generation subrou-
tine satises the tests of uniformity and serial correlation with high
resolution. The cycle length of the random number generator was
219937 -1.
2.2.2. Step 2: Multi-objective Optimization by Genetic Algorithm
(GA)
Genetic Algorithms (GAs) are effective stochastic optimization
tools used in different elds of science and technology [17,15]. Each
genetic algorithm process is initiated by randomly generating a
population of predened chromosomes. A chromosome, as a well-
the genetic information. Each chromosome as a genotype is related
to a phenotype as its corresponding physical object. Genetic algo-
rithm, as an evolutionary articial intelligence technique, applies
genetic operators, such as mutation and crossover, to the best
members from the previous generation in order to establish the
new generation. These operators stochastically adjust/control the
gene(s) values to help the population evolve and produce optimum
chromosomes. Mutation operator, as an exploration tool, is utilized
to randomly introduce new information from the search space (i.e.,
gene pool) and to move the solution(s) towards the global optimum.
On the other hand, crossover operator, as an exploitation tool, com-
bines information from existing chromosomes and conducts the
solution(s) towards local optima.
In the present study, genetic algorithm was applied to determine
the appropriate design inputs that lead to optimum operational
energy and environmental impacts. To formulate the genetic algo-
rithm, the design inputs are codied into a chromosome depicted
in Fig. 4.
An initial population of the chromosomes is randomly pro-
duced. The tness of each generated chromosome is calculated by
recalling the trained ANNs of the rst optimization step. Each chro-
mosome is identied by six tness values (objectives) including one
for operational energy and ve for environmental impacts. So, for
each chromosome all six trained ANNs were separately recalled
to calculate the corresponding tness. To do this, the genes of
each chromosome were fed into the input layer of the previously
trained ANNs and the outputs of the ANNs were reported as the
530 R. Azari et al. / Energy and Buildings 126 (2016) 524534

Fig. 2. Architecture of proposed ANN.

Fig. 3. The ow chart describing the genetic algorithm based training of ANNs.

Insulaon Frame Wall R- South North Glazing

Material Material Value WWR WWR Type

Fig. 4. Schematic representation of the chromosome consists of design inputs.

