Content uploaded by Patrick Janssen

Author content

All content in this area was uploaded by Patrick Janssen on Jan 12, 2016

Content may be subject to copyright.

81

Evolutionary

Optimisation of Semi-

transparent Building

Integrated Photovoltaic

Facades

Thian-Siong Choo and Patrick Janssen

issue 1, volume 12international journal of architectural computing

82

Evolutionary Optimisation of Semi-transparent

Building Integrated Photovoltaic Facades

Thian-Siong Choo and Patrick Janssen

The optimisation of semi-transparent building

integrated photovoltaic facades can be challenging

when attempting to find an overall balance

performance between conflicting performance criteria.

This paper presents a three-phase design optimisation

method that maximises overall electricity savings

generated by these types of facades by simulating the

combined impact of electricity generation, cooling load,

and daylight autonomy.Two demonstrations are

performed, with the difference being that the second

demonstration uses an enhanced model for calculating

daylight savings that takes into account the use of

blinds to counteract glare. For both demonstrations,

the three-phase optimisation method significantly

reduces optimisation run times. Comparing the design

variants evolved by the two demonstrations, the use of

the enhanced daylight savings model results in a total

electricity savings that is more accurate but in terms of

visual differentiation, the difference between the

optimized design variants is relatively small.

1. INTRODUCTION

With the current global emphasis on sustainable design, there is a trend to

design multifunctional semi-transparent building integrated photovoltaic

(BIPV) facades. Such facades integrate PV cells into conventional facade

glazing systems. Semi-transparent BIPV facades can provide good daylight

availability, reduce the solar heat gain through the building envelope, and

also have the ability to generate electricity to supplement the building’s

electricity consumption. It has been shown that semi-transparent BIPV

facades are effective in improving energy efficiency and reducing the overall

electricity consumption of a building [1][2].

The challenge of designing semi-transparent BIPV facades is to optimize

the multiple conflicting performance criteria. Unlike typical roof mounted

photovoltaic systems, where performance is only focused on the amount of

electricity generated, the design of semi-transparent BIPV facades has an

impact on a wider range of factors, including solar radiation and daylight

penetration into the rooms in the building.To optimize the performance of

such facades, optimisation systems can be used that leverage existing

simulation tools for performance evaluation. However, the types of

simulations that are required are often complex in their own right, and may

take a significant amount of time to execute. Optimisation systems typically

need to execute these simulations many times in an iterative manner and as

a result, optimisation systems can have very long run times, with a single run

possibly taking weeks to complete [3].

In order to test the run time of an evolutionary algorithm, a base-case

optimisation of a semi-transparent BIPV facade was conducted with the

performance criteria being the maximisation of electricity savings taking into

consideration electricity generation, cooling load, and daylight savings.This

base-case optimisation does not use parallel processing but uses the

standard software tools and computer systems available to most practising

architects.A parametric model was created and an evolutionary algorithm

was used to evolve an optimized population of designs.The evolutionary

algorithm ran for almost 14 days.This long run-time is a major barrier for

practitioners to adopt multi-objective optimisation methods into their

workflow [4]. Such run times clearly do not align well with a designer’s

process of working.

This paper proposes a three-phase optimisation method as an

alternative method for the optimisation of semi-transparent BIPV facades.

Section 2 gives an overview of the proposed three-phased method for semi-

transparent BIPV façades. In addition, it describes the design scenario, based

on a typical office façade in Singapore, used in the two demonstrations.

Section 3 presents the first demonstration, which maximizes total energy

savings taking into consideration electricity generation, cooling load, and

daylight savings.Section 4 presents the second demonstration, which also

maximises total electricity savings, but in this case an enhanced model for

83Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

calculating daylight savings is used that takes into account the use of blinds

to counteract glare [5]. Section 5 discusses and compares the results from

both demonstration 1 and 2. Lastly, Section 6 briefly draws conclusions and

highlights avenues for further research.

2.THREE-PHASE OPTIMISATION METHOD

The proposed method is based on a general design method developed by

Janssen and Kaushik [6].This method is adapted to the design of semi-

transparent BIPV facades using an evolutionary optimisation approach.The

method consists of three phases: 1) calibration, 2) optimisation, and 3)

validation as shown in Figure 1. In the calibration phase, simulation models

are selected and simulation programs are configured and tested. Simulation

models that are deemed too slow are replaced by faster proxy models,

which are calibrated in order to ensure that appropriate trade-offs are

achieved between speed and accuracy. In the optimisation phase, the proxy

models are used within the iterative optimisation process in order to

explore design variants with improved performance. Finally, in the validation

stage, selected optimal design variants on the Pareto front are analysed and

evaluated in more detail.Any calibrated proxy models are now replaced by

the more accurate simulation models in order to verify the performance

improvements.

