PERFORMANCE-BASED

FACADE DESIGN TOOL:

Approach for Automated and Multi- Mahsa Minaei
Objective Simulation and Optimization University of Massachusetts Amherst

Abstract: Buildings have a considerable impact on the environment, and it is crucial to consider environmental
and energy performance in building design. Buildings account for about 40% of the global energy consumption
and contribute to over 30% of the CO2 emissions. A large proportion of this energy is used for meeting occupants’
thermal comfort in buildings, followed by lighting. The building facade forms a barrier between the exterior and
interior environments; therefore, it has a crucial role in improving energy efficiency and building performance.
In this regard, decisionmakers are required to establish an optimal solution, considering multi-objective problems
that are usually competitive and nonlinear, such as energy consumption, financial costs, environmental performance,
occupant comfort, etc. Sustainable building design requires considerations of many design variables and multiple,
often conflicting objectives, such as the initial construction cost, energy cost, energy consumption, and occupant
satisfaction. One approach to address these issues is the use of building performance simulations and optimization
methods.
This research presents a novel method for improving building facade performance, taking into consideration occupant
comfort, energy consumption and energy costs. The research discusses development of a framework, which is
based on multi-objective optimization and uses a Genetic Algorithm (GA) and machine learning in combination with
building performance simulations. The framework utilizes the EnergyPlus simulation engine and custom scripts
using Python programming to implement optimization algorithm analysis and decision support. The framework
is automated in all steps: generating design scenarios, sending scenarios to the simulator, collecting the specific
output, and decision making in optimization phase. So, the framework enhances the process of performance-based
facade design, couples simulation and optimization packages, and provides a flexible and fast supplement in the
facade design process by rapid generation of design alternatives.
The study describes the components and functionality of this framework in detail, as well as a two-step optimization
technique, which is a new technique that combines GA and Machine Learning. This technique improves the
framework speed, performance, and stability of an artificial neural network (ANN) and reduces the sensitivity.
The case study for a test cell presents, illustrating how the framework is used to test a variety of design possibilities
and validation of this framework, as well as its application for facade design in different climates.

Keywords: Performance-based facade design, simulation-based optimization, machine learning, minimum viable
product

INTRODUCTION

The buildings and building construction sectors improving energy efficiency and building performance.
combined are responsible for 36% of global final Therefore, this research focuses on performance-based
energy consumption and nearly 40% of total direct and facade design, appropriate simulation and optimization
indirect CO2 emissions (IEA 2019). Energy demand tools, and methods for design analysis and support.
from buildings and building construction continues to Building performance simulation (BPS) provides
rise, driven by improved access to energy in developing relevant design information by indicating potential
countries, greater ownership and use of energy- (quantifiable) directions for design solutions. BPS
consuming devices, and rapid growth in global buildings’ tools and applications facilitate the process of design
floor area, at nearly 3% per year (IEA 2019). A large decision-making by providing quantifiable data about
proportion of this energy is used for meeting occupants’ building performance. BPS tools are an integral part
thermal comfort in buildings, followed by lighting. The of the design process for energy efficient and high-
building facade forms a barrier between the exterior performance buildings, since they help in investigating
and interior environments, and has a crucial role in design options and assess the environmental and energy

233
Performance-Based Facade Design Tool

impacts of design decisions (Attia 2013). The important METHODOLOGY

aspect is that simulation does not generate design
The new framework for performance-based design
solutions, instead, it supports designers by providing
approach, aiming to minimize building energy
feedback on performance results of design scenarios.
consumption and energy cost, while considering
Optimization is a method for finding a best scenario
occupants’ comfort level, was developed as part of this
with highest achievable performance under certain
research. This is a modular framework, consisting of
constraints and variables. There are different methods
independent scripts that represent modules, steps, and
for optimization, requiring use of computational
functions of application under test. The modules are
simulation to achieve optimal solution, or sometimes
used in a hierarchical fashion to apply the framework,
requiring analysis or experimental methods to optimize
consisting of five steps:
building performance without performing mathematical
optimization. In the BPS context, however, the term 1. Defining goals, performance criteria, facade design
optimization generally indicates an automated process variables, and their properties, acceptable ranges
that is entirely based on numerical simulation and for high-performance facade design.
mathematical optimization (Nguyen 2014). Integrating 2. Generating the database that includes all possible
BPS and optimization methods can form a process design scenarios based on the variables with
for selecting optimal solutions from a set of available permutation in Python and selected outputs after
alternatives for a given design problem, according to a simulation in EnergyPlus. This is module 1.
set of performance criteria. 3. Coupling Python script with simulation engine
This paper focuses on developing a new framework (EnergyPlus) to automatically perform simulations
for performance-based facade design. The framework for scenarios from database (measurements
considers energy consumption, occupant comfort, methods) to quantify variables and generate the
and energy cost optimality, and implements BPS and needed outputs. This is module 2.
relevant optimization methods for performance-based 4. Optimization phase by implementing Python script,
facade design. The components and development of the genetic algorithm (GA) and Batch Normalization to
framework are discussed in detail. evaluate outputs and find the optimal scenarios.
This is module 3.

