See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361107709

Synthetic Machine Learning for Real-time Architectural Daylighting

Prediction

Conference Paper · April 2022

DOI: 10.52842/conf.caadria.2022.1.313

CITATIONS READS

4 298

8 authors, including:

Rutvik Deshpande Maciej Nisztuk

Digital Blue Foam Wroclaw University of Science and Technology
7 PUBLICATIONS 6 CITATIONS 17 PUBLICATIONS 69 CITATIONS

SEE PROFILE SEE PROFILE

Cesar Cheng Ramanathan Subramanian

digital blue foam Singapore University of Technology and Design
6 PUBLICATIONS 8 CITATIONS 9 PUBLICATIONS 19 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Rutvik Deshpande on 04 November 2022.

The user has requested enhancement of the downloaded file.

SYNTHETIC MACHINE LEARNING FOR REAL-TIME
ARCHITECTURAL DAYLIGHTING PREDICTION

RUTVIK DESHPANDE1, MACIEJ NISZTUK2, CESAR CHENG3,

RAMANATHAN SUBRAMANIAN4, TEJAS CHAVAN5, CAMIEL
WEIJENBERG6 and SAYJEL VIJAY PATEL7
1,2,3,4,5,6,7
Digital Blue Foam.
1,2,3,4,5,6,7
{rutvik|maciej|cesar|ram|tejas|camiel|sayjel}@digital-
bluefoam.com

Abstract. “Synthetic Machine Learning” offers a revolutionary leap

in real-time environmental analysis for conceptual architectural design.
By integrating automatic synthetic data generation, artificial neural
network (ANN) training and online deployment, Synthetic Machine
Learning offers two main advantages over conventional simulation;
First, it reduces the analysis time for a reference simulation from
minutes to seconds; Second, it is possible to deploy ANN as a web
service in an online design environment, which therein increases
accessibility, significantly reducing simulation costs and setup time.
The application of Synthetic Machine Learning to perform Daylight
Autonomy (DA) and Spatial Daylight Autonomy (sDA) studies to
maximise building daylighting for a given use, window to wall ratio,
and floorplan arrangement is showcased through a preliminary
demonstration work. Comparatively the use of algorithmically
generated synthetic data versus real-world data is becoming ubiquitous
in other disciplines, the advantages of this approach to the building
design process are further discussed.

Keywords. Daylight Autonomy, machine learning, building energy

performance, synthetic data-sets; SDG 7; SDG 11.

1. Introduction
Synthetic Machine Learning is part of a revolution in environmental analysis for
architectural design. For the first time, no-code Machine Learning (ML) platforms,
such as Google Cloud, allow anyone to use their data to make fast, automatic design
predictions. With the easing of barriers to ML model creation, non-experts will be able
to access, create, share, and apply a vast set of bespoke ML tools and services.
Presently, simulations are the go-to solution for building performance analysis.
Despite broad acceptance, there are several technical obstacles that designers must
overcome when performing a simulation; without a domain-level understanding of the
simulation mathematical models, the results are vulnerable to misinterpretation.

POST-CARBON, Proceedings of the 27th International Conference of the Association for Computer-Aided
Architectural Design Research in Asia (CAADRIA) 2022, Volume 1, 313-322. © 2022 and published by the
Association for Computer-Aided Architectural Design Research in Asia (CAADRIA), Hong Kong.
314 R. DESHPANDE ET AL.