Figure 1: Schematic

of the three-phase

method

84 Thian-Siong Choo and Patrick Janssen

2.1 Performance Objectives

In the tropics, semi-transparent BIPV facades affect electricity savings in

three distinct ways, namely electricity generation, cooling load and daylight

savings. Electricity generation is the electricity produced by the PV cells in

the semi-transparent BIPV facade. Maximising electricity generation will

reduce external electricity consumption from the grid. Cooling load is the

amount of electricity required to cool a room to a set temperature.

Cooling load is affected by the amount of solar radiation entering through

the facade. Minimising solar radiation will reduce the electricity

consumption for cooling. Daylight savings is the amount of electricity saved

by using daylight instead of artificial lighting in order to light a room to a set

minimum illuminance level.The daylight savings is affected by the daylight

autonomy, which is the percentage of occupied hours per year when the

minimum illuminance level can be maintained in a room by daylight alone

[7]. Maximising daylight autonomy will also maximize daylight savings.

In the case of daylight savings,an additional factor that may be taken into

consideration is that occupants will typically close window blinds in order

to counteract glare. Since closing blinds reduces daylight autonomy, daylight

savings will also be reduced.

2.2 Calibration Phase

For each of these components, an appropriate simulation model will be

chosen and tested. If the execution time of the simulation is excessively

slow,a proxy model will then be used that is faster to execute, but that

nevertheless still has sufficient accuracy to guide the evolutionary process.

The total electricity saved is defined by Equation 1:

ES= EG+ DS– CL(1)

where ESis the total electricity saving (kWh·a-1), EGis the electricity

generated (kWh·a-1), DSis the daylight savings (kWh·a-1) and CLis the

cooling load (kWh·a-1).

2.3 Optimisation Phase

For the optimisation phase, an evolutionary algorithm is used to evolve a

population of optimized design variants.The evolutionary algorithm requires

three key procedures to be defined: development, evaluation, and feedback.

The development procedure generates design variants using a parametric

model. Genes in the genotype are then associated with the parameters in

the model.

The evaluation procedure evaluates design variants.For both

demonstrations, the evaluation procedure calculates the total electricity

savings, which includes the daylight savings, cooling load and electricity

generation.The second demonstration uses an enhanced model for

calculating daylight savings that takes into account the use of blinds to

counteract glare.

The feedback procedure kills design variants that perform badly and

reproduces design variants that perform well.The evaluation scores for the

85Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

different performance criteria are combined to create a single fitness score

as shown in Equation 1.

2.4 Validation Phase

For the validation phase, the best design variants that emerge from the

evolutionary process are analysed and re-simulated.At this final stage, the

slower but more accurate original simulations should replace the proxy

simulations. Performance characteristics of the final design variants can then

be verified.

2.5 Demonstrations

Both demonstrations involve optimising the pattern of PV cells on a semi-

transparent BIPV facade in order to maximize the total electricity savings as

shown in Figure 2.The PV pattern affects both the solar radiation and the

daylight penetrating into the room through the glazing.

A typical north oriented office space for one person occupancy with 4

m (width) x 4 m (depth) x 3 m (height) is modelled for the experiment, as

shown in Figure 2 (top left).The facade is separated into four zones: vision

glass panels 1, 2, 3 and a spandrel glass panel. Each zone is independent from

one another.Three discrete valued genes – cell height, cell width and cell

spacing – define the PV cell pattern for each independent zone, as shown in

Figure 2 (bottom left). Cell height and width vary from 5 – 15.5 cm at 0.5

cm steps. Cell spacing varies from 0.5 – 5 cm at 0.5 cm steps.All the cells of

the semi-transparent BIPV facades will be similar in shape.The pattern

occupies a facade with a height and width of 4 m.

Figure 2.Top Left:Simulation model

with sensors (a = 1.50m, b = 0.85m, c

= 0.50m). Bottom Left: Schematic of

cell arrangement for semi-transparent

BIPV façade with gene 1_x, 2_x and

3_x, where x is the glass panel

numbers (1-3). Right. Section of a

typical office space.