Figure 1: Conceptual diagram, showing components of the

framework. (Author 2019)

234
 Mahsa Minaei

5. Developing a front-end for user to test the Step 2: Creating the Database (MYSQL)
framework and collect the data for next step, which
After setting variables and parameters for facade
is implementing deep learning after collecting
design, all possible scenarios are generated using
enough data.
Python programming. With the permutation in Python
The next sections discuss the model development, script, design scenarios are generated and added to the
then components of the framework and its implementation database with a specific scenario ID. In this study, we
in detail will be illustrated. have 38,400 scenarios to investigate for the test cell,
described in the next section. After running simulation
Step 1: Defining Optimization Objectives, Performance in EnergyPlus, all outputs in Step 3 are populated in this
Criteria and Facade Variables database with identical scenario ID. EnergyPlus provides
Figure 1 shows the components of the framework. a wide range of outputs, but, for this purpose, the
Performance-based facade design requires a holistic following results are obtained:
approach, considering performance indicators, such • Cooling, heating, and lighting loads, Energy Use
as energy performance and human comfort. These Intensity (EUI) for electricity and gas, PMV and
performance requirements (variables) must be PPD, and total energy costs for electricity and gas.
quantified. The goals (optimization objectives) for • Module 1 is responsible for generating all
this framework are to aid the design decision making scenarios with defined variable and populating
process, where energy consumption and cost are these scenarios in the database.
minimized, and occupant comfort (thermal and visual) • Module 2 is responsible for automatically sending
is maximized. The energy requirements for heating, these scenarios to the simulation engine and for
cooling, and lighting of buildings are strongly driven by populating the selected outputs in the database.
the performance of the facade, especially the glazing. • Data Flow Diagram (DFD) in figure 2 shows the
The objectives for reducing energy consumption overview of the framework system that represents
are to reduce heating, cooling, and lighting loads. the flow of data through this process.
Performance requirements (variables) to meet
this objective are window to wall ratio (WWR), wall The database manages all scenarios’ inputs
assembly, insulation, solar control, and glazing system. and outputs and is MYSQL, which is an open source
Performance-based facade design objectives that are relational database management system (RDBMS)
related to human factors and contribute to occupant that uses Structured Query Language (SQL) for adding,
comfort and satisfaction in buildings include thermal accessing and managing content in the database. The
comfort and visual comfort. The variables that relate advantages of this type of database for the purpose
to facade design include air temperature, mean radiant of this research is the quick processing time, proven
temperature, air movement, relative humidity, clothing reliability, open source, ease and flexibility of use.
layers and activity levels. The Predictive Mean Vote
Step 3: Coupling Python scripts with Simulation
(PMV) suggested by Fanger (1970) predicts the effects
Engine (EnergyPlus)
of these six factors on thermal comfort. Predicted
Percentage of Dissatisfied (PPD) persons predicts the EnergyPlus 8.5 is used in this research as an energy
percentage of people who would feel discomfort with modeling engine. EnergyPlus has been chosen as
certain thermal conditions. This research investigates a BPS tool for two main reasons: (a) this program
how objectives are treated, what approach is more allows reliable modeling of both building and HVAC
desirable and how to deal with constrained problems. systems, and (b) it works with text-based inputs and

