Sustainable Digitization
Using Artificial Intelligence to Predict Building Daylight Autonomy
The goal of sustainable design is to reduce the negative environmental impacts of buildings while boosting the health and comfort of building occupants. While architects design buildings to be comfortable, too often this is dependent on energy-intensive systems such as heating, air-conditioning, ventilation, and lighting.
In this article, we discuss recent research looking at using a data-driven approach that uses Machine Learning to predict daylighting for building layouts. Compared to other simulation-based methods, this approach allows us to instantly analyze thousands of design options in seconds to determine the best fenestration and building layout to maximize daylighting for a given project.
Daylighting Principles
One of the key factors in a sustainable design is “Daylighting”. Daylight plays a critical role in sustainable building strategies because it is a highly cost-effective and renewable strategy to reduce energy use, cost, and boost occupant well-being, visual comfort, productivity, and health.
Some of the key strategies that are followed by designers and architects to daylight into design efficiently are: optimizing building orientations for maximum sun exposure, using correct indoor and outdoor flooring materials to achieve the right reflectance, placement of windows, light shafts, skylights, and atriums.
Here are the rules of thumb for daylighting in a design[1].
In this experiment, the idea was to develop a data-driven process to design spaces with sufficient daylight.
The Data Approach
The daylighting data was generated using Ladybug and Honeybee, the environmental plug-ins for Grasshopper3D. Parameters provided for the Box were, height of the box, width of the box, length of the box, and Window to Wall Ratio (WWR) of the glazing provided just on the southern façade. The main feature used for daylight analysis was “Daylight Autonomy”.
Daylight autonomy is the percentage of time that daylight levels are above a specified target illuminance within a physical space or building. The calculation is based on annual data and the predetermined lighting levels. — Wikipedia
The city chosen for this analysis was the Indian city of Mumbai, which is a moderately hot tropical location.
The daylight simulations were computationally intensive, so input parameters were limited to certain values. In our experiment, a total of 117 design options and simulation daylight simulations were created. The dataset contained geometric and glazing information for each design option.
Data Analysis
Here is the correlation of all the features with Daylight Autonomy (DA):
1. Volume and Average DA = .22
2. Floor Area and Average DA = -0.04
3. Width and Average DA = .23
4. Length and Average DA = -0.48 (Strong negative correlation)
5. Height and Average DA = 0.53 (Strong positive correlation)
6. WWR and Average DA = 0.34
When two sets of data are strongly linked together we say they have a High Correlation. Correlation is Positive when the values increase together, and. Correlation is Negative when one value decreases as the other increases.
Results of Machine Learning
The daylighting data was split into two sub-datasets: train data, and test data. The simple linear regression model was trained on the train data with the length, width, height, WWR, floor area as the model features, and Daylight Autonomy, as the target variable. The Daylight Autonomy was predicted for the test data using the trained ML model.
The machine learning process is mentioned below.
Process → Split → Train → Predict → Evaluate
Just by using simple linear regression to predict the Daylight Autonomy, the model was able to achieve a correlation of 96.7% , which is actually great. more features can be added to this like the location, glazing on other façades, and much more. This can be a more data-driven approach to sustainable designing and can even cut the time taken for heavy daylight simulations.
“Designers should be able to use sunlighting strategies without creating a conspicuous ‘sunlighting building’ look.” — William Lam
N O T E S
1 “Daylighting Design Principles” https://ocw.mit.edu/courses/architecture/4-401-environmental-technologies-in-buildings-fall-2018/lecture-slides-1/MIT4_401F18_lec8.pdf
2 “Daylight Rules-of-Thumb Experimentally Examined” https://www.asee.org/documents/sections/2017/Mid-West/32-Mansy-Daylight%20Rules-of-Thumb%20Experimentally%20Examined.pdf
3 “Daylighting Buildings” ; http://www.ewinlight.com/wp-content/uploads/2017/09/grtdaylightingpresn2009-11-12-100730124931-phpapp01.pdf
4 “Daylight design rules of thumb” https://www.irbnet.de/daten/iconda/CIB_DC23487.pdf
About the Author
Rutvik Deshpande is a machine learning and data science research intern at Digital Blue Foam and Architecture Student at @NITRR. He has interests in Data Science, Machine Learning, and Artificial Intelligence for the AEC industry. He is also a Computational Design, Urban Analytics enthusiast and a Kaggler too.
About Digital Blue Foam
At Digital Blue Foam we believe buildings and cities must reduce their dependence on fossil fuels in all aspects of the process design. This includes materials, construction processes, and operations. Presently, we are researching new ways to reduce energy use , and ultimately carbon emissions, and boost occupant comfort by leveraging tried and tested passive design principles in our design software.