86 Thian-Siong Choo and Patrick Janssen

For the first demonstration, the model used for calculating daylight

savings assumes that no blinds are used. For the second demonstration, the

enhanced daylight savings model assumes that occupants will close the

blinds to counteract visual discomfort due to glare.A roller blind with an

openness factor of 3 is modelled behind vision glass panel 2 as shown in

Figure 2.The blind is activated whenever the daylight glare probability is

above 0.35, which is when glare is perceptible [8].

2.6 Software Tools

The three-phase method shown in Figure 1 is not dependent on any

particular software. It can be implemented with a range of different

software tools for parametric modelling and performance simulation. For

the two demonstrations, commonly used software tools for six different

tasks were considered.They are 1) parametric modelling; 2) performance

simulation of electricity generation, EG,of BIPV system; 4) performance

simulation for daylight savings, DS;3) performance simulation of cooling

load, CL;5) performance simulation of daylight glare probability and 6)

evolutionary algorithm for design optimisation.

For both demonstrations, the same sets of software tools are used. For

parametric modelling, Grasshopper was chosen mainly due to its popularity

in the design community [3]. Grasshopper is a plugin for the Rhinoceros

computer aided design modelling software [9]. It is a Visual Dataflow

Modelling (VDM) system that allows designers untrained in scripting to

generate parametric models quickly [10].

A number of specialist Grasshopper components and custom

components listed in Table 1 are used for running optimisation algorithms

and executing simulations. For evolutionary optimisation, the Galapagos

component is used [3].This component is an evolutionary optimisation

solver, which can be used to optimize designs for single fitness metrics. For

electricity generation, a simplified model is implemented in Grasshopper as

a mathematical equation. For cooling load, a model is defined that uses the

DIVA-Viper Grasshopper component that links to the EnergyPlus energy

simulation engine [11][12]. For the daylight savings model, a model is defined

Table 2: Selected

software for

demonstration 1 and 2.

87Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

Processes Selected Software

Parametric Modelling Grasshopper (GH)

Evolutionary Algorithm Galapagos (a GH component)

Simulation for EG Custom GH component

Simulation for DS DIVA-Daylight (a Grasshopper plugin that has a wrapper to DAYSIM)

Simulation for CL DIVA-Viper (a Grasshopper plugin that has a wrapper to EnergyPlus)

Simulation of DGP Custom GH component developed as a wrapper to Evalglare

that uses the DIVA-Daylight component that links to the Daysim simulation

engine [7][11]. In this case, Daysim is used to calculate daylight autonomy,

which is then used as a basis for calculating daylight savings. Both the

cooling load model and the daylight savings model use the EnergyPlus

weather data file for Singapore [13].

For the first demonstration, Figure 3 shows the workflows for the

various software tools and how they relate to the three-phase method. For

the second demonstration, an additional customized simulation component

is developed within Grasshopper to link to Evalglare, a radiance-based tool

for glare simulation and analysis [5].The addition of the Evalglare simulation

is shown in Figure 6 in Section 4.

Figure 3:Workflow of different

software tools for demonstration 1

88 Thian-Siong Choo and Patrick Janssen

3. DEMONSTRATION 1

The first demonstration calculates electricity savings by taking into

consideration electricity generation, cooling load and daylight savings. In this

case, the model used for calculating daylight savings assumes that no blinds

are used.

3.1 Calibration Phase

For the calibration phase, the three components of the total electricity

savings calculation will be explained below.

Electricity Generation Model

In order to obtain a precise prediction of the annual electricity generation

of any particular photovoltaic module, the module first needs to undergo

extensive testing in the laboratory in order to obtain a set of electrical

characteristics. Using this data, EnergyPlus can then be used to simulate

annual electricity generation using a model called the ‘equivalent one diode

model’ [14].

