You are on page 1of 14

BUILD SIMUL

DOI 10.1007/s12273-016-0334-z

An approach for building design optimization using design of


experiments

Research Article
Jay Dhariwal1,2 (), Rangan Banerjee1

1. Department of Energy Science and Engineering, Indian Institute of Technology Bombay, Mumbai, Maharashtra 400076 India
2. Department of Architecture, Massachusetts Institute of Technology, Cambridge, MA 02139, USA

Abstract Keywords
Building simulation based optimization involves direct coupling of the optimization algorithm to factorial design,
a simulation model, making it computationally intensive. To overcome this issue, an approach is sensitivity analysis,
proposed using a combination of experimental design techniques (fractional factorial design and response surface methodology,
response surface methodology). These techniques approximate the simulation model behavior surrogate modelling,
using surrogate models, which are several orders of magnitude faster than the simulation model. genetic algorithms,
Fractional factorial design is used to identify the significant design variables. Response surface optimization

methodology is used to create surrogate models for the annual cooling and lighting energy with
the screened significant variables. The error for these models is less than 10%, validating their
Article History
Received: 18 June 2016
effectiveness. These surrogate models speed up optimization with genetic algorithms, for single-
Revised: 10 October 2016
and multi-objective optimization problems and scenario analyses, resulting in a better solution. Thus,
Accepted: 19 October 2016
optimization becomes possible within reasonable computational time with the proposed methodology.
This framework is illustrated using the case study of a three-storey office building for New Delhi.
© Tsinghua University Press and
Springer-Verlag Berlin Heidelberg
2016

1 Introduction as they are computationally intensive. A typical optimization


analysis can take years to iterate over 20 design variables,
Space cooling using air-conditioning and lighting are the each with three parametric levels, assuming seven seconds
main sources of energy consumption in buildings in hot per simulation, on an Intel® Core™ i7-2670QM processor.
climates. This energy consumption can be reduced in a Sensitivity analysis (Tian 2013; Nguyen and Reiter 2015)

Building Thermal, Lighting,


cost-effective manner by choosing the right combination of and optimization methods (Nguyen et al. 2014; Machairas

and Acoustics Modeling


passive building design options using building simulation et al. 2014), coupled with building simulation programs,
based optimization. Building simulation programs, e.g. have been used to reduce the computational run time for
EnergyPlus (2014), RADIANCE (2016), etc. can be used to building design optimization.
accurately estimate the cooling energy and daylighting levels
in buildings. These programs can evaluate the building 1.1 Sensitivity analysis
energy and daylighting performance in a few seconds to a
few hours. Determination of the optimal building designs for Sensitivity analysis (SA) can be used to find the design
a single-objective or the Pareto front for desired objective variables that have the largest influence on the response
functions, e.g. lifecycle cost (LCC) and energy use intensity variable. Cooling energy, heating energy and lighting
(EUI) traditionally utilize parametric analyses i.e. exhaustive energy are some of the common response variables in the
search using building simulation based optimization. These context of building simulation. Local SA focuses on finding
analyses require a large number of simulations to find the the effect of the change in design variables around a point
optimal building designs and hence, they may not be viable or the base case. Gong et al. (2012) used orthogonal arrays
E-mail: jayd@mit.edu
2 Dhariwal and Banerjee / Building Simulation

of fractional factorial design for local SA to improve upon GenOpt in handling multi-objective optimization problems.
the base case. Global sensitivity analysis (GSA) spans the Nguyen and Reiter (2014) used GenOpt with EnergyPlus
entire input space, considering the effect of the different and suggested that if the direct coupling between building
combinations of design variables on the response. GSA is simulation and optimization engine is time-consuming, then
preferred as it makes the analysis base case independent. surrogate based optimization should be explored. Nguyen
Iooss and Lemaitre (2015, pp. 101–122) compared the GSA et al. (2014) stated that surrogate or meta-models can offer
methods graphically in terms of the number of model a faster fitness function evaluation to address this important
evaluations and the complexity of the model. Their analysis limitation of GA or other evolutionary algorithms. In the
suggests that fractional factorial designs (FFDs) and the context of building simulation, this means that the direct
Morris method are among the techniques requiring fewer coupling of the building simulation model with the GA
model evaluations. The Morris method can only rank the to evaluate the fitness function may be computationally
input factors but cannot quantify their effect on the outputs intensive. The use of a computationally efficient but appro-
so it is not possible to know how much of the variance of ximate fitness evaluation, in place of an exact evaluation,
the output is explained (Tian 2013). FFDs can quantify can enhance the use of GA to solve complex real world
the contribution of the variance of each design variable to problems since it enables exploring a larger search space
explain the response. The most significant design variables in less time.
can be selected based on this quantification. FFDs can
easily be used to analyze 50 or more design variables in a 1.3 Surrogate models
reasonable amount of time.
A surrogate model (SM) can be used to approximate the
1.2 Optimization methods simulation model behavior. It is created using a limited
number of simulation runs over the search space of interest
Optimization methods can be more effective in finding the for predicting the simulation output. Magnier and Haghighat
optimal building design after eliminating the less important (2010) used an artificial neural network (ANN) for surrogate
variables using GSA and thereby reducing the search space. modelling and coupled the ANN model with GA to optimize
Search methods based on genetic algorithm (GA) are most for thermal comfort and energy consumption. The time taken
popular (Machairas et al. 2014). Nassif (2014) used GA for to train the ANN model was 3 weeks and the optimization
optimization to reduce the energy consumption in HVAC time was short, whereas with a direct coupling between
systems. Evins et al. (2012) used full factorial design based simulations and GA could have taken several years.
GSA with Pareto optimization using NSGA-II algorithm Eisenhower et al. (2012) used the Support Vector Machines
for building simulation based optimization. The Pareto (SVM) method to create many SMs of a 30-zone EnergyPlus
front can be helpful when analyzing the trade-offs between building model. The authors stated that the optimization
the objective functions to choose the optimal solution. using the SVM model lead to results similar to those
Most researchers have used up to two objective functions obtained with the EnergyPlus model. ANN (Magnier and
only for Pareto optimization due to the computational cost. Haghighat 2010) and SVM (Eisenhower et al. 2012) have
Jin and Overend (2012) used three objective functions been the most commonly used machine learning models
corresponding to cost, carbon and comfort for Pareto to develop surrogate or meta-models (Machairas et al. 2014)
optimization. However, in their case, the final solution for building design. MATLAB and Dakota (Adams et al.
would not have been significantly different if considering 2014) are the optimization toolkits offering surrogate
two objectives at a time. Gerber et al. (2015) used multi- optimization capabilities, but they require more advanced
agent systems (Gerber and Ibanez 2014) to extend the skills to be used (Nguyen et al. 2014). Tresidder et al. (2012)
capabilities of multi-disciplinary design optimization in used Kriging SM for single-objective and multi-objective
the fields of architectural design, building engineering optimization and found its performance to be better over
and construction. Their approach combined design inputs the evolutionary algorithms searching the main model
and optimization methods with the real time data for user directly. However, they also advised to limit the number
preferences to inform simulation behavior. An optimization of design variables to 10 for Kriging based SM. Response
program, GenOpt (Wetter 2009), has implemented several surface methodology (Montgomery 2007) can also be used
optimization algorithms and can be coupled with building to form an empirical model of the variables for the responses
simulation programs. Karaguzel et al. (2014) coupled of interest, e.g. cooling energy prediction. Response surface
EnergyPlus with GenOpt to minimize LCC for building models are the most popular SMs in general. Robertson
materials and operational energy consumption. Nguyen et et al. (2015) used response surface models to predict domestic
al. (2014) discussed the limitation of the current versions of energy use. The application of response surface methodology
Dhariwal and Banerjee / Building Simulation 3