tness values for that chromosome. Then, the chromosomes were
sorted based on predened target(s). If the optimization target was
a single-objective optimization that minimized only one of the
environmental outputs, e.g. the Eutrophication Potential (EP), then
the chromosomes were totally ordered according to the EP tness
values calculated using the corresponding ANN for each chromo-
some. On the other hand, in case of a multi-objective optimization
that simultaneously minimizes two or more targets/objectives; e.g.,
operational energy and/or other environmental impacts altogether,
the Non-dominated Sorting Genetic Algorithm II (NSGA-II) is used
[7]. Unlike classical multi-objective optimization methods, the evo-
lutionary algorithms especially multi-objective versions of genetic
algorithms are much more effective and faster optimization tools.
The main difference between NSGA-II and conventional GA is the
sorting mechanism. Indeed, the chromosomes in NSGA-II are sorted
based on the concept of domination. In general, the chromosome
x dominates the chromosome y if it is not worse than the chromo-
some y for all objectives and strictly better than it for at least one
objective:
i , fi (x) fi (y)
(3)
j , fj (x) < fj (y)
 R. Azari et al. / Energy and Buildings 126 (2016) 524534
Fig. 5. Genetic Algorithm ow chart.
Table 6
Main controlling parameters of the applied genetic algorithm.
Optimization Parameter Value
Initial population size 50
Selection operator Merge, Sort, and Truncate
Mating operator Roulette wheel selection
Crossover operator Single-point crossover
Mutation rate 5% of new generation
where, fi (x) is the tness value of the chromosome x respect to
objective i. It is clear that i and j are representatives of environ-
mental outputs, i.e. OE, GWP, AP, ODP, EP, and SFP.
Having compared all possible pairs of population, the NSGA-II
inspects and evaluates the chromosomes based on the number of
non-dominations and assigns a number/rank to each chromosome
accordingly. So, the chromosomes are precisely classied and posi-
tioned within a set of fronts. The rst front consists of completely
non-dominated set of chromosomes in the current population and
the second front comprise of chromosomes are dominated by the
members in the rst front only and the front goes on. Hence, the
chromosomes in the rst front are given a rank value of 1 and those
in the second are assigned the rank value of 2 and so on. Then,
the chromosomes within each front are sorted according to their
crowding distance. The crowding distance is a measure of how close
a chromosome is to its neighbors. The chromosome located in the
lesser crowded region, i.e. the region with large average crowding
distance, results in better diversity in the population and is more
preferred.
In the next stage, the selection, mating, crossover, and muta-
tion operators were applied on the sorted population to select
and reproduce new generation. The adjustable genetic algorithm
parameters applied in this optimization step are listed in Table 6.
Also, the ow chart of this optimization process is shown in Fig. 5.
3. Results
The MSE variations of the best chromosome in training the data
for single-objective optimization of the data revealed a decline of
errors as the number of iterations increased. The Fig. 6, which illus-
trates the MSE variations in training the EP data, shows a decline
531
6
5
Training Error (%)
4
3
2
1
0 100 200 300 400 500 600 700 800 900 1000
Iteraon (-)
Fig. 6. Iteration-dependent error variations of the ANN model for optimization of
the responding variable; EP in this case.
from 5.67% to 1.00% within 946 iterations. Relatively large number
of epochs (iterations) and small errors shows that the developed
ANN model is able to properly predict the behavior of the data.
Table 7 lists the statistical details of the ANN model training and
testing for the entire data.
Number of epochs in Table 7 indicates the number of times that
the genetic algorithm is repeated to learn successfully how to fore-
see the response variable (EP, for instance) through the developed
network. Reasonably low training and test MSE values are indica-
tors of how well the models explain the data. Table 7 also reports
other statistical parameters such as correlation coefcient, coef-
cient of efciency, goodness of t, and coefcient of determination
for each environmental output. Coefcient of determination, for
instance, shows the proportion of variations in the response vari-
ables (EP, for instance) that can be explained by the model.
The performance of the developed ANN model was further
checked through comparison of predicted values of each response
variable with actual experimental data. Fig. 7 veries the accu-
532 R. Azari et al. / Energy and Buildings 126 (2016) 524534

Table 7
Statistical details of the ANN model training and testing in single-objective optimization.

Response SFP EP ODP AP GWP OE

Number of epochs 8720 946 171 5485 10281 17224

Training MSE 0.000399 0.000399 0.000399 0.000398 0.000399 0.009998
Test MSE 0.000672 0.000243 0.000249 0.001154 0.001279 0.009718
Training Error (%) 0.999794 0.999186 0.999707 0.998365 0.999628 4.999595
Test Error (%) 1.297002 0.780105 0.790548 1.698941 1.788496 4.929027
Max Training Error (%) 3.637830 (85) 2.722563 (2) 2.134754 (24) 3.848782 (14) 4.006225 (14) 14.881263 (18)
Max Test Error (%) 2.714232 (15) 1.758720 (59) 1.784440 (41) 4.561337 (16) 4.427776 (17) 15.275572 (74)
Correlation Coefcient 0.978599 0.981646 0.992016 0.969047 0.968478 0.931649
Coefcient of Efciency 0.957123 0.961080 0.983467 0.938673 0.937833 0.866035
Goodness of Fit (%) 79.293394 80.271999 87.141996 75.235793 75.066775 63.398793
Coefcient of Determination 0.957656 0.963630 0.984097 0.939053 0.937949 0.867970

Fig. 7. Relationship between the experimental data and the optimized ANN model outputs for OE, GWP, AP, SFP, ODP and EP.