In terms of execution time, the equivalent one diode model in

EnergyPlus is sufficiently fast. However, in this case, the problem is that the

electrical characteristics of the photovoltaic modules being evolved are not

known.The developmental process generates alternative configurations of

photovoltaic modules and it is therefore clearly impossible to test these in a

laboratory.A simplified model is therefore used that can estimate the annual

electricity generation with reasonable accuracy using the mathematical

equation shown in Equation 2:

P = Asx Fax Gtx Effcell x Effinvert (2)

where P is the electrical energy produced by photovoltaic modules (kWh),

Asis the net area of photovoltaic module (m-2), Fais the fraction of surface

area with active solar cells, Gtis the total annual solar radiation energy

incident on BIPV (which is set at a DIVA-calculated value of 561 kWh.m-2),

Effcell is the semi-transparent BIPV facades module efficiency (which is set at

12%) and Effinvert is the average inverter efficiency (which is set at 90%).To

find Gt,the annual solar radiation energy incident on the BIPV, a one-time

simulation is done with DIVA-Daylight.

In order to verify the accuracy of the simplified model, a set of

commercially produced photovoltaic modules (for which the electrical

characteristics were already known) were simulated using both the

equivalent one diode model in EnergyPlus and the simplified model. In total,

simulations for 16 different modules for the four different cardinal

directions were carried out,resulting in 64 variants.The two sets of data

were plotted on a graph using Microsoft Excel and a linear trend line was

calculated.This resulted in an R2coefficient of determination of 0.98,

thereby confirming that the simplified model has a reasonably good

correlation with the equivalent one diode model in EnergyPlus.

Daylight Savings Model

Daylight savings is calculated based on the daylight autonomy for the room,

which can be simulated using DIVA-Daylight. Since the Daysim simulation in

DIVA-Daylight is already an optimized simulation model, the simulation

executed relatively quickly and there was no need to create a proxy in this

case [7][11].

With reference to a recent study on various lighting standards around

the world by Halonen et al, it was found that minimum illuminance for

interior spaces ranges from 200lx to 500lx [15]. Hence, for the simulation

of daylight autonomy, the minimum illuminance level of 500lx is set for the

simulation.Working hours are set from 9:00 to 18:00.A 3 x 3 nodal grid of

daylight sensors are drawn 0.85m from the floor and 0.25m away from the

vertical walls (Figure 2, top). Since daylight autonomy is more critical for

89Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

areas further way from the windows,only the back 2 rows of 6 daylight

sensor nodes are used for the daylight autonomy simulation.

Default materials from the material library in DIVA-Daylight are used for

the simulation. Based on research done by, which shows that a typical

photovoltaic module has a reflectance of below 10%, the photovoltaic layer

is assigned a reflectance of 10% [16].

The following settings were used in DIVA-Daylight: ab = 2, ad = 1000, as

= 20, ar = 300 and aa = 0.1, where ab is ambient bounce, ad is ambient

divisions, ar is ambient resolution and aa is ambient accuracy.The detailed

explanation of the settings is beyond this paper.They can be found in the

Radiance manual. Daylight savings were then calculated according to the

following formula:

DS= (DAsim/100) * LPB * FA * WH (5)

where DSis the total daylight savings (kWh·a-1), DAsim is the simulated

daylight autonomy (%),LPB is the lighting power budget (kW.m-2), FA is the

floor area of simulation model (which is 16m2) and WH is the working

hours per year. LPB is set based on the Code of Practice, which

recommends an LPB for offices of 0.015kW.m-2 [17]. WH is set based on 5

work days per week with 9hrs of work per day, which results in 2,340hrs

per year.

Cooling Load Model

The cooling load for the room is simulated using DIVA-Viper. For

simplification, the study considers the heat gain through the semi-

transparent BIPV facades but not the internal heat gains from lights,

equipment and occupants. Default materials from the material library in

DIVA are used.With reference to Figure 2, the walls, floor and ceiling are

assigned as “adiabatic” and spandrel glass panels are assigned as “opaque

spandrel glass”.A window module is defined using a typical 6 mm thick clear

glass window with a U-value of 5.8, a solar heat gain coefficient (SHGC) of

0.82 and a visible transmittance (VT) of 0.88 [18].

Since the PV cells are on the surface of the glass, they will reduce the

solar radiation entering the window,essentially acting as small shading

devices.There are two different approaches to modelling this shading effect

with different trade-offs between speed and accuracy.The slower but more

accurate approach is to model PV cells as individual shading elements.The

calculation of the amount of solar radiation entering the window is

dependent on the time and location of the sun, and it therefore needs to be

simulated for each time-step for an entire year. Depending on the design

variant, the total number of PV cells can range from 156 to 4224. Each cell

is assigned a solar reflectance of 0.1 and visible reflectance of 0.1.The large

number of shading elements resulted in a relatively long time taken for the

simulation to run, with the longest simulation taking approximately 30

90 Thian-Siong Choo and Patrick Janssen

minutes.This caused the optimisation of the base case, mentioned in the

introduction, to run for almost 14 days.