(RSM) for surrogate modelling can help to avoid using for these models and Pareto optimization uses a variant of
artificial intelligence techniques, e.g. ANN or SVM. These NSGA-II algorithm. These experimental design techniques
techniques require expert knowledge and have been the also offer an understanding of the functional relationship
main reason for the limited use of SMs for building design of the design variables with the response, unlike GA based
(Machairas et al. 2014). Originally, RSM was used to model optimization applied over the simulation model (Reeves
experimental responses but later on, its use was extended and Wright 1996). The SMs developed in this paper are
for computer experiments as well (Alvarez 2000). Once the also tested to assess their accuracy. This methodology is
SMs are created and validated, they can be used as part of illustrated using a case study of a three-storey, three-zone
the optimization algorithms for design (Nguyen et al. 2014; office building, with a floor area of 189 m2 per floor, based
Wetter and Wright 2004). on the composite climate of New Delhi.

2 Objectives of this study 3 Methodology

To overcome the high computational run-time limitation Figure 1 shows the flowchart of the design methodology used
of optimization methods coupled with building simulation to optimize the building design. As shown in the figure,
directly, a systematic approach is proposed in an optimization FFD is used for GSA to find the most significant variables
framework based on experimental design techniques, explaining the response. For these identified variables,
namely FFD and RSM. Although both these techniques are RSM is used for surrogate modelling to approximate the
useful when applied in isolation, these techniques are more simulation model. This RSM model is tested for error. If
effective when used together as it is done in other domains. the error is greater than 10% then another RSM design, e.g.
FFD is among the GSA techniques requiring fewer model central composite design (Montgomery 2007), can be fit.
evaluations to identify the most significant variables. After Otherwise, GA or other evolutionary algorithms are used
characterizing the response using FFD, RSM can then be for single- or multi-objective optimization. What-if scenario
used to fit an SM for these significant variables only. This analyses are performed for the changes in cost or values of
allows the use of RSM to fit a model with fewer variables the design variables to arrive at the best solution and design.
thereby enhancing its accuracy. RSM is easier to understand If the number of response variables is greater than one,
than artificial intelligence techniques and can increase the then the GSA and the SM parts in the flowchart need to
use of SMs for building design. Thus, a simulation model be applied to each of the response variables and then all
can be approximated to an SM for the responses of interest the SMs should be used together for optimization. This
using FFD and RSM. The use of FFD and RSM techniques methodology is illustrated through the case study.
together for building design optimization of a single-storey
building was shown by us earlier (Dhariwal and Banerjee
2015). This work is being extended for a multi-zone,
multi-storey building. The use of GA based optimization
with SMs is compared to GA directly coupled to simulation
in this paper to demonstrate the efficacy of the proposed
methodology.
The use of GA with these approximate models has the
potential to be orders of magnitude faster than using GA
based optimization on the simulation model directly. The
speed of SMs can help in analyzing a greater search space
to find a better design solution. This becomes more useful
for multi-dimensional Pareto optimization as it would be
much more computationally intensive to use optimization
algorithm coupled to the simulation model directly.
Approximate model based GA also offers the possibility for
scenario analyses around the design solutions to fine tune
the building design with the quick evaluation of SMs. Hence,
this approach is compared with GA optimization and GA
after GSA optimization, both directly coupled with building
simulation, for the results obtained for the same com-
putational time. GA is used for single-objective optimization Fig. 1 Flowchart for building design optimization methodology
4 Dhariwal and Banerjee / Building Simulation

3.1 Theory of design of experiments by calculating the corresponding treatment sum of squares.
The factors or the interactions with the highest treatment
The theory of design of experiments used in the methodology sum of squares explain the maximum variance in the response
is presented in brief in this section. Factorial designs can be and therefore, are the most significant. This is illustrated
used to find the effect of many factors or variables on the for the GSA of one of the responses from the case study in
output or the response. Factorial designs investigate all the appendix.
possible combinations of the levels of the factors. In the
k
context of building simulation, as an example, the design y = α0 + åαi xi + ååαij xi x j + ε (1)
variable “overhang depth” is considered as a factor and the i =1 i¹ j

different levels of overhang depth of 0 m and 1 m may be


é LO LO LO ù
considered as the two levels of this factor. These designs ê ú
can be effective for factor screening by using a small ê LO LO HI ú
ê LO HI LO úú
number of runs, putting all the k factors at two levels each. ê
ê LO HI HI úú
Equation (1) shows the regression model interpretation of E = êê (2)
ê HI LO LO úú
the two-factor factorial design, relating the response, y,
ê HI LO HI ú
with the factors, where ε represents the noise or error in ê ú
ê HI HI LO ú
response y. In Eq. (1), αi is the coefficient corresponding to ê ú
the main effect for xi and αij is the coefficient for the second êë HI HI HI úû
order interaction between the factors, xi and xj. The matrix
of experiments, E, shown in Eq. (2) for a 23 factorial design
åå( y
i j
ij - y.. )2 = åni ( yi . - y.. )2 + åå( yij - yi . )2 (3)
i i j