Furthermore, simulations require tremendous computational resources and can be

extremely time-intensive. This conflicts with the early-stage design process, where
designers perform several rounds of design iteration and refinement.
In recent years, research interest in using ML is rapidly gaining momentum. It is
being explored as an alternative to simulation and as a method for building
performance analysis (Fedorova et al., 2021, Seyedzadeh et al., 2018). This is because
ML models rely on observations of known outcomes to make fast, automatic
predictions of building performance metrics, rather than applying the expert-crafted,
time-intensive mathematical models of simulations.
Despite these benefits, ML algorithms require access to large data sets to learn and
predict outcomes. Unlike digital businesses like finance, which generate billions of
gigabytes of data every day, the architecture, engineering, and construction (AEC)
industries face a unique problem: a majority of building-related datasets are confined
in the real world. As a result, new approaches for acquiring building-related data sets
are intended to better utilize ML's potential. Synthetic Machine Learning aims to fill
this gap by incorporating synthetic data (Rubin, 1993), which is algorithmically
generated data via computer simulation, into the ML model's inception.
The conceptual framework and technological implementation of a novel Synthetic
Machine Learning system are showcased in the paper. The proposed system offers two
major advantages over conventional simulations by coupling automatic data-set
generation with neural network training: First, it reduces analysis time for a benchmark
simulation from several minutes to a few seconds; Second, it can be deployed as a web
service within an online design environment, increasing accessibility by significantly
reducing simulation cost and setup time. A preliminary concept of developing a web
service to perform Daylight Autonomy (DA) and Spatial Daylight Autonomy (sDA)
analysis (Reinhart et al., 2006) within an online design platform — Digital Blue Foam
(DBF) — is presented to demonstrate its capabilities and features. DBF has the
capabilities to run sustainability analysis based on CFD, generative design and urban
insights. Existing daylighting tools are cumbersome and difficult to set up. The
platform has been built with the intent of the integrated design, reduced computation
time and collaboration in mind. The advantage of this approach is informed design
recommendations with minimum effort and flexibility offered to the user in generating
context-specific massing options.

2. Related Work
The use of synthetic data and ML in architectural design is becoming more prominent.
The background and recent work on the application of such methodologies in the
assessment of building daylighting are presented in this section. It also encompasses a
comprehensive research summary on the application of ML and daylighting study to
assess building energy performance.

2.1. DAYLIGHT AUTONOMY

 SYNTHETIC MACHINE LEARNING FOR REAL-TIME 315
ARCHITECTURAL DAYLIGHTING PREDICTION

The DBF platform can perform traditional analysis (shadow and sun hours studies) to
understand the daylighting performance of the building exterior and its urban context.
To better understand the building internal performance, the DA and sDA analysis were
used. DA is a standard metric to describe the quality of building daylighting. DA is the
percentage of time a given space remains above the threshold illuminance (300 lux for
general office occupancy) throughout the year (Reinhart et al., 2001). sDA is a metric
to support DA. sDA is the percentage of the floor area of the given space, which
receives sufficient daylight annually. Sufficient daylight is generally a minimum of 300
lux for at least 50% of the annual occupied hours (Heschong et al., 2012). The DA and
sDA analysis is directly linked to the size of the windows. The “window-to-wall ratio”
(WWR) directly determines the amount of sunlight entering the building's interior. The
connection of DA and sDA with the size of the windows allows for the practical use of
the analysis into building's energy performance assessment as illustrated in Figure 1.
At any time during a given day, the amount of daylight entering a room can be tailored
by suggesting the optimal size, location, and relative positioning of windows. For
example, placing smaller windows in hot-dry climates would be preferable for
sufficient daylight and to avoid glare and solar heat gain. The impact of those factors
combined, allows economically justified decisions from the start of the design stage.

Figure 1. Daylight Autonomy and Spatial Daylight Autonomy as metrics to quantify daylighting

2.2. DAYLIGHT SIMULATIONS

Presently building designers work with expert consultants who use advanced
simulation software to calculate DA and sDA through computational simulation with
the required parameters. This approach uses a local climate file such as an EnergyPlus
weather file (EPW file), running hour-by-hour calculation of annual illuminance
(amount of light received on the surface) in space with lighting simulation software.
The typical workflow uses the existing libraries available in tools, like Rhinoceros-
grasshopper's Ladybug plugin for Sketchup De Luminae (De Luminae, 2021)
extension, to simulate the DA values. Both DA and sDA are quite common and used
widely, but sDA has become popular due to its inclusion in LEED v4 and the WELL
Building Standard. This approach can be time-consuming, laborious, and might require
expertise because the user needs to design the building models, add materials, set
targets, and correct radiance parameters. When the setup is ready, the simulation can
be performed, which might take several minutes based on the input parameters and the
complexity of geometry (Solemma LLC., 2021). The Synthetic Machine Learning
system allows the calculation of DA and sDA to be performed in real-time.

2.3. SYNTHETIC DATA GENERATION

316 R. DESHPANDE ET AL.

The physical nature of the construction industry limits access to data. This requires
different methods of obtaining information. As a result, an example of novel
methodology was devised for synthetic data generation pipeline that can generate an
arbitrary amount of 3D data, as well as the relevant 2D and 3D annotations (Fedorova
et al., 2021). Machine learning algorithms (MLAs) have been widely employed to
predict various building performance parameters since the early 2000s. A few studies
that deployed MLA for design appraisal in the field of architecture through
performance parameters such as visual comfort (Chatzikonstantinou & Sariyildiz,
2016; Nasrollahi & Shokri, 2016; Navada et al., 2016), internal daylighting
(Kazanasmaz et al., 2009; Lorenz & Jabi, 2017; Yoon et al., 2016), artificial lighting
(Bellocchio et al., 2011; Şahin et al., 2015), and building energy performance
(Chakraborty & Elzarka, 2019; Singh et al., 2019).