In view of the slow simulation speed, a faster proxy model is used.With

this approach, the solar heat gain coefficient (SHGC) and visible light

transmittance (VLT) for the facade are adjusted to take into account the

effect of the PV cells.The equations for SHGC and VT used in the proxy

simulation are shown in Equations 3 and 4:

SHGC = Apv/A x SHGCvg (3)

where SHGC is the solar heat gain coefficient of semi-transparent BIPV

facade (Wm-2K-1),Apv is the total area of PV cells (m2),A is the area of

semi-transparent BIPV facade (m2), and SHGCvg is the solar heat gain

coefficient of the vision glass panel (which is part of the semi-transparent

BIPV facade but without photovoltaic cells) (Wm-2K-1).

VT = Apv/A x VTvg (4)

where VT is the visible transmittance of the semi-transparent BIPV facade

(Wm-2K-1),Apv and A are as in Equation (3), and VTvg is the visible

transmittance of the vision glass panel (which is part of the semi-

transparent BIPV façade but without PV cells).

In order to check the accuracy of the proposed proxy model, 164

cooling load simulations for different BIPV facades were conducted using

both the accurate simulation and the proxy model.The two sets of data

were plotted on a graph using Microsoft Excel and a linear trend line was

calculated.This resulted in an R2coefficient of determination of 0.93,

thereby confirming that the simplified model has a reasonably good

correlation with the slower and more detailed cooling load simulation.

3.2 Optimisation Phase

For executing the evolutionary algorithm, the Grasshopper Galapagos

component is used. Galapagos is executed with a population of 30, initial

boost of 2%, and the maintain level is set at 10% and inbreeding at 75%.The

system was executed on a single computer with an i5 Intel core CPU of

3.5GHz with 8GB of RAM, on 64 bits Windows.

The optimization process ran for 2 days,17 hours, generating a total of

1471 design variants.The design variant with the best fitness is variant

number 753, with a total electricity savings of -131 kWh·a-1 (shown as a

black dot on Figure 5).

91Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

3.3 Validation Phase

In the final validation phase, a set of designs from the Pareto front were

selected and analysed (see Figure 5). In order to verify the performance

improvements in cooling load, these designs were re-simulated using the

slow cooling load simulation. In order to better understand the

relationships between the design variants, all designs are ranked using

Pareto ranking, and those on the Pareto front are then plotted on a 3D

graph (see Figure 5 left). On this graph, the optimal design is shown as a

black dot and certain selected other designs are shown as grey dots.

In all cases, the results from the slow mode simulation confirmed the

performance improvements. Compared to the base case design shown in

Figure 4, the optimal design variant shown in Figure 5 has an overall

improvement for the total electricity savings of 61%.

Figure 4. Left: Base case design

variant with standard PV cell

arrangement. Right: Bar chart of ES,E

G,

DSand CLfor base case design variant.

Figure 5. Left: 3D plot of results for

optimisation assuming no blinds are

used. Right:

design variant 753, the optimized

design variant for demonstration 1.

92 Thian-Siong Choo and Patrick Janssen

4. DEMONSTRATION 2

The second demonstration calculates total electricity savings using the same

approach as in the first demonstration, taking into consideration electricity

generation, cooling load, and daylight savings. In this case, an enhanced

daylight savings model is used.

4.1 Calibration Phase

For electricity generation and cooling load, exactly the same models are used

as in the first demonstration. For daylight savings, an enhanced model is

developed that assumes that occupants will close the blinds to counteract

visual discomfort due to glare.This will reduce the daylight savings that can be

achieved since closing blinds will reduce daylight autonomy.The modified

equation for calculating the total electricity generated is shown in Equation 6.

ES= EG+ DS*– CL(6)

where ESis the total electricity saving (kWh·a-1), EGis the electricity

generated (kWh·a-1), CLis the cooling load (kWh·a-1), DS*is the enhanced

daylight savings (kWh·a-1).The enhanced daylight savings model is described

in more detail below.

Enhanced Daylight Savings Model

For enhanced daylight savings, the model must take into account the use of

blinds to counteract glare. One approach to this would be to perform an

annual simulation that calculates glare at every point in time. However, a

base-case simulation took approximately 6 hours and as a result this

approach was clearly too slow.