includes the elements eij, which are the input factors xj for
RSM is a collection of mathematical and statistical
the experiment i. The simulation runs are performed for
techniques for empirical model building where the response
the “LO” and the “HI” values of the factors, corresponding
of interest is influenced by several design variables. For
to the lowest and the highest values of the range for each
example, the response of interest, cooling energy (y), may
factor, for the two-level factorial design. As the number of
be described as a function of independent variables, e.g.
factors increases, the number of runs using this factorial
window to wall ratio (x1), roof reflectivity (x2), overhang
design can also rapidly outgrow the computational resources
depth (x3) in Eq. (4) where ε represents the noise or error in
of most experimenters. If the experimenter can assume
response y. The expected response, E(y)=f (x1, x2, x3), is
the higher order interactions to be negligible, then the
called a response surface. Usually, the form of the relationship
information on the main effects and the low order interactions
between the response and the independent variables is
can be obtained with only a fraction of the factorial design
unknown so a polynomial model is fit as per Eq. (5) using
runs. The use of FFD for GSA for building design is a
the method of least squares. For the polynomial model to
reasonable assumption as the higher order interactions
be a better approximation of the true functional relationship,
can be neglected. A single replicate factorial design strategy
the search space is first reduced using FFD for screening
is used for building design and the negligible effects are
and then RSM is used to create an SM using the screened
combined for an estimate of error (Montgomery 2007, p.
variables. Box-Behnken design (Montgomery 2007) and
246). After setting up the experimental design, the response
central composite designs (Montgomery 2007) are the
is calculated for each experiment and the analysis of variance
standard RSM designs used. Using one of these RSM designs
(ANOVA) is performed. Equation (3) shows the partitioning
ensures that a large number of runs or levels of the design
of the sum of squares for a one-factor ANOVA model,
variables are not needed. Further details on FFD and RSM
where yij is the response for the i-th level of the factor and
can be found in (Montgomery 2007).
the j-th replicate, ȳ.. is the grand mean, ȳi. is the mean for
the i-th factor level and ni is the number of replicates at y = f ( x1 , x2 , x3 ) + ε (4)
each factor level. The term on the left hand side of Eq. (3) is k k
called the “total sum of squares”, the first term on the right y = β0 + åβi xi + åβii xi2 + ååβij xi x j + ε (5)
hand side is called the “treatment sum of squares” due to i =1 i =1 i¹ j

the mean deviation of the i-th level of the factor from the
grand mean and the second term on the right hand side 4 Case study
is the “residual sum of squares”. As shown for a one-factor
ANOVA model, the model can be expanded to consider 4.1 Experiment setup
multiple factors and interactions. Through this analysis, the
contribution of all the factors and the interactions is found The simulation model for the base case is built in EnergyPlus.
Dhariwal and Banerjee / Building Simulation 5

The design variables are estimated for their ranges, thermal Table 1 shows the 26 design variables and their ranges
properties and cost data. GSA, SM and GA are applied used for the simulation runs. The window to wall ratio
as per the framework in Fig. 1. All the simulation runs for (WWR) lower limit was fixed as per the requirements
computing the cooling and the lighting energy are carried for LEED IEQ credit 8.1 (IGBC 2011) using daylighting
out on an annual basis. Design-Expert® software (Stat-Ease simulation in EnergyPlus. Glasswool with density of 48 kg/m3
2014) is used for the FFD, Box-Behnken RSM design and is used as the thermal insulation. Water is used as the
ANOVA. VBA programming language and MATLAB are additional thermal mass design variable, above the base case
used for running the simulations in batch mode. All the thermal mass. The range for WWR for the north facade
computational analyses are performed on a computer using (WWR-N) is kept higher than WWR for the other facades
an Intel® Core™ i7-2670QM processor. as the direct solar gain through this facade would not be
The methodology is illustrated for a three-storey office there at New Delhi’s latitude. This offers the possibility of
building in Fig. 2 for New Delhi. Each floor has two blocks having better views, while keeping the energy consumption
of 9 m by 9 m by 3m, connected by a common circulation low. Visible Light Transmittance (VLT) would help to
space. This area on each floor is considered as a thermal understand its impact on the lighting energy consumption.
zone and is centrally air-conditioned. This leads to three The response variables are annual cooling and lighting energy
zones on three floors with air-conditioning. The blocks on consumption for each zone. A high WWR would reduce
the north and the south side of the circulation space on the lighting energy due to the use of daylighting controls
each floor and the roof are the staircase and lift shafts. but increase the cooling energy and vice versa. The heating
There are neighbouring buildings on the south side and the energy is neglected as it is very minimal for this case.
east side, which cause solar shading. The building geometry
and the internal gains data are taken from the IBPSA student Table 1 Design variables and their ranges
modelling competition (IBPSA 2015). The equipment EUI Design variable Range
is 58 kWh/(m2·yr). Air-conditioning set points are assumed
15% to 75%, 15% to 50%,
to be 22 °C – 27 °C (Manu et al. 2014). The structural WWR
15% to 50%, 15% to 40%*
material for the base case is taken as concrete for it to be Window SHGC 0.2 to 0.8 for each facade*
representative of the Indian buildings. The data for the Window VLT 0.2 to 0.8 for each facade*
construction material for the base case of the envelope is Overhang depth 0 to 1 m for each facade*
taken from (Ramesh et al. 2012). Daylighting controls are Fin depth 0 to 1 m for each facade*
used in each zone to reduce the lighting energy consumption. Window U-value 1 to 5.8 W/(m2·K)
Roof reflectivity (ρ) 0.25 to 0.85
Wall insulation thickness (t) 0–0.1 m
Roof insulation thickness (t) 0–0.2 m
Wall additional thermal mass thickness (t) 0–0.1 m
Roof additional thermal mass thickness (t) 0–0.2 m
*North (N), east (E), west (W), south (S) facades respectively.

4.2 GSA

For the 26 design variables with their ranges as shown in


Table 1, the FFD chosen for GSA is a resolution V design.
This resolution V design can estimate the main effects and
the two factor interactions. A total of 352 simulation runs
based on this design were automated using VBA, calling the
EnergyPlus batch file for each simulation taking 9.5 hours
on an Intel® Core™ i7-2670QM processor.
Figure 3 shows the design variables that have the most
significant effects for the lighting energy GSA. Window
VLT, WWR for each facade and the interactions involving
them explain more than 80% of the total variance in the
model for the lighting energy response variable. The results
Fig. 2 Three-storey office building in New Delhi are similar for each zone. Figure 4 shows that WWR and
6 Dhariwal and Banerjee / Building Simulation