2.4. RESEARCH QUESTION

This paper presents a conceptual framework and an implementation of Synthetic
Machine Learning. As a result, this paper explores several questions: What advantages
does Synthetic Machine Learning offer over conventional environmental simulations?
What are the trade-offs? What are possible applications of Synthetic Machine Learning
within the building design process? What method allows for the creation and
consolidation of useful synthetic data sets for analysis? How does Synthetic Machine
learning unlock analysis to a broader audience of designers?

Figure 2. An overview of Synthetic Machine Learning system and summary of applied technologies

3. Methodology

3.1. OVERVIEW
Figure 2 provides an overview of the Synthetic Machine Learning system and
summarizes the applied technologies. The system consists of four parts: [1]
a parametric model; generating synthetic data based on environmental simulation; [2]
a ML model, trained on the generated synthetic data, to predict building DA and sDA;
[3] a web service containing a pre-trained ML model; and [4] a web-based user
interface that allows the interaction with the ML mechanics.

3.2. TECHNICAL IMPLEMENTATION

 SYNTHETIC MACHINE LEARNING FOR REAL-TIME 317
ARCHITECTURAL DAYLIGHTING PREDICTION

This section describes the technical implementation of the Synthetic Machine Learning
system for DA and sDA prediction. Figure 3 illustrates the data flow within the system.

Figure 3. The framework of Synthetic Machine Learning system for DA and sDA prediction

3.2.1. Automatic Data Generation

To generate a synthetic dataset for ML model training, the first step is to set up a
parametric model. Developed in Rhino 6, Grasshopper3D, with the LadyBug tools
plugin, the parametric model automates the generation of DA and sDA data for
different locations and design configurations. The method creates the simulation data
based on a spatial module representing the simplified building. The system creates data
that randomly cover the space of possible solutions, including the boundary conditions.
The ML model can then map between different extreme inputs and create the predicted
solutions. The spatial module used for simulation was defined by the following
parameters: width, length, height, WWR ratio for cardinal directions (North, East,
South & West), module rotation, and location (based on EPW file).
Complex typologies such as L or H shape are represented as a set of spatial
modules. The analysis final value is then computed as an average value of sub-values
of all spatial modules. With this representation, the system is also able to analyse
individual units of the building as shown in Figure 5. This building volume description
covers most of the design spectrum options at a scale of the early design stage process.
Because of that, currently, there is no option to upload custom 3D models. As this
interpretation is cuboid-based, currently the organic typologies are not supported.
One of the main challenges of the data generation process is optimizing
computation time. This is done in several ways; First, the system implements an
efficient pipeline that allows data to be created in randomly distributed batches and
processed across several computers. Second, Accelerad, (Jones, 2019; Jones &
Reinhart, 2017), a daylighting simulation package that makes use of GPU-compute
architecture to accelerate simulation time. Finally, the system simplifies building
designs to a cuboid unit-based representation.

3.2.2. Model Training

Based on research by (Lee et al., 2019), three different ML models have been tested:
Linear Regression, Random Forest Regression, and Artificial Neural Networks (ANN)
(Walczak & Cerpa, 2003). In terms of accuracy and performance with outliers and
extremities, ANN outperformed the other two models, and similar findings have been
reported in the literature (Ekici et al., 2021; Lorenz et al., 2019; Lorenz & Jabi, 2017).
318 R. DESHPANDE ET AL.

When the simulation stage was finished, the Pandas library (McKinney, 2010;
Reback et al., 2020) was used to import, clean, manipulate the data, and structure the
ML model dataset. As mentioned, the ANN was applied as the core learning
mechanism. Usage of Keras (Chollet, 2015), the deep learning API within the
TensorFlow framework (Abadi et al., 2015) facilitated the ML model setup. The Keras
API is a simplified interface between the user and core TensorFlow logics, minimizing
steps that had to be performed to create functional ML models.