racy of ANN in prediction of all environmental impacts as response
variables.
The single-objective optimization results shown in Table 8
reveal that the operational energy (OE) consumption is mini-
mized in a design scenario with about 60% south WWR, about
10% north WWR, and the use of triple-glazed windows, berglass
window frame, R-17 wall insulation, and FG batt and polystyrene
insulations. Compared with the optimum scenario for OE, the
GWP-optimized design scenario incorporates lesser R-value (R-
13), higher north WWR (about 60%; which in turn results in lesser
use of insulation in building skin), and using wood vinyl as win-
dow frame material. Among the optimized scenarios in Table 8,
only GWP- and SFP-optimized design scenarios entail relatively
large WWR percentages in both south and north. This in turn
means lesser use of insulation mass in building skin dened for
these scenarios. Hydrouorocarbon (HFC) blowing agents used in
extruded polystyrene production make signicant contribution to
global warming and smog formation and the low GWP and SFP
in design scenarios with lesser use of insulation can be associated
with the impact of HFCs. HFCs, however, do not contain chlorine
and have essentially zero ozone depletion potential. This means
that the mass of insulation does not have any effect on the ODP of
scenarios in Table 8.
On the other hand, the design scenarios with lowest WWR per-
centages result in minimum ODP and EP. In other words, the lowest
potentials for ozone depletion and eutrophication occur in cases
with the lowest mass of glass in building skin.
The result for multi-objective optimization is included in
Table 9. As the table shows, the optimum design scenario that
yields the optimum low environmental impacts is the one that
incorporates triple-glazed windows with berglass frame on south
elevation as about 60% WWR and on north elevation as about 10%
WWR. Interestingly, this optimum design scenario does not con-
tain the highest R-value, but R-17. The is in line with diminishing
return law based on which increasing R-value above a certain opti-
mum point does not have a major impact on improving the building
performance.
Another observation that needs to be veried with other data
is that the optimum design combination for all environmental
outputs is the same as design combination that yields the least
operational energy use. If veried by other data and future stud-
ies, this can imply that architects wishing to lower the life-cycle
 R. Azari et al. / Energy and Buildings 126 (2016) 524534 533

Table 8
Results of single-objective optimization. Design inputs and environmental outputs associated with optimum (minimum) scenarios.

Min OE case Min GWP case Min AP case Min ODP case Min EP case Min SFP case

Design Inputs
Insulation Material FG Batt+ Ext. FG Batt+ Ext. FG Batt+ Ext. FG Batt+ Ext. FG Batt+ Ext. FG Batt
Polystyrene Polystyrene Polystyrene Polystyrene Polystyrene
Window Frame Material Fiberglass Wood Vinyl Aluminum Fiberglass Fiberglass Wood Vinyl
Wall R-value hr.ft2 . F/Btu 17 13 11 21 21 11
m2 K/W 2.99 2.28 1.93 3.69 3.69 1.93
South WWR (%) 60 60 59 10 10 60
North WWR (%) 10 60 10 10 10 50
Glazing Type Triple Glazing Triple Glazing Triple Glazing Double Glazing Double Glazing Triple Glazing

Environmental Outputs
OE (MJ) 191.123 192.119 191.884 197.894 197.893 195.434
GWP (kg CO2-eq) 1804.6700 1695.6097 1819.9874 1912.5770 1912.5771 1746.4217
AP (kg SO2-eq) 6.6764 6.6290 6.0070 6.71600 6.7159 6.5406
ODP (kg CFC-11-eq) 9.0704E-07 9.1478E-07 8.9582E-07 8.2400E-07 8.2405E-07 9.5233E-07
EP (kg N-eq) 0.0898 0.0910 0.0869 0.0852 0.0852 0.0920
SFP (kg O3-eq) 18.2865 17.9231 18.3499 20.3440 20.3440 17.6045