window SHGC for each facade, window U-value and wall 4.3 Surrogate modelling
insulation are significant in explaining the cooling energy
for each of the three zones. Additionally, roof ρ and roof Box-Behnken RSM designs are used in Design-Expert® for
insulation are also important for zone 3. These variables the creation of SMs for lighting energy and cooling energy
and the interactions involving them explain more than 80% for the three zones. These designs use three levels for each
of the total variance in cooling energy. Figure 4 also shows factor, taking the lower, the middle and the upper limit from
that WWR-N and SHGC-N are among the most important the range of the design variables in Table 1. The designs
design variables for the three zones for the GSA for cooling lead to 113 simulation runs for the lighting energy response,
energy. Roof ρ is the most important design variable for the 161 simulation runs for the cooling energy response for
zone 3 cooling energy GSA. zone 1 and zone 2 and 193 simulation runs for the cooling
Overhang depth, fin depth, wall and roof additional energy for zone 3. The same simulation runs could be used
thermal mass are the design variables excluded for the for finding the response for the zones 1 and 2 as the
creation of the SMs. These variables are set at the most significant design variables for these zones are identical.
appropriate values for optimization in the simulation The total time taken for the simulation runs is about 12 hours
model. Overhang depth is set at 1 m for all the facades on an Intel® Core™ i7-2670QM processor. These designs are
except the north facade as it has a significant effect and it is used in Design-Expert® to fit a second order polynomial to
an inexpensive energy efficiency option. The other three get the SM after the response variables are obtained from
design variables are removed from the model. Hence, eight the simulation runs. The SMs can be represented as per
design variables, namely window VLT and WWR for each Eqs. (6), (7), (8). The appendix shows an example of the
facade are chosen for the three zones for creating the SMs data used for the creation of an SM using the Box-Behnken
for the lighting energy. Window SHGC, WWR, window design for the zone 3 cooling energy.
U-value and wall insulation are chosen for all three zones Annual lighting energy for each zone =
to create the SMs for cooling energy. In addition to these
f (VLT-N, VLT-E, VLT-W, VLT-S, WWR-N, WWR-E,
variables, roof ρ and roof insulation are also chosen for the
zone 3 cooling energy SM. WWR-W, WWR-S) (6)

Fig. 3 GSA results for lighting energy consumption

Fig. 4 GSA results for cooling energy consumption


Dhariwal and Banerjee / Building Simulation 7

Annual cooling energy for zone 1 and zone 2= all the zones. The prediction error is on the positive side for
f (SHGC-N, SHGC-E, SHGC-W, SHGC-S, WWR-N, the lighting energy prediction and on the negative side for
the cooling energy prediction but is within 10%. These
WWR-E, WWR-W, WWR-S, Window U -value,
results validate the effectiveness of the SMs.
Wall insulation) (7)

Annual cooling energy for zone 3 = f (SHGC-N, 4.4 Optimization analyses


SHGC-E, SHGC-W, SHGC-S, WWR-N,WWR-E,
The optimization analyses are carried out for single-objective
WWR-W, WWR-S, Window U -value, Wall insulation, and Pareto optimization. The single-objective optimization
Roof reflective coating, Roof insulation) (8) minimizes the incremental LCC of the building. The
incremental LCC is the LCC over the initial fixed component
4.3.1 Validation of SMs of the construction cost of the building. Multi-objective or
Pareto optimization considers incremental LCC and EUI
Latin hypercube sampling and Monte Carlo sampling are
as the two objectives. Three cases are evaluated for the
each used to generate 125 test cases for the lighting and
optimization analyses, namely GA coupled with EnergyPlus
cooling energy SMs for the three zones since the number of
simulation (called “GA only” case), GA coupled with
design variables is large. The appendix shows an example
EnergyPlus simulation after GSA (called “GSA+GA” case)
of the data used to validate the SMs for the zone 3 cooling
and GA using SM formed with RSM after GSA (called
energy response variable. Figure 5 shows the predicted and
“GSA+SM+GA” case). The “GSA+SM+GA” case uses GA
simulated results for the lighting and cooling energy for the
with the SMs obtained for the cooling and lighting energy
three zones for the 250 test cases. Figure 6 shows the Box
estimation for each zone. The comparison of these cases
and Whisker Plot for the percentage error for the lighting
helps to understand whether GSA and SM add value for
and the cooling energy prediction. The error for the lighting
optimization. The number of simulation runs is limited to
energy for all the zones is kept together as it is similar for
the simulation runs needed for the “GSA+SM+GA” case to
compare these three cases. The total number of simulation
runs required for GSA, SM and the validation is about
1000 as discussed in Sections 4.2 and 4.3. The population
size for GA was estimated by running GA to minimize LCC
for the “GSA+SM+GA” case. In this case, the number of
model evaluations (the product of the number of generations
and the population size) was kept to be 1000. The population
size of 30 minimized the fitness function value and was
robust in the vicinity, thus it was assumed for all the cases.
Hence, the number of generations used for “GA only” case
1000
would be   ~ 33 . “GSA+GA” case requires 352 simulation
30
runs for GSA as per Section 4.2 so the number of generations
1000 - 352
for this case would be ~ 22 .
30
Fig. 5 Predicted and simulated results for each SM Table 2 shows the cost functions assumed for these
analyses. The analyses for LCC are in constant dollars. The
discount rate of 10% is used and the life of the house is
assumed to be 50 years. The cooling capacity for air-
conditioning required for each zone was estimated by
fitting regression models with cooling capacity (in W) as a
function of cooling energy (in kWh/(m2·yr)) as per Eqs. (9),
(10), (11). SHGC, VLT and window U-value are related as
shown in Eq. (12) using the data available for glazing products
in the Indian market. The cost data for plywood, additional
Fig. 6 Box and Whisker plot of % error for lighting energy (LE) thermal mass and insulation in Table 2 was obtained from
and cooling energy (CE) prediction local shops in India. For views, a constraint is kept for the
8 Dhariwal and Banerjee / Building Simulation

Table 2 Cost functions of parameters for optimization (1 US $ =


66 INR in 2016)
Life
Material Cost (in INR) (in years)
Windows (incremental 1300 (5.8 – Window U-value) 25
cost) per m2
Grid electricity 10 per kWh + 200 per month 1
(MERC 2014) (demand charge)
Roof ρ (GCCA 2015; 57.3 (Roof ρ)2 + 19.7 (Roof ρ) − 10
SSEF 2014) 4.1 per sq. ft.
Insulation (roof or wall) 3200 per m of thickness per m2 50
area
Air-conditioning 40K per TR of capacity 10
Plywood 1000 per m2 25 Fig. 7 Comparison of the minimum LCC solutions for the “GSA+
Additional thermal mass 5000 per m3 50 SM+GA” case