3.2.3. ML Web Service

After the training process, the ML model was ready to use. The model was then
deployed on Google Cloud, a cloud computing service offered by Google. The
platform offers existing infrastructure and interfaces created for these types of use
cases. This significantly accelerates the implementation process of ML models. Using
the Google Cloud API, the ML model was directly connected to the DBF platform
which served as the user interface. Through the connection, the model evaluates the
3D building models generated by the DBF platform logic.

Figure 4. DBF tool GUI for varying WWR on various facades to perform DA and sDA analysis

3.2.4. Digital Blue Foam

The user can perform analysis on the buildings generated by the DBF platform. The
prediction is presented as a clear 3D visualization and contains information on building
energy efficiency. The colour-coded report presents the estimated DA and sDA values,
for selected WWR ratios for each facade of the 3D building. In addition, the WWR
values are represented graphically by window sizes on the corresponding building
parts. Figure 4 illustrates how DBF visualizes the results of model prediction to the
user. Figure 5 demonstrates the user interface which allows a designer to access the
web service. Figure 6 shows the use of DBF with the DA score to evaluate 3 design
options. The DBF tool was used to evaluate various forms and orientations that
maximize daylight, reduce solar heat gain, and accommodate the high-angle sun in the
tropical city of Singapore. Different design options were tested and the ML Daylight
analysis was run to compare DA & sDA metrics for the solutions.
 SYNTHETIC MACHINE LEARNING FOR REAL-TIME 319
ARCHITECTURAL DAYLIGHTING PREDICTION

Figure 5. DBF building DA prediction system is based on ML and dramatically reduces daylight
analysis time compared to traditional simulation-based approaches. The building score is given on
the UI. The building parts performance is represented as coloured units visualized on the 3D model

Figure 6. Illustration of an initial example study to demonstrate the application of the Synthetic
Machine Learning system to a hypothetical design scenario. The case study objective was to design
an office space with two 15 stories buildings located in the Downtown Singapore region. To find the
optimal building orientation, size, and WWR facade ratios, the ML-Daylighting feature was used.
The DA improved from 44% to 76% and the sDA improved from 48% to 81%. Selecting the last
option would result in less energy consumption from artificial lighting during the daytime

4. Results
The section presents the effectiveness of the ML system in measuring the building solar
energy performance. The accuracy of ML model prediction, as well as the execution
time, were measured relative to simulation. The section also discusses how the results
of model prediction are being visualized to the user.

4.1. ML PREDICTION EFFECTIVENESS

The system was created through two iterations, using different datasets and ML
models. To formulate the problem correctly, our initial spatial module had only four
features: width, length, height, and WWR ratio for the southern elevation. We
performed the simulation for the city of Mumbai. As the number of data points,
approximately 3500, is relatively low, we used Linear Regression as a learning
mechanism. With this setup, we compared predicted results with actual, simulated ones
using correlation metric, which was as good as 96.7%. Performance measurement is
visualised in Figure 7.
320 R. DESHPANDE ET AL.

In the second iteration, we added more parameters (WWR for each elevation,
rotation of the spatial module) which assured more varied data. We used the Neural
Network Regression ML algorithm as a learning method. The data was trained with 6
to 10 Neural Network layers with 100.000+ total trainable parameters. We achieved a
training loss of Mean Absolute Error (MAE) of 0.009 and Mean Squared Error (MSE)
of 0.0002. The performance kept on improving with more diverse data. We ensured
higher variability of the dataset by the greater granularity of randomly generated inputs.

Figure 7. ML DA model's prediction accuracy

4.2. COMPUTATION TIME AND SCALE-ABILITY

In the first iteration of the system, the computation time for a model was more than a
month. Optimizing the computational pipeline, simplifying building representation,
and leveraging Accelerad's GPU capabilities were shown to drastically improve
computation time. The current system can automate synthetic data generation and
model training for a single city within approximately 1 day, using a typical workstation
computer: equivalent to around 3500 data points. This number ensures the high
accuracy of our neural network model. To further improve the computation time, the
next iteration of the system will utilize distributed computing using rendering farms.

5. Discussion
Our work aimed to develop a system that allows real-time assessment of the building's
daylighting performance without the need for time-consuming simulations. We
focused primarily on the universal application of the system and the ease of its
expansion. The system uses a compact set of simulation data that can predict the DA
and sDA correctly even with a small size. Our research has shown that the data-set size
has the most significant impact on the accuracy of ML system predictions. It is possible
that when simulating other parameters, the precision of the system will not be constant.
If the accuracy is not satisfactory, we will increase the number of simulations for a
given parameter. While the system answers to the general assumptions of the
conceptual building design stage (operating on basic shapes), the current limitation can
be the simplified building representation with the cuboid units. Further work is needed
to correctly calculate daylight and generate data for irregular and organic typologies.

