Sarith No Hyp

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 158

The Pennsylvania State University

The Graduate School


Department of Architectural Engineering

PARAMETRIC MODELING STRATEGIES FOR EFFICIENT

ANNUAL ANALYSIS OF DAYLIGHT IN BUILDINGS

A Dissertation in

Architectural Engineering

by

Sarith Subramaniam

© 2018 Sarith Subramaniam

Submitted in Partial Fulfillment


of the Requirements
for the Degree of

Doctor of Philosophy

May 2018
The dissertation of Sarith Subramaniam was reviewed and approved* by the following:

Richard G. Mistrick
Associate Professor of Architectural Engineering
Graduate Program Officer of Architectural Engineering
Dissertation Advisor
Chair of Committee

Kevin W. Houser
Professor of Architectural Engineering

Donghyun Rim
Assistant Professor of Architectural Engineering

Ute Poerschke
Professor of Architecture

Moses Ling
Associate Professor of Architectural Engineering
Special Signatory

*Signatures are on file in the Graduate School

ii
ABSTRACT

Daylighting simulations are an essential part of the modern building design process. Results
generated from such simulations influence various aspects of building design such as facade
orientation, design of electric lighting and lighting control systems, selection and sizing of glazing,
and choice of internal or external shading systems for glare control. Recent advances in the
development of visual scripting mediums for popular 3D modeling platforms have made it possible
to easily set up hundreds of such simulations for the design of a single building. Most
contemporary mainstream daylight simulation tools, however, are not conducive to such large-
scale studies. These tools are based on nearly two-decade old raytracing algorithms, which in
addition to being computationally inefficient, also rely on simplifying assumptions that compromise
the precision of the simulations.
This dissertation research investigates simulation workflows which can improve the computational
efficiency and precision of parametric daylighting simulations. These workflows leverage newly
introduced tools within the Radiance raytracing system. The principal research is documented in
five separate chapters that cover four hypothesis-driven numerical studies and an open source
software development initiative. The first two studies focus on novel approaches to employ
daylight coefficients as a means of calculating precise values of illuminance and luminance for
annual climate-based simulations. The remaining studies investigate the potential of employing
multi-phase simulations as a means of reducing the computational runtime for illuminance-based
simulations. The workflows employed for organizing and automating the simulations for the
numerical studies were scripted using a custom-written software. This software, written in the
Python programming language, was eventually assimilated within an open source building
simulation library that has now been publicly released.

iii
Table Of Contents

List of Figures ............................................................................................................................. vii


List of Tables............................................................................................................................... xv
Preface....................................................................................................................................... xvii
Acknowledgements ................................................................................................................... xviii
Chapter 1. Introduction .............................................................................................................. 1
1.1 Daylighting ...................................................................................................................... 1
1.2 Predicting daylight availability ......................................................