average WWR for external facades to be greater than 40% “GSA+SM+GA” case. Each such iteration for the “GA only”
as per Eq. (13). The overhang depth and the fin depth for and “GSA+GA” cases would have taken about 27 hours, so
the north facade is fixed at 0 m for aesthetics. it would have been computationally intensive. Figure 8
compares the “GA only”, “GSA+GA” and “GSA+SM+GA”
Zone 1 cooling capacity (in W) = 634 ´ CEzone1 + 749 (9)
cases for single-objective optimization. The figure shows
Zone 2 cooling capacity (in W) = 603 ´ CEzone2 - 3264 that the “GSA+GA” case uses 352 simulation runs to screen
significant variables before GA is run. The GSA helps this
(10)
case with a better initial solution, leading to the final minimum
Zone 3 cooling capacity (in W) = 495 ´ CEzone3 + 84 (11) LCC solution for this case to be better than that for the “GA
SHGC = -0.16 + 0.65 ´ VLT + 0.13 ´ (Window U -value) only” case. “GSA+SM+GA” takes about 1000 simulations for
GSA and SM creation and validation but after that it is
- 0.01´ (Window U -value)2
orders of magnitude faster and gives a better solution than
(12) the “GSA+GA” and “GA only” case solutions.
WWR-N + WWR-E + WWR-W + WWR-S
³ 40% (13) 4.4.2 Multi-objective optimization
4
Multi-objective optimization is explored to find the Pareto
4.4.1 Single-objective optimization
front with respect to LCC and EUI. The “GA only”,
The “GA only” and the “GSA+GA” cases take about 27 hours “GSA+GA” and “GSA+SM+GA” cases are compared for
for 1000 simulations. GA implemented in MATLAB is used the spread of the solutions on the Pareto front and the
for single-objective optimization. Since the “GSA+SM+GA” convergence to the best Pareto front. A variant of NSGA-II
case uses the SMs developed in Section 4.3, it is able to algorithm (Deb 2001) implemented in MATLAB is used
search through several hundred generations in less than a
minute to converge on an optimal solution. Due to the speed
of the “GSA+SM+GA” case, it is possible to find the effect
of varying different options in GA, e.g. initial population
seeds or crossover fraction, quickly and to choose the best
solution among them. Figure 7 compares three solutions
for “GSA+SM+GA” case. The first solution uses the same
random seed as the “GA only” and “GSA+GA” cases. For the
second solution, the upper bound of the ranges of design
variables is used as the initial population and for the third
solution, a different random seed is chosen as compared to
the second solution. As shown in the figure, the third solution
converges after ~700 generations and it is better than the
other two solutions. The simulated solution for this case is Fig. 8 Comparison of the minimum LCC solutions for the three
about 0.7% different than the predicted solution for the cases
Dhariwal and Banerjee / Building Simulation 9

for Pareto optimization. The number of simulation runs


used for all the cases is 1000 as discussed in Section 4.4.1.
Hence, for the “GA only” case, the number of generations
for 1000 simulation model evaluations would be 33 and for
the “GSA+GA” case, the number of generations would be
22. For the “GSA+SM+GA” case, once the SMs are created,
the run time would be in the order of minutes as compared
to the “GA only” and “GSA+GA” cases, in which the
simulation runs would have to be carried out again for
approximately 27 hours.
Since the “GSA+SM+GA” case is fast due to the fitness
function evaluation with the SMs, it is possible to explore
different options, e.g. initial population seeds or crossover Fig. 10 Comparison of the Pareto front solutions for the three
fraction, to find a better solution. Figure 9 compares the cases
four solutions explored for this case. The first solution uses
the minimum LCC solution as the initial population, the 4.4.3 What-if scenario analyses for the minimum LCC
second solution uses the upper bound of the design variables solution
as the seed, the third solution uses the lower bound of the What-if scenario analyses are performed around the
design variables as the seed and the fourth solution seed is minimum LCC optimum design solution of the “GSA+
the same as the third solution except for a different Pareto SM+GA” case from Section 4.4.1 to find the sensitivity of
fraction. The first and the second solutions are among the the design to the changes in the costs of the design variables
best solutions while the third solution has a much poorer and other constraints. Figure 11 compares the design solutions
distribution than the other three solutions. This suggests obtained for different what-if scenarios. Each of these
that it is possible for the GA to converge to a local minimum scenarios is optimized with GA using the SMs developed in
solution, which cannot be known before running the Section 4.3. Similar to the previous cases, the time taken by
optimization. Faster SMs provide the ability to explore multiple SM based GA is less than 5 minutes on an Intel® Core™
solutions thereby helping in eliminating worse solutions. i7-2670QM processor.
Figure 10 compares the three cases for the Pareto front. Higher WWR for the north facade would offer better
The “GA only” solution has a poorer spread and convergence views as there are no neighbouring buildings in this
as compared to the “GSA+GA” and “GSA+SM+GA” cases. direction. The scenario with the window costs halved leads
The “GSA+GA” case spans the entire range except in the to the change in the design solution from single-glazed
area of EUI of 87 to 91 kWh/(m2·yr) in terms of the spread. to triple-glazed windows. Different scenarios show that the
The “GSA+GA+SM” case solution has the best spread and solution is largely insensitive to the changes in the costs
convergence among the three cases. The simulated solution and other constraints. These analyses would have been
for the “GSA+SM+GA” case is also better than those of the computationally intensive using the “GA only” or “GSA+GA”
other cases.

Fig. 9 Comparison of the Pareto front solutions for the “GSA+ Fig. 11 What-if scenario analyses for the “GSA+SM+GA” case
SM+GA” case LCC solution
10 Dhariwal and Banerjee / Building Simulation

cases as the optimization for each case would have taken or scenario analyses without incurring any extra simulation
several more hours. These are all further advantages of using runs. Running several hundred generations of GA using
SM based GA analysis over GA coupled with the simulation the SMs takes less than 5 minutes. On the other hand, for
models directly. every analysis, the “GA only” and “GSA+GA” cases would
need to be coupled with the main simulation model
4.5 Discussion making them computationally intensive.
 The analyses performed for single-objective optimization