A proxy model is therefore used that calculates a reduction factor to

daylight savings.The reduction factor is defined by the glare coefficient, α,

which is added as a factor to daylight savings, DS:

DS*= αDS(7)

where DS*is the enhanced daylight savings (kWh·a-1) taking into account

glare, DSis the daylight savings (kWh·a-1), and αis the glare coefficient.The

full equation for calculating enhanced daylight savings is shown in Equation

8, and with the exception of the glare coefficient, it is exactly the same as

Equation 5 used by the daylight savings model for the first demonstration.

DS*= (αDA/100) *LPB *FA *WH (8)

Glare Coefficient

The key to the enhanced daylight savings model is therefore calculating the

glare coefficient.As a starting point to calculating this coefficient, it is noted

that it may be considered to be a factor that reduces daylight autonomy due

to the use of blinds.Therefore, if daylight autonomy can be calculated both

with blinds and without blinds, then the appropriate factor for reducing

93Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

daylight autonomy can be calculated.This factor would then be the glare

coefficient.

Glare can be approximated using a measure called ‘daylight glare

probability’ (DGP), which represents the probability of people being visually

disturbed due to the following factors: level of vertical eye illuminance,

luminance of glare source, solid angle of glare source and position index.

Glare is considered imperceptible when DGP is less than 0.35. Hence for

the enhanced daylight autonomy calculation,it is assumed that the blinds

will be close at all times when the DGP rises above 0.35.

DGP can be simulated using the Evalglare program [5]. Since Evalglare is

not available as a Grasshopper component, a customized Grasshopper

component has been developed by Choo et al [19].This allows DGP to be

easily calculated at any point in time for any design variant. However, the

problem is once again that the annual simulation is very slow, and calculating

the annual DGP for every point in time for a whole year would be

prohibitively slow. In order to overcome this, DGP for a whole year can be

feasibly calculated for a range of design variants with just a single point in

time. In such a case, it is important to choose a point in time that is

representative of a range of design variants and different glare conditions.To

determine this point in time, 20 runs of the annual glare simulation were

performed for a range of design variants in order to determine the day and

time when the blinds for all design variants were subject to a DGP of

greater than 0.35.The selected point in time was the 21st June, 12:00PM

which is the solstice.

The methodology for calculating the glare coefficient can therefore be

summarised as follows. First, a set of 125 design variants were generated

from the parametric model. For each of these design variants, daylight

autonomy without blinds and daylight autonomy with blinds was calculated

for the 21st June, 12:00PM.The two sets of results for daylight autonomy

(without blinds and with blinds) were then plotted on a graph in Excel and a

linear regression was then performed.The glare coefficient is then equal to

the regression coefficient, which in this case was 0.69.

Figure 6 shows the workflows for the various software tools, including

Evalglare.This figure may be compared to Figure 3.

94 Thian-Siong Choo and Patrick Janssen

4.2 Optimisation Phase

The same settings are used as for the first demonstration. Galapagos is

executed with a population of 30 and initial boost of 2%.The maintain level

is set at 10% and inbreeding at 75%.The system was executed on a single

computer with an i5 Intel core CPU of 3.5GHz with 8GB of RAM, on 64

bits Windows.

Figure 6:Workflow of different

software tools for demonstration 2,

including the Evalglare software tool.

Figure 7: Left: 3D plot of results for

optimisation with enhanced daylight

savings. Right: Optimized design variant

for demonstration 2.

95Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

The optimization process ran for 14 days 17 hours,generating a total of

1771 design variants.The design variant with the best fitness is variant

number 861, with a total electricity savings of -674 kWh·a-1 (see Figure 7

right).As with the previous demonstration, a 3D Pareto graph is plotted in

order to better understand the relationships between the design variants

(see Figure 7 left).

4.3 Validation Phase

Similarly to demonstration 1, in the final validation phase, the designs from

the Pareto front were selected and analysed (shown as grey dots on Figure

7).The results confirmed the performance improvements with respect to

the base case shown in Figure 4.

5. DISCUSSION

The results from the two demonstrations are compared at three different

levels: in terms of the performance of design variants; in terms of the visual

differentiation of the design variants; and in terms of the speed of execution

of the optimisation algorithm. In order to aid in the comparison of design

variants, selected designs from the Pareto fronts for the two

demonstrations are shown in Figure 8.