6. Conclusion
This paper demonstrates a Synthetic Machine Learning system to provide a real-time,
quantitative prediction of daylighting for architectural floor plan designs. The research
presented here makes the following contributions: (1) implementation of novel
 SYNTHETIC MACHINE LEARNING FOR REAL-TIME 321
ARCHITECTURAL DAYLIGHTING PREDICTION

automatic synthetic data generation pipeline, using a typical render farm configuration
and architectural software; (2) conceptual framework for deploying ML models as web
services online to assess the performance of architectural proposals; (3) comparison of
Synthetic Machine Learning analysis versus benchmark simulation, showing
significantly faster DA analysis from several minutes to seconds (Solemma LLC.,
2021) with an accuracy of 96.7% predicted results; (4) an initial demonstration of the
system to address daylight analysis of an architectural proposal in DBF. The novelty
of automatic, synthetic data generation of DA pipeline relies on: (1) the aggregation of
site-specific datasets created for different cities and implemented directly onto a
generative design tool; (2) automatic selection of the appropriate dataset based on the
chosen location; (3) the integration of this data with the DBF tenancy editor.

6.1. FUTURE WORK

This paper describes the application of Synthetic Machine Learning to create a real-
time and scalable system for performing DA and sDA assessments. Future
development of the project will expand the analysis to predict the maximum acceptable
value of WWR for each orientation, and suggest how this influences other parameters
like heat gain and cooling load. Additional evaluators of solar energy performance
within the system will also enrich the analysis, including: information on primary solar
data such as direct sun hours, annual irradiance, and radiation metrics; lighting based
real estate planning and marketing; photovoltaics performance metrics; daylight
parameters for floor plans; floor plans internal circulation patterns; urban heat island
modelling; and finally pedestrian wind comfort. While this study focuses on daylight
analysis, the approach can be applied to create a potentially limitless set of ML-based
web services for building energy, comfort, daylighting, and CFD analysis.

References
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis,
A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia,
Y., Jozefowicz, R., Kaiser, L., Kudlur, M., … Zheng, X. (2015). TensorFlow: Large-
Scale Machine Learning on Heterogeneous Distributed Systems. 19.
Bellocchio, F., Ferrari, S., Lazzaroni, M., Cristaldi, L., Rossi, M., Poli, T., & Paolini, R.
(2011). Illuminance prediction through SVM regression. 2011 IEEE Workshop on
Environmental Energy and Structural Monitoring Systems, 1–5.
Chakraborty, D., & Elzarka, H. (2019). Advanced machine learning techniques for building
performance simulation: A comparative analysis. Journal of Building Performance
Simulation, 12(2), 193–207. https://doi.org/10.1080/19401493.2018.1498538
Chatzikonstantinou, I., & Sariyildiz, S. (2016). Approximation of simulation-derived visual
comfort indicators in office spaces: A comparative study in machine
learning. Architectural Science Review, 59(4), 307–322.
Chollet, F. (2015). Keras: The Python deep learning API. https://keras.io/
De Luminae. (2021). Daylight Autonomy Extension. De Luminae. Retrieved from
https://deluminaelab.com/dl-light-manual/en/daylightautonomy.html
Ekici, B., Kazanasmaz, Z. T., Turrin, M., Taşgetiren, M. F., & Sariyildiz, I. S. (2021). Multi-
zone optimisation of high-rise buildings using artificial intelligence for sustainable
metropolises. Part 1: Background, methodology, setup, and machine learning results.
Solar Energy, 224, 373–389. https://doi.org/10.1016/j.solener.2021.05.083
 322 R. DESHPANDE ET AL.