Table 1: List of fixed parameters

235
Performance-Based Facade Design Tool

outputs, and these facilitate the interaction with Python is a combination of batch normalization, which is an
scripts. EnergyPlus can investigate discussed variables Artificial Neural Network (ANN) technique, and flood
as inputs and simulate envelope related outputs in the fill algorithm. ANNs are effective methods that imitate
study. Thermal comfort is calculated based on PMV and the complex relationship of the network to solve multi
PPD. The formulas for both PMV and PPD are built into objectives and non-linear problems. ANNs resemble the
EnergyPlus and their values can be obtained directly biological neural system, composed of layers of parallel
from the simulation output file. neurons and weighted links. They learn the relationship
Initial simulation test cell considered a single between the input and output variables by studying
office space (40’x40’x10’), located in Atlanta, Georgia. previously recorded data. The layer that produces the
The south-facing facade was used to develop network output are usually called the output layer, and
different design scenarios, varying WWR, materials, all other layers are called hidden layers.
glazing system, and shading control. Defining related The optimization method in this study is a
parameters as inputs and setting data needed for combination of GA and ANN, which is a machine
outputs are the primary method for connecting design learning technique. The GA in combination with flood
scenarios in the database with the simulation engine. fill algorithm and batch normalization creates a new
Python script works as an interface to call scenarios technique to find a relation between the outputs, to
from the database and to send them to the simulator. assign weights, and dynamically adjust the target
Each parameter must identify a well-defined relation position to find optimal scenarios.
with discussed variables, which reveals facade behavior A batch normalization technique is used for the
in relation to performance aspects being analyzed. first phase of optimization. This is a technique for
improving speed, performance, and stability of artificial
Step 4: Optimization Phase by Implementing GA and neural network (ANN) by adjusting and scaling the
Machine learning activations. The batch normalization was introduced
In building optimization studies using GA, the in 2015 by Ioffe and Szegedy. The intention behind
computational time is generally reduced by two batch normalization is to optimize network training.
methods. First, is the use of a very simplified model Several benefits of this methods are: faster training,
instead of complete simulation. This simplification higher learning rates, reduced sensitivity, and easier
can cause inaccurate modeling of the building. The methods to initialize and produce better results (Hinton
other method is selecting a very small size for GA 2012). This technique, combined with a flood field
populations or a relatively small number of generations. algorithm, facilitates the optimization by sorting the
One efficient solution to reduce the computational time highest indicators and decides which scenarios must be
associated with GA algorithm is use of machine learning simulated.
techniques to reduce time and increase the accuracy of The flood fill algorithm takes three parameters:
the results. The machine learning used in this research start node, target, and replacement, and determines the

Figure 2: Data Flow Diagram (DFD) of the framework. (Author

2019)

236
 Mahsa Minaei

area connected to our target. This algorithm facilitates

the optimization by sorting the highest indicators and
decides which scenarios must be simulated based on
the specific scenario ID. Using this algorithm decreases
the process time, because it is not necessary to
simulate all scenarios—rather, only scenarios that are
closer to the target. The comparison is based on the
assigned indicator value. In a dynamic system, it is
necessary to scale indicators to represent the impact
of the indicators so as to configure following tasks
and converge the results to the goal based on these
scores. Figure 3 shows a sample for scoring total EUI
electricity, EUI gas, PMV and energy cost indicators.
These indicators work as fitness functions in genetic
algorithms, which is a particular type of objective
function to summarize and guide the simulations
towards optimal design solutions. These indicators or
fitness functions must correlate closely to the goals
and must be computed quickly, because it needs to be Figure 3: Total EUI-Electricity (MJ/m2), EUI-Gas, PMV and Energy
iterated many times in order to produce usable results. Cost indicator scores. (Author 2019)
The initial population is generated randomly, based
on the range of possible design scenarios. It is sent to
the simulator to run the initial calculations, and then
results are returned to the database to compare with
the goals and standards. Then, design scenarios that
have results closer to the goals are kept, and others
are removed. In this framework, the goal is summation
of three indicators, for energy consumption, comfort,
and cost. The indicators are dynamically updated
based on the range of results. The top left chart in
figure 3 shows an example, where indicators from 6 to
-3 are used for the initial test cell energy consumption
Figure 4: Total Indicators vs. Scenario IDs (for 2,061 scenarios).
results. Occasionally, the solutions may be “seeded” in (Author 2019)
areas where optimal solutions are likely to be found.
Individual solutions are selected through a fitness-based
process, where fitter solutions (as measured by a fitness
function) are typically more likely to be selected. This
method accelerates the simulation process and the
results give us clusters of optimized scenarios for
analysis in next phase of optimization. Figures 4 and 5
show how the optimization algorithm selects and sorts
the fitted results for this framework.
Figure 4 shows the results before applying
optimization for processing 2,061 scenarios and figure
5 shows the result of 18,103 scenarios with assigning
the first step of optimization. In this case, we have Figure 5: Total Indicators vs. Scenario IDs (for 18,103 scenarios).
1,627 scenarios that scored 20 and more than 20 (Author 2019)
(1,591 scenarios at 20 and 36 more than 20). Using
this process decreases the processing time, because it The next step of optimization is applying
is not necessary to simulate all scenarios—rather, only integrated correlation matrix clustering as a dropout
scenarios that are closer to the target. After running all technique, then comparing the results. This dropout
scenarios (38400 ID) and applying a batch normalization is a regularization technique for reducing overfitting in
technique, 3,164 scenarios are selected. The next step neural networks by preventing complex co-adaptations
focuses on comparison of results. on training data. Dropout refers to dropping out units