5.1 Performance of design variants

The total electricity savings has a range from worst to best of - 2,102

kWh·a-1 to -131 kWh·a-1.For the first demonstration, the total electricity

savings for the design variants on the Pareto front ranged from -131 kWh·a-

1 to -1,104 kWh·a-1.For the second demonstration, the total electricity

savings for the design variants on the Pareto front ranged from -647 kWh·a-

1 to -1,682 kWh·a-1.The base case shown in Figure 4 has a total electricity

savings of -795 kWh·a-1.

The key observation from these results is the significant and consistent

drop in the fitness score for the second demonstration.This is primarily due

to a drop in daylight autonomy as a result of occupants closing blinds to

counteract glare. For the first demonstration, the best design variant had a

daylight autonomy of 68.2%,while for the second demonstration, the best

design variant had a daylight autonomy of only 18.7%.This reduction in

daylight autonomy is the main reason why the best design variant from the

first demonstration has a total electricity savings that is 81% greater than

the best design variant from the second demonstration.

This highlights that without considering glare in the optimisation

process, there is an overestimation of 81% of total electricity savings.The

second demonstration using the enhanced daylight savings model therefore

gives a more accurate indication of final performance that is closer to the

reality.

96 Thian-Siong Choo and Patrick Janssen

Figure 8.Top:Selected design variants on the Pareto front and corresponding performance bar charts for demonstration 1. Bottom:

Selected design variants on the Pareto front and corresponding performance bar charts of demonstration 2.

97Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

5.2 Visual differentiation of design variants

Besides considering the building performances, the aesthetical implications

of the different design variants are also likely to be of interest to designers.

In general, the design variants on the Pareto front vary considerably from

one another. For example, Figure 9 shows two design variants from the

second demonstration; design variant 27 shows a more open view out of

the space compared to design variation 861.The designer may decide to

choose design variant 27 despite the lower performance.

Figure 9. Left: Interior perspective

of the 861st design variation from

demonstration 2. Right: Interior

perspective of the 27th design

variation from demonstration

98 Thian-Siong Choo and Patrick Janssen

With respect to visual differentiation, it is noteworthy that for the

optimal design from the first and second demonstrations, the difference

between the two design variants is relatively small, with both designs having

very densely packed PV cells (see Figures 5 and 8).This suggests that it may

not be necessary to use the enhanced daylight savings model that takes into

account occupants’ use of blinds, since in the end, both daylight savings

models converge on similar designs.

5.3 Speed of Execution

For both demonstrations, the use of the three-phase method led to

significant reduction in run times. In the case of the first demonstration, a

base-case optimisation was run.The total run time for the base case was 14

days, while the total run time using the three-phase method was 2 days and

17 hours.This is a reduction of 81%.

In the case of the second demonstration, no base case was run as it

would have taken a prohibitively long time. Based on the execution times of

the individual simulations, it is clear that an optimisation run not using proxy

models would have taken years to complete.The total run time using the

three-phase method was 14 days and 17 hours.

Comparing the two demonstrations, the run times using the three-phase

method differed significantly.The enhanced daylight savings model,and in

particular the Evalglare simulation used within that model resulted in a

543% increase in run time.

This increase in run time further brings into question the need for the

enhanced daylight simulation model. If a more accurate prediction total

electricity savings is required,it would still be possible to optimize using the

normal daylight savings model (as performed in the first demonstration), and

to then re-simulate the final selected designs using the enhanced daylights

savings model.

6. CONCLUSION

The use of the three-phase method for optimizing semi-transparent BIPV

facades with conflicting performance objectives has been demonstrated.The

two demonstrations highlighted the complexities involved in trying to

reduce the run time of the evolutionary optimisation algorithms.The results

from the demonstrations suggest that the most feasible approach may be to

optimise designs without taking into account the use of blinds, and then to

simulate the use of blinds to counteract glare during the validation stage.

However, even without taking into account the use of blinds,the run time of

the evolutionary algorithm was still over two and a half days long. For most

designers, this would be considered very long, especially if the intention was

to apply these methods in early stage design. In order to ensure that such

optimisation methods can be effectively applied, both the complexity of

setting up the models and the overall run times of the optimization

algorithms will need to be significantly reduced. Future research will

investigate parallel execution of the optimisation algorithm and replacement

of the simulations with surrogate models.