Fedorova, S., Tono, A., Nigam, M. S., Zhang, J., Ahmadnia, A., Bolognesi, C., & Michels, D.
L. (2021). Synthetic 3d data generation pipeline for geometric deep learning in
architecture. The International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, XLIII-B2-2021, 337–344.
Heschong, L. (Chair), Wymelenberg, V. D., (Vice-Chair), K., Andersen, M., Digert, N.,
Fernandes, L., Keller, A., Loveland, J., McKay, H., Mistrick, R., Mosher, B., Reinhart, C.,
Rogers, Z., & Tanteri, M. (Ed.). (2012). Approved method: Ies spatial daylight autonomy
(Sda) and annual sunlight exposure(Ase). IES - Illuminating Engineering Society.
Jones, N. (2019). Accelerad. https://nljones.github.io/Accelerad/index.html
Jones, N. L., & Reinhart, C. F. (2017). Experimental validation of ray tracing as a means of
image-based visual discomfort prediction. Building and Environment, 113, 131–150.
Kazanasmaz, T., Günaydin, M., & Binol, S. (2009). Artificial neural networks to predict
daylight illuminance in office buildings. Building and Environment, 44(8), 1751–1757.
Lee, J., Boubekri, M., & Liang, F. (2019). Impact of Building Design Parameters on
Daylighting Metrics Using an Analysis, Prediction, and Optimization Approach Based on
Statistical Learning Technique. Sustainability, 11(5), 1474.
Lorenz, C.-L., & Jabi, W. (2017). Predicting Daylight Autonomy Metrics Using Machine
Learning. Sustainable Design of the Built Environment SDBE 2017, London, UK.
Lorenz, C.-L., Spaeth, A., Bleil de Souza, C., & Packianather, M. (2019, June 28). Machine
Learning in Design Exploration: An Investigation of the Sensitivities of ANN-based
Daylight Predictions. CAAD Futures 2019, Daejeon.
McKinney, W. (2010). Data Structures for Statistical Computing in Python. 56–61.
Navada, S. G., Adiga, C. S., & Kini, S. G. (2016). Prediction of daylight availability for visual
comfort. International Journal of Applied Engineering Research, 11(7), 4711–4717.
Nasrollahi, N., & Shokri, E. (2016). Daylight illuminance in urban environments for visual
comfort and energy performance. Renewable and Sustainable Energy Reviews, 66, 861–
874. https://doi.org/10.1016/j.rser.2016.08.052
Reback, J., McKinney, W., jbrockmendel, Bossche, J. V. den, Augspurger, T., Cloud, P.,
gfyoung, Sinhrks, Klein, A., Roeschke, M., Hawkins, S., Tratner, J., She, C., Ayd, W.,
Petersen, T., Garcia, M., Schendel, J., Hayden, A., MomIsBestFriend, … Mehyar, M.
(2020). pandas-dev/pandas: Pandas 1.0.3. Zenodo.
Reinhart, C. F., & Walkenhorst, O. (2001). Validation of dynamic RADIANCE-based
daylight simulations for a test office with external blinds. Energy and Buildings, 33(7),
683–697. https://doi.org/10.1016/S0378-7788(01)00058-5
Reinhart, C. F., Mardaljevic, J., & Rogers, Z. (2006). Dynamic Daylight Performance Metrics
for Sustainable Building Design. LEUKOS, 3(1), 7–31.
Rubin, B. (1993). Statistical disclosure limitation. Journal of official Statistics, 9(2), 461-468.
Şahin, M., Oğuz, Y., & Büyüktümtürk, F. (2015). Approximate and three-dimensional
modeling of brightness levels in interior spaces by using artificial neural
networks. Journal of Electrical Engineering and Technology, 10(4), 1822–1829.
Singh, M. M., Singaravel, S., & Geyer, P. (2019). Improving Prediction Accuracy of Machine
Learning Energy Prediction Models. Proceedings of the 36th CIB W, 78, 2019.
Seyedzadeh, S., Rahimian, F. P., Glesk, I., & Roper, M. (2018). Machine learning for
estimation of building energy consumption and performance: A review. Visualization in
Engineering, 6(1), 5. https://doi.org/10.1186/s40327-018-0064-7
Solemma LLC. (2021). Why is ClimateStudio so fast? Solemma. Retrieved August 26, 2021,
from https://www.solemma.com/climatestudio/speed
Walczak, S., & Cerpa, N. (2003). Artificial Neural Networks. In Encyclopedia of Physical
Science and Technology (pp. 631–645). https://doi.org/10.1016/B0-12-227410-5/00837-1
Yoon, Y., Moon, J. W., & Kim, S. (2016). Development of annual daylight simulation
algorithms for prediction of indoor daylight illuminance. Energy and Buildings, 118, 1–
17. https://doi.org/10.1016/j.enbuild.2016.02.030

View publication stats