The base case simulation model is created in EnergyPlus of minimizing LCC show that the “GSA+GA” case yields
for both the cooling and lighting energy simulations. a better solution than the “GA only” case showing that
Within the proposed framework, it is also possible to use a GSA adds value. The “GSA+SM+GA” case yields the best
different simulation engine for each of the cooling and solution among the three cases, showing the value of SMs
lighting energy simulations, e.g. EnergyPlus for cooling with GSA. As the SMs offer a quick evaluation, it is
energy and RADIANCE for lighting energy. GSA is used to possible to explore and analyze multiple solutions for this
identify the significant variables as the number of variables case. The SM based GA also helps to avoid converging to
used to create the SMs is recommended to be low in the a local minimum solution.
literature for RSM. FFD is used for GSA as it was found to  The analyses for multi-objective Pareto optimization

be a method requiring fewer number of model evaluations. using LCC and EUI as the objective functions show that
But this could be replaced with other GSA techniques as the convergence and spread over the Pareto front of the
well. Similarly, RSM was used for surrogate modelling as it “GSA+SM+GA” case is better than in the other two cases.
was found to be an easier technique to understand, which The “GA only” case has a poor spread on the Pareto front
could aid in the increased use of SMs for building design for the case study. For Pareto optimization for two or
optimization. The prediction error for the SMs was less higher dimensions, “GSA+SM+GA” seems to be a better
than 10%. If the error is greater than 10% then another RSM option than the other cases, in which the direct coupling
design, such as central composite design (Montgomery to simulation model is needed.
2007), can be fit. Another strategy to reduce the error  What-if scenario analyses around the optimal solution

can be to create an SM with fewer design variables. Other using the SM based case are quick and help in choosing a
commonly used models in this domain, e.g. ANN, SVM or better design, considering the impact of changes in costs
Kriging, can also be applied to create an SM. and other constraints. This would have been very time
The application of SMs is shown for their use for consuming with the “GA only” and “GSA+GA” cases.
optimization in this paper. In order to show the advantages  SM based techniques can be made faster using

of the SMs, one of the most used optimization algorithms, parallelization of multiple processors as the respective
GA, is compared for three cases, namely “GA only” where simulation runs for GSA or surrogate modelling do not
the simulation model is directly coupled to GA, “GSA+GA” have to be done sequentially. This is in contrast to GA
where the simulation model is coupled to GA after GSA coupled with simulation, which has to be run generation
and “GSA+SM+GA” where the GA is run over the SM, by generation. The SM creation process using parallelization
which makes it much faster. “GSA+SM+GA” case requires can also be automated and could be useful for cloud
about 1000 simulation runs to create the SM from the computing based solutions.
simulation model so the number of simulations runs for Apart from the use of SMs with evolutionary algorithms,
the other two cases is also restricted to 1000. While keeping the SMs would also be valuable to study the interrelationships
the time similar for the three cases, the solution quality is between the different design variables for the cooling or the
compared. Sections 4.4.1 and 4.4.2 compare the three cases, lighting energy. For example, a high WWR would reduce
namely, “GA only”, “GSA+GA” and “GSA+SM+GA” for the lighting energy due to the use of daylighting controls
single-objective optimization and Pareto front. Section 4.4.3 but increase the cooling energy and vice versa. For educational
performs what-if scenario analyses around the minimum purposes, the SMs could be created for the reference buildings
LCC optimum design solution of the “GSA+SM+GA” case and put up on the web for public use for optimization
from Section 4.4.1. The following observations are made: without having to run the computationally intensive
 SMs address an important limitation of GA i.e. to make simulations. This can help in providing a real-time feedback
the fitness function evaluation less expensive in terms of to the building simulation community on the impact of
time, thereby improving GA’s effectiveness for building their choices. The use of SMs could also be explored for
simulation based optimization. other computationally intensive simulation studies, e.g. to
 SMs have to be created only once and then they could be approximate the wind pressure coefficients with the CFD
used for single-objective or multi-objective optimization simulations using the prominent wind directions for
Dhariwal and Banerjee / Building Simulation 11

calculating the annual natural ventilation potential. The “GSA+SM+GA” case is compared to the “GA
only” and “GSA+GA” cases for single objective and Pareto
4.6 Limitations of this work optimization to ascertain the effectiveness of GSA and SM.
“GSA+SM+GA” case yields better solutions for both the
Since the FFDs use two levels of each of the design variables LCC minimization and the convergence and the spread
for GSA, the FFDs assume a monotonic relationship of on the Pareto front. It also performs the what-if scenario
the design variables with the response. However, this may analyses quickly, helping to select a better design for the
be a reasonable assumption for the initial screening of the changes in costs and other constraints. The SMs enhance
design variables for building simulation based optimization. the effectiveness of GA with their speed and are able to
SMs reduce the computational time for building design explore a larger search space to find better solutions.
optimization as compared to the use of simulation models The SMs developed through this analysis can be used
at the expense of incurring a prediction error of up to 10%. for mass housing projects for a given climate by architects
For a sensitive objective function, where a small deviation or other building industry professionals, who do not have
from the optimum values can result in the significant a background in building simulation. The speed of SMs
reduction in the optimum, the SMs may be less useful. The would make this methodology valuable for calibration using
use of multiple SMs for optimization may also increase the experimentation, with the uncertainty in input parameters
risk of accumulating errors in the optimization process. To to the simulation models used to form statistical distributions
overcome these issues, emphasis should be put into the of outputs instead of point estimates. It would also be
automation of the process of validating the SMs. The SMs interesting to apply this methodology to urban scale energy
selected should minimize the prediction error. modelling of neighbourhoods to make urban design more
There have only been limited studies of the use of SMs interactive with the speed up. This approach opens up
for building design and more studies are needed to see if possibilities for building design optimization analysis.
a significant difference between the SMs and the actual
simulation models exists. Acknowledgements

5 Conclusions The authors are grateful for the constructive feedback from
the anonymous reviewers; it helped to considerably improve
A building simulation based optimization methodology this article. The authors also acknowledge Alonso Dominguez
using design of experiments is proposed for building Espinosa and Susan Spilecki from MIT for proofreading
design and is illustrated for the case study of a three-storey, this paper.
three-zone air-conditioned office building in New Delhi
climate. Annual cooling energy and lighting energy for each References
zone are the responses for GSA and surrogate modelling
for the case study. FFD screens the most important design Adams BM, Mohamed SE, William JB, Keith RD, John PE, Kenneth
variables for which the SMs are created. The SMs are TH, Lara EB, et al. (2014). Dakota, a multilevel parallel object-
orders of magnitude faster than the simulation model oriented framework for design optimization, parameter estimation,
by approximating the simulation model behavior. The uncertainty quantification, and sensitivity analysis. Version 6.0
optimization with GA using these SMs leads to quick analyses theory manual. Albuquerque, NW, USA: Sandia National
of single- and multi-objective optimization and what-if Laboratories.
scenarios, resulting in a better solution than GA directly Alvarez LF (2000). Design optimization based on genetic programming.
coupled to the simulation model. Thus, the optimization PhD Thesis, University of Bradford, UK.
becomes possible within reasonable computational time Deb K (2001). Multi-Objective Optimization Using Evolutionary
with the proposed methodology. Algorithms. Chichester, UK: John Wiley & Sons.
FFD screens VLT and WWR for each facade as the Dhariwal J, Banerjee R (2015). Building simulation based optimization
significant variables for the lighting energy GSA. Window through design of experiments. Paper presented at 2nd IBPSA-
SHGC and WWR for each facade, window U-value, wall Italy Conference (BSA2015), Bozen-Bolzano, Italy.
insulation, roof ρ and roof insulation are found to be the Eisenhower B, O’Neill Z, Narayanan S, Fonoberov VA, Mezić I (2012).
most significant out of the 26 design variables for cooling A methodology for meta-model based optimization in building
energy GSA. The SMs created for the lighting and cooling energy models. Energy and Buildings, 47: 292–301.
energy are validated with Latin hypercube sampling and EnergyPlus (2014). EnergyPlus Engineering Reference, the Reference
Monte Carlo sampling and are found to have a prediction to EnergyPlus Calculations. Available at http://www.energyplus.gov.
error of less than 10% for the entire search space. Accessed 15 Nov 2014.
12 Dhariwal and Banerjee / Building Simulation