Table 9 the architecture and construction research communitys need to

Results of multi-objective optimization showing the characteristics of optimum
design guidelines that reduce the operational energy consump-
scenario.
tion and environmental impacts simultaneously. For this purpose,
Design Inputs the methodology that we used in this research can be followed to
Insulation Material FG Batt + Ext.
develop design guidelines. More than its results, the signicance
Polystyrene
Window Frame Material Fiberglass
of our research is in the methodology that we applied. A future
Wall R-value hr.ft2 . F/Btu 17 direction of the research in this area could be to develop a soft-
m2 .K/W 2.99 ware tool that can help conduct the entire methodology within one
South WWR (%) 60 tool. Such tools need to include all design and construction inputs
North WWR (%) 10
beyond what we studied in our research. Also, it is important to
Glazing Type Triple Glazing
note that we treated all design inputs as well as environmental
Environmental Outputs outputs equally with regard to their signicance in design decision-
OE (MJ) 191.123
GWP (kg CO2-eq) 1804.6700
making. In real-life situations, however, architects may assign more
AP (kg SO2-eq) 6.6764 weight to some design inputs or environmental impacts because of
ODP (kg CFC-11-eq) 9.0704E-07 design priorities, or in order to overcome the conicting effects of
EP (kg N-eq) 0.0898 inputs or impacts, so future methodologies and tools should pro-
SFP (kg O3-eq) 18.2865
vide capability to assign weight to factors in their decision support
mechanism.
We also encourage future research to address some of the
environmental impacts of the building envelopes should focus limitations we faced in doing this study. This study focused on Seat-
on reducing operational energy of building more than any other tle as the location of project, mainly because the city is among
impact category. In real practice and in the context of global envi- the few cities that the version of Athena Impact Estimator that
ronmental challenges, however, some environmental impacts such was used in this research supported with regard to inventory
as global warming pose more immediate threats to the environ- modeling. Future research could use other LCA software, such
ment compared with other impacts. In this sense, a minimum-GWP as SimaPro, to cross-examine the results of the present research
design combination currently can be more important than an opti- in the context of other geographical locations. Indeed, different
mum design combination in achieving an eco-friendly building. locations enjoy different climates potentially, which would affect
We emphasize that the results of this research should be inter- energy performance of buildings. Various geographical locations
preted with caution because there are many other design-related also incorporate different manufacturing practices, electricity gen-
and location-related factors not captured in this study that can eration and distribution, transportation and logistic systems, and
affect the environmental impacts of building envelopes. consuming patterns, which all would affect the environmental life-
cycle impacts of buildings and their components.
Future research can also expand the scope of this study and
4. Conclusion examine more design characteristics or other building components
and systems. Finally, future research should also focus on devel-
While signicant research has been conducted on the impact opment of architect-friendly software and tools that can work as
of building envelopes on energy performance of buildings, the plugin to architectural modeling tools to optimize the environmen-
knowledge is still relatively vague about their life-cycle environ- tal impacts of design.
mental impacts. In the present research, we considered 6 basic
design characteristics of building envelopes and used quantita-
tive and simulation-based techniques to explore combinations of 6 References
design characteristics that yield optimum environmental impacts
in 6 categories of operational energy, global warming potential, [1] A. Asdrubali, C. Baldassarri, V. Fthenakis, Life cycle analysis in the
acidication potential, ozone depletion potential, eutrophication construction sector: guiding the optimization of conventional Italian
buildings, Energy Build. 64 (2013) 7389.
potential, and smog formation potential. [2] I.E. Athena, Athena Impact Estimator (IE) User Guide, Athena Sustainable
The results of this research once validated by further study Materials Institute, 2014 http://calculatelca.com/wp-content/uploads/2014/
in this area by other researchers have the potential to address 10/IE4B v5 User Guide September 2014. pdf.