237
Performance-Based Facade Design Tool

Figure 6: Correlation Matrix. (Author 2019)

(hidden or visible) in the neural network. In machine

learning, correlation clustering provides a method for
clustering a set of objects into an optimum number of
clusters based on their similarity. So, in a correlation
matrix, the relationship between the objects (variables)
are known instead of the actual representations of the
objects. Figure 6 shows the correlation matrix based
on output data, integrated with optimization method to
sort the results. Figures 7 and 8 show the final results of
all scenarios with both techniques implemented. Figure
Figure 7: Results for all scenarios with applying batch
7 represents the first phase of optimization, and figure normalization and flood field algorithm (Phase 1 optimization).
8 shows the second phase after applying a correlation (Author 2019)
matrix and batch normalization. Results show that
process time, performance, and accuracy are improved
by using this method.

Step 5: Developing a front-end for user to test the

framework and collect the data for next step which is
implementing deep learning after collecting enough
data.
This research studies simulation-based optimization
methods and develops the framework for facade design
Figure 8: Results for all scenarios with applying correlation
that is automated and couples simulation engine with matrix and cluster eliminating (Phase 2 optimization). (Author
optimization algorithms. This framework can be used 2019)

238
 Mahsa Minaei

Figure 9: Data Flow Diagram L0 (Author 2019)

as a backend for web or mobile application (or any user

interfaces) and, eventually after collecting enough data,
the deep learning can be implemented to predict the
configurations related to outputs.
Figure 10 represents the whole model development
for this facade design tool. The focus of this model
and framework is developing the MVP (Minimum
Viable Product), which is the core feature to effectively
deploy the product to the users. MVP can be part of
Figure 10: MVC flow chart (Author 2019)
the process directed toward making and selling this
product to the users (Raddof 2014). It works as object
method are providing clean separation of concern (SoC)
in this iterative process of generation, presentation,
and enabling test driven development (TDD).
data collection and analysis and learning. Creating MVP
Model represents the shape of data of the
will allow collecting the maximum amount of validated
application. It is a central component that directly
learning data from users.
manages the data, logic, and rules of application and is
In this research, MVP is a web application product
independent of the user interface. It receives user input
for users to test and collect data for a next development
from the controller. View is a user interface that displays
that needs big data for deep learning. This study mainly
data, so users are able to modify it. View actually works
focuses on developing a framework as a back end and
as the front-end in this case. All result representations,
a simple front end for users’ interface. Figure 9 shows
such as charts, diagrams, tables, or any other specific
the level 0 data flow diagram (DFD), which is known as a
forms can be displayed here. Controller handles the user
context diagram and shows a data system as whole and
request and renders the appropriate view with the model
the way it interacts with external entities.
data as a response. In other words, it accepts inputs and
In order to develop the product based on this
data and converts them to commands for the model
research, test the iteration, and collect the results and
or view. Figure 10 illustrates MVC architecture and the
outputs, we need a user interface as the front-end
flow of users’ requests and responses in this case study.
that is connected to a developed backend. For this
The interaction between the front end and back end with
purpose, the MVC method is applied. MVC is a software
simulation engine and creating the configurations.
and application design pattern used for developing
interfaces and stands for Model, View, Controller as
CONCLUSION AND FUTURE WORK
three separated interconnected elements. MVC is a
lightweight highly testable framework as compared to This research discussed the role of simulations and
traditional ASP.NET web forms. This method separates optimization in the design decision-making process.
content from presentation and data processing from Then, a novel performance-based facade design
content. In other words, design pattern keeps the display framework was described, where different performance
and data separate to allow each to change without criteria and variables have been defined for achieving
affecting the other and enable full control over the energy efficiency, occupant comfort, and cost optimality.
rendered HTML. So, the main advantages of the MVC The framework has been implemented by coupling