Evins R, Pointer P, Vaidyanathan R, Burgess S (2012). A case study Nguyen AT, Reiter S (2014). Passive designs and strategies for low-
exploring regulated energy use in domestic buildings using design cost housing using simulation-based optimization and different
of experiments and multi-objective optimisation. Building and thermal comfort criteria. Journal of Building Performance Simulation,
Environment, 54: 126–136. 7: 68–81.
GCCA (2015). A Practical Guide to Cool Roofs and Cool Nguyen AT, Reiter S, Rigo P (2014). A review on simulation-based
Pavements. Global Cool Cities Alliance, Available at http:// optimization methods applied to building performance analysis.
www.coolrooftoolkit.org/. Accessed 15 Nov 2015. Applied Energy, 113: 1043–1058.
Gerber DJ, Ibanez M (2014). Paradigms in Computing: Making, Nguyen AT, Reiter S (2015). A performance comparison of sensitivity
Machines, and Models for Design Agency in Architecture. Los analysis methods for building energy models. Building Simulation,
Angeles: eVolo. 8: 651–664.
Gerber DJ, Pantazis E, Marcolino L, Heydarian A (2015). A multi agent RADIANCE (2016). The RADIANCE 4.2 Synthetic Imaging System.
systems for design simulation framework: Experiments with Available at http://radsite.lbl.gov/radiance/. Accessed 10 Oct 2016.
virtual physical social feedback for architecture. In: Proceedings Ramesh T, Prakash R, Shukla KK (2012). Life cycle energy analysis of
of Symposium on Simulation for Architecture & Urban Design a residential building with different envelopes and climates in
(SimAUD), Part of Spring Simulation Multiconference (SpringSim), Indian context. Applied Energy, 89: 193–202.
Reeves CR, Wright C (1996). Genetic algorithms and the design of
Alexandria, VA, USA.
experiments. In: Davis LD, De Jong K, Vose MD, Whitley LD
Gong X, Akashi Y, Sumiyoshi D (2012). Optimization of passive design
(eds), Evolutionary Algorithms. New York: Springer. pp. 207–226.
measures for residential buildings in different Chinese areas.
Robertson JJ, Polly BJ, Collis JM (2015). Reduced-order modeling and
Building and Environment, 58: 46–57.
simulated annealing optimization for efficient residential building
IBPSA (2015). BS2015 student competition. Available at http://
utility bill calibration. Applied Energy, 148: 169–177.
www.ibpsa.org/?page_id=464. Accessed 15 Nov 2015.
SSEF (2014). Cool Roofs for Cool Delhi. Shakti Sustainable Energy
IGBC (2011). Green Building Rating System for New Construction
Foundation, Available at http://shaktifoundation.in/initiative/
and Major Renovations, LEED 2011 for India. Hyderabad, India:
cool-roofs-for-cool-delhi/. Accessed 15 Nov 2014.
Indian Green Building Council.
Stat-Ease (2014). Design-Expert® software, version 9. Available at http://
Iooss B, Lemaître P (2015). A review on global sensitivity analysis
www.statease.com.
methods. In: Dellino G, Meloni C (eds), Uncertainty Management
Tian W (2013). A review of sensitivity analysis methods in building
in Simulation-Optimization of Complex Systems: Algorithms and energy analysis. Renewable and Sustainable Energy Reviews, 20:
Applications. New York: Springer. pp. 101–122. 411–419.
Jin Q, Overend M (2012). Facade renovation for a public building Tresidder E, Zhang Y, Forrester AIJ (2012). Acceleration of building
based on a wholelife value approach. In: Proceedings of Building design optimisation through the use of Kriging surrogate models.
Simulation and Optimization Conference, Loughborough, UK. Paper presented at First Building Simulation and Optimization
Karaguzel OT, Zhang R, Lam KP (2014). Coupling of whole-building Conference (BSO12), Loughborough, UK.
energy simulation and multi-dimensional numerical optimization Twiga (2014). Twiga Insul. Available at http://www.twigafiber.com/
for minimizing the life cycle costs of office buildings. Building twigainsul.php. Accessed 15 Nov 2014.
Simulation, 7: 111–121. Wetter M (2009). GenOpt, Generic Optimization Program—User
Machairas V, Tsangrassoulis A, Axarli K (2014). Algorithms for Manual, Version 3.0.0. Technical Report LBNL-2077E. Berkeley,
optimization of building design: A review. Renewable and USA: Lawrence Berkeley National Laboratory.
Sustainable Energy Reviews, 31: 101–112. Wetter M, Wright J (2004). A comparison of deterministic and
Magnier L, Haghighat F (2010). Multiobjective optimization of building probabilistic optimization algorithms for nonsmooth simulation-
design using TRNSYS simulations, genetic algorithm, and Artificial based optimization. Building and Environment, 39: 989–999.
Neural Network. Building and Environment, 45: 739–746.
Manu S, Shukla Y, Rawal R, de Dear R, Thomas L (2014). India Model Appendix
for Adaptive (thermal) Comfort-IMAC 2014. Ahmedabad:
CARBSE. Available at http://www.carbse.org/wp-content/uploads/
2014/09/Introduction_IMAC2014.pdf. A1 GSA for cooling energy
MERC (2014). Electricity Tariff for Reliance Infrastructure— 
Distribution Consumers. Maharashtra Electricity Regulatory Table A1 shows the values of the design variable for three
Commission, Available at http://www.rinfra.com/pdf/Tariff_for_ out of 512 runs for GSA for the case study. The values of
uploading_010414.pdf. Accessed 15 Nov 2014. these variables correspond to their lower and upper limits
Montgomery DC (2007). Design and Analysis of Experiments. New in Table 1. For example, design variable window SHGC-N
Delhi: John Wiley & Sons. for the north facade windows has a low value of 0.2 and
Nassif N (2014). Modeling and optimization of HVAC systems using
a high value of 0.8. Based on the combination of the 26
artificial neural network and genetic algorithm. Building Simulation,
7: 237–245.
variables, for the first simulation run in Fig. A1, the response,
Nayak JK, Prajapati JA (2006). Handbook on Energy Conscious cooling energy for zone 3, is 33.88 kWh/(m2·yr). ANOVA
Buildings. IIT Bombay and Solar Energy Centre, Ministry of table in Table A2 shows the significant effects and the
Non-Conventional Energy Sources, Government of India: R & D contribution of these effects in explaining the variance in
project no. 3/4(03)/99-SEC. response, i.e. cooling energy for zone 3.
Dhariwal and Banerjee / Building Simulation 13