REFERENCES

1. Fung,T.Y.Y. and Yang, H., Study on thermal performance of semi-transparent building-

integrated photovoltaic glazings,Energy and Buildings, 2008, 40(3), 341-350.

2. Robinson, L.E. and Athienitis,A.K., Design methodology for optimisation of electricity

generation and daylight utilization for facades with semi-transparent photovoltaics,

Proc. Building Simulation, 2009.

3. Rutten, D., Evolutionary Principles Applied to Problem Solving”,

http://ieatbugsforbreakfast. wordpress.com/ 2011/03/04/epatps01 [15-10-2012].

4. Attia, Shady, Mohamed Hamdy,William O’Brien, and Salvatore Carlucci, Assessing

gaps and needs for integrating building performance optimization tools in net zero

energy buildings design,Energy and Buildings, 2013, 110 - 124.

5. Wienold, J. and Christoffersen, J., Evaluation methods and development of a new

glare prediction model for a new glare prediction model for daylight environments with

the use of CCD cameras,Energy and Buildings, 2006, 38,743-757.

6. Janssen, P. and Kaushik,V., Iterative Refinement through Simulation: Exploring trade-offs

between speed and accuracy,Proc. 30th eCAADe Conference - Vol. 1, Czech

Technical University in Prague, Faculty of Architecture (Czech Republic) 12-14

Sep 2012, 555-563.

99Evolutionary Optimisation of Semi-transparent Building Integrated Photovoltaic Facades

7. Reinhart, C.F., Tutorial on the Use of Daysim Simulations for Sustainable Design,

Harvard University, Cambridge, 2010.

8. Wienold, J., Dynamic daylight glare evaluation,Proc. 11th International IBPSA

Conference, Glasgow, Scotland, 2009, 944-951.

9. McNeel, R., Rhinoceros – NURBS Modelling for Windows (version 4),McNeel North

America, Seattle, 2010.

10. Janssen, P., and Chen,K.W., Visual Dataflow Modelling: a Comparison of Three

Systems,Proceedings of the 14th International conference on Computer Aided

Architectural Design, CAAD Futures, Liège: Les Éditions de l’Université de Liège,

2011, 801-816.

11. Jakubiec, A., and Reinhart, C., DIVA-FOR-RHINO 2.0: Environmental parametric

modeling in Rhinoceros/Grasshopper using Radiance, Daysim and EnergyPlus,Building

Simulation 2011, Sydney.

12. Crawley, D.B., Lawrie, L.K.,Winkelmann, F.C., Buhl,W.F., Huang,Y.J. and Pederson,

C.O., EnergyPlus: creating a new generation building energy simulation program,

Energy and Buildings,2001, 33(4),319-331.

13. NREL, Weather Data,

http://apps1.eere.energy.gov/buildings/energyplus/weatherdata_about.cfm [15-10-

2012]

14. EnergyPlus, Energy: Photovoltaic,Input-Output Reference, EnergyPlus

Documentation, U.S. Department of Energy, Berkeley, 2011.

15. Halonen, L.,Tetri, E. and Bhusal,P., Guidebook on Energy Efficient Electric Lighting for

Buildings,Espoo, Sweden, 2010.

16. Protogeropoulos, C., Zachariuu,A., Photovoltaic Module Laboratory Reflectivity

Measurements and Comparison Analysis with Other Reflecting Surfaces,Proc.25th

European Photovoltaic Solar Energy Conference 2010,Valencia, Spain.

17. SPRING, Singapore Standard SS530 : 2006 – Code of Practice for Energy efficiency

standard for building services and equipment,SPRING, Singapore, 2006.

18. Pilkington, Pilkington Glass Handbook 2010,Nippon Sheet Glass Pte Ltd, Japan,

2010.

19. Choo,T. S., Ouyang, J. and Patrick Janssen, 2013,Multi-objective Optimisation of

Semi-transparent Building Integrated Photovoltaic Facades,Proceedings for

Sustainable Building Conference 2013, Singapore.

100 Thian-Siong Choo and Patrick Janssen

Thian-Siong Choo and Patrick Janssen

Department of Architecture

School of Design and Environment

National University of Singapore

4 Architecture Drive

Singapore 117 566

Thian-Siong Choo, thiansiong.choo@nus.edu.sg, choothiansiong@yahoo.co.uk;

Patrick Janssen, patrick@janssen.name