239
EnergyPlus as a simulation engine, and custom scripts interface. In addition to developing the user interface
using Python programing language. Then a user and web application, this product will be developed to
interface was developed with an MVC method to serve accept any IDF files, the users will be able to choose
as a front end, in order to test and collect data. The their variable for optimization and then the rest of the
research describes the components and functionality process will be automated and results represented.
of this framework in detail, as well as the two-step After releasing the open source application,
optimization technique. A case study for a test cell was other important developments for this research will be
presented, illustrating how the framework is used to test collecting the data for implementing deep learning on
a variety of design possibilities. the data and outputs from different iterations, so as to
Future work will focus on the application for facade enable the correlation matrix to generate automatically.
design in different climates and developing the user

REFERENCES
Aksamija, Ajla. 2013. Sustainable Facades: Design Methods for High-Performance Building Envelopes. New York: John Wiley & Sons.
Attia, Shady, Mohamed Hamdy, William O’Brien, and Salvatore Carlucci. 2013. “Assessing Gaps and Needs for Integrating Building
Performance Optimization Tools in Net Zero Energy Buildings Design.” Energy and Buildings 60 (May): 110-124.
Evins, Ralph. 2013. A Review of Computational Optimization Methods Applied to Sustainable Building Design. Renewable and
Sustainable Energy Reviews 22 (June): 230-245.
Evins, Ralph, Philip Pointer, and Stuart Burgess. 2012. “Multi-Objective Optimization of a Modular Building for Different Climate
Types.” In First Building Simulation and Optimization Conference, Loughborough, 173-180.
Fanger, P.O. 1970. Thermal Comfort. Analysis and Applications in Environmental Engineering. Copenhagen: Danish Technical Press.
Gosavi, Ab. 2015. Simulation-Based Optimization. Berlin: Springer.
Hinton, Geoffrey E., Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R. Salakhutdinov. 2012. “Improving Neural Networks
by Preventing Co-Adaptation of Feature Detectors.” arXiv preprint arXiv:1207.0580.
Huang, Yu, and Jian-lei Niu. 2016. “Optimal Building Envelope Design Based on Simulated Performance: History, Current Status, and
New Potentials.” Energy and Buildings 117 (April): 387-398.
Ioffe, Sergey, and Christian Szegedy. 2015. “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate
Shift.” arXiv preprint arXiv:1502.03167.
Li, Xiwang, Jin Wen, and Er-Wei Bai. 2016. “Developing a Whole Building Cooling Energy Forecasting Model for On-Line Operation
Optimization Using Proactive System Identification.” Applied Energy 164 (February): 69-88.
Machairas, Vasileios, Aris Tsangrassoulis, and Kleo Axarli. 2014. “Algorithms for Optimization of Building Design: A Review.”
Renewable and Sustainable Energy Reviews 31(March): 101-112.
Nguyen, Anh-Tuan, Sigrid Reiter, and Philippe Rigo. 2014. “A Review on Simulation-Based Optimization Methods Applied to Building
Performance Analysis.” Applied Energy 113 (January): 1043-1058.
Ostergård, Torben, Rasmus L. Jensen, and Steffen E. Maagaard. 2016. “Building Simulations Supporting Decision Making in Early
Design–A Review.” Renewable and Sustainable Energy Reviews 61 (August): 187-201.
Ries, Eric. 2011. The Lean Startup: How Today’s Entrepreneurs Use Continuous Innovation to Create Radically Successful Businesses.
New York: Crown Books.
Shi, Xing, Zhichao Tian, Wenqiang Chen, Binghui Si, and Xing Jin. 2016. “A Review on Building Energy Efficient Design Optimization
from the Perspective of Architects.” Renewable and Sustainable Energy Reviews 65 (November): 872-884.
Wang, Haidong, and Zhiqiang John Zhai. 2016. “Advances in Building Simulation and Computational Techniques: A Review Between
1987 and 2014.” Energy and Buildings 128 (September): 319-335.