534 R. Azari et al. / Energy and Buildings 126 (2016) 524534
[3] R. Azari, Integrated energy and environmental life-cycle assessment of ofce
building envelopes, Energy Build. 82 (2014) 156162.
[4] R. Azari, Y.W. Kim, Comparative assessment of life cycle impacts of curtain
wall mullions, Build. Environ. 48 (2012) 135145.
[5] P.J. Braspenning, F. Thuijsman, A.J.M.M. Weijters, Articial Neural Networks:
An Introduction to ANN Theory and Practice, Springer, Berlin, Germany, 1995.
[6] G. Chiandussi, M. Codegone, S. Ferrero, F.E. Varesio, Comparison of
multi-objective optimization methodologies for engineering applications,
Comput. Math. Appl. 63 (2012) 912942.
[7] C.A. Coello, D.A. Van Veldhuizen, G.B. Lamont, Evolutionary Algorithms for
Solving Multi-Objective Problems, Kluwer Academic Publishers, New York,
2002.
[8] Cohon, D.H. Marks, A review and evaluation of multiobjective programming
techniques, Water Resour. Res. 11 (1975) 208220.
[9] K. Deb, Optimization for Engineering DesignAlgorithms and Examples, PHI
Publishing, 2012.
[10] A. Dodoo, L. Gustavsson, R. Sathre, Lifecycle carbon implications of
conventional and low-energy multi-storey timber building systems, Energy
Build. 82 (2014) 194210.
[11] EIA, 2015. Monthly energy review. Energy Information Administration. http://
www.eia.gov/totalenergy/data/monthly/pdf/sec2 3.pdf.
[12] EN 15978, 2011. Sustainability of Construction Works. Assessment of
Environmental Performance of BuildingsCalculation Method. European
Committee for Standardization.
[14] EPA, 2009. Buildings and the Environment: a Statistical Summary. http://
ww2.harford.edu/faculty/eaugusti/Environ%20Health/
green%20building%20stats.pdf.
[15] M. Gen, R. Cheng, Genetic Algorithms and Engineering Design, John Wiley and
Sons Inc., 1997.
[16] K. Gurney, An Introduction to Neural Networks, Taylor & Francis, London, UK,
2005.
[17] R.L. Haupt, S.E. Haupt, Practical Genetic Algorithms, 2nd edition, John Wiley &
Sons Inc., 2004.
[18] S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice Hall PTR,
Upper Saddle River, NJ, USA, 1998.
[19] R. Heijungs, S. Suh, Computational Structure of Life Cycle Assessment, Kluwer
Academic Publishers, The Netherlands, 2002.
[20] Hirsch, J. (2009). The Quick Energy Simulation Tool (eQuest) [software tool].
Available from: http://www.doe2.com/equest/.
[21] International Organization of Standards, ISO 14040, Environmental
ManagementLife Cycle AssessmentPrinciple and Framework, National
Standard Authority of Ireland, Ireland, 2006.
[22] International Organization of Standards, ISO 14044, Environmental
ManagementLife Cycle AssessmentRequirements and Guidelines, National
Standard Authority of Ireland, Ireland, 2006.
[23] International Organization of Standards, ISO 16745, Environmental
Performance of Buildings-Carbon Metric of a Building, Use Stage, National
Standard Authority of Ireland, Ireland, 2015.
[24] S. Junnila, A. Horvath, Life-cycle environmental effects of an ofce Building, J.
Infrastruct. Syst. 9 (4) (2003) 157166.
[25] S. Junnila, A. Horvath, A.A. Guggemos, Life-cycle assessment of ofce buildings
in Europe and the United States, J. Infrastruct. Syst. 12 (1) (2006) 1017.
[26] L. Kosareo, R. Ries, Comparative environmental life cycle assessment of green
roofs, Build. Environ. 42 (7) (2007) 26062613.
[27] M. Matsumoto, T. Nishimura, Mersenne twister: a623-dimensionally
equidistributed uniform pseudo-random number generator, ACM Trans.
Model. Comp. Simul. 8 (1) (1998) 330.
[28] M. Ottele, K. Perini, A.L.A. Fraaij, E.M. Haas, R. Raiteri, Comparative life cycle
analysis for green facades and living wall systems, Energy Build. 43 (12)
(2011) 34193429.
[29] R.M. Pulselli, E. Simoncini, N. Marchettini, Energy and emergy based
costbenet evaluation of building envelopes relative to geographical
location and climate, Build. Environ. 44 (5) (2009) 920928.
[31] F. Stazi, A. Mastrucci, P. Munafo, Life cycle assessment approach for the
optimization of sustainable building envelopes: an application on solar wall
systems, Build. Environ. 58 (2012) 278288.
[32] D.D. Tingley, A. Hathway, B. Davison, An environmental impact comparison of
external wall insulation types, Build. Environ. 85 (2015) 182189.
[33] G. Treloar, R. Fay, B. Ilozor, P. Love, Building materials selection: greenhouse
strategies for built facilities, Facilities 19 (3/4) (2001) 139149.
[34] G. Treloar, R. Fay, B. Ilozor, P. Love, An analysis of the embodied energy of
ofce buildings by height, Facilities 19 (5/6) (2001) 204214.
[35] Y.G. Yohanis, B. Norton, Life-cycle operational and embodied energy for a
generic single-storey ofce building in the U.K, Energy 27 (2002) 7792.