Table A1 Design variable values for 3/512 runs for GSA


Design variable Run 1 Run 2 Run 3
Overhang-N Low Low Low
Overhang-E High High Low
Overhang-W Low Low High
Overhang-S High Low High
WWR-N Low High Low
WWR-E High Low Low
WWR-W Low High Low
WWR-S High Low High
Window SHGC-N High Low Low
Window SHGC-E Low High Low
Fig. A1 Cooling energy for the zones for 3/512 runs for GSA
Window SHGC-W Low Low Low
Window SHGC-S High Low Low
A2 SM for the zone 3 cooling energy
Roof ρ High Low Low
Wall thermal mass Low High High Equation (8) shows the 12 design variables chosen to create
Roof thermal mass High High Low an SM for the zone 3 cooling energy. The Box-Behnken
Window U-value High Low High design is created for these 12 design variables to fit an SM.
Wall insulation High Low High This design involves 12 blocks, in each of which four design
Roof insulation High High Low variables are varied through the four possible combinations
Fin-N Low Low Low of high and low, while the other eight design variables are
Fin-E High High High kept at middle values. This leads to 12 × 24=192 runs. The
Fin-W High Low High
193rd run is created by keeping all 12 design variables at the
middle values. The low and the high values for each design
Fin-S Low High High
variable are set from the lower and upper limits of Table 1.
VLT-N High Low High
The middle value is the average of the low and the high values.
VLT-E High High Low
Table A3 shows the 24=16 runs for one of the blocks out
VLT-W Low High Low
of the 193 runs. SHGC-N, SHGC-E, WWR-N and WWR-W
VLT-S High Low High
are varied in combinations of their low and high values,

Table A2 ANOVA table for GSA for zone 3 cooling energy for the case study
Variance ratio, F-value, Contribution (in %),
Variables Sum of squares, SS Degrees of freedom, f Mean SS, SS/f (Mean SS/Mean error SS) SS/Total SS
Model 99661 32 3114 575
Roof ρ 14554 1 14554 2687 14%
Window SHGC-N 12852 1 12852 2373 13%
WWR-N 9975 1 9975 1841 10%
Window SHGC-S 7954 1 7954 1468 8%
Roof ρ × Roof insulation 7389 1 7389 1364 7%
WWR-N × Window SHGC-N 4628 1 4628 854 5%
Window SHGC-E 3968 1 3968 733 4%
Roof insulation 3901 1 3901 720 4%
Window SHGC-W 3799 1 3799 701 4%
WWR-S 3751 1 3751 692 4%
WWR-E 3556 1 3556 657 4%
WWR-W 3531 1 3531 652 3%
Residual 1728 319 5
Total 101389 351
14 Dhariwal and Banerjee / Building Simulation

Table A3 16/193 runs of Box-Behnken design for SM for zone 3 cooling energy*
Run number SHGC-N SHGC-E WWR-N WWR-W Zone 3 cooling energy (in kWh/(m2·yr))
1 0.8 0.2 15% 15% 30.54
2 0.2 0.8 15% 15% 30.47
3 0.2 0.2 75% 15% 29.30
4 0.8 0.8 75% 15% 44.70
5 0.2 0.2 15% 50% 30.03
6 0.8 0.8 15% 50% 36.19
7 0.8 0.2 75% 50% 44.29
8 0.2 0.8 75% 50% 34.88
9 0.2 0.2 15% 15% 27.28
10 0.8 0.8 15% 15% 33.65
11 0.8 0.2 75% 15% 42.02
12 0.2 0.8 75% 15% 32.29
13 0.8 0.2 15% 50% 33.17
14 0.2 0.8 15% 50% 33.12
15 0.2 0.2 75% 50% 31.95
16 0.8 0.8 75% 50% 46.93
*Other 8 of the 12 design variables are set at middle values for these 16 runs.

according to their ranges from Table 1. For example, for the using the formula for each design variable in Table A4.
simulation run number 1, SHGC-N is at its high value of This formula uses the lower and upper limits of the design
0.8, WWR-N is at its low value of 15%, etc. Rest of the eight variables from Table 1. For example, the window SHGC
design variables, i.e. SHGC-W, SHGC-S, WWR-E, WWR-S, varies between 0.2 and 0.8 for each facade so the test cases
wall insulation, roof insulation, roof ρ and window U-value would also lie between these two extremes. For these test
are set at middle values. For example, the roof ρ would be set cases, the zone 3 cooling energy is compared for its values
0.25 + 0.85 from the simulation runs and the SM prediction to calculate
at = 0.55 from its range in Table 1. Simulations the error.
2
are carried out to calculate the response variable, zone 3
Table A4 Test cases for zone 3 cooling energy
cooling energy for these 193 runs. Table A3 shows the
Design variable Test case data
response for simulation run 1 to be 30.54 kWh/(m2·yr).
WWR-N (in %) 15 + (75−15)U(0, 1)*
Using the Box-Behnken design from Table A3, the method
WWR-E (in %) 15 + (50−15)U(0, 1)*
of least squares is fit to estimate the parameters of the
second order polynomial, which is the SM. WWR-W (in %) 15 + (50−15)U(0, 1)*
WWR-S (in %) 15 + (40−15)U(0, 1)*
Window SHGC 0.2+(0.8−0.2)U(0, 1)* for N, E, W, S facades
A3 Generation of test cases for the validation of the SM
Window U-value 1.0+(5.8−1.0)U(0, 1)*
for zone 3 cooling energy
Wall insulation (in m) (0.1) U(0, 1)*
Latin hypercube sampling and Monte Carlo sampling Roof insulation (in m) (0.2) U(0, 1)*
methods are each used to generate 125 test cases for the 12 Roof ρ 0.25 + (0.85−0.25) U(0, 1)*
design variables between 0 and 1 for zone 3 cooling energy * U(0,1) generates a random number between 0 and 1 as per Latin hypercube
validation. These values between 0 and 1 are converted sampling or Monte Carlo sampling.
to values falling within the ranges of the design variables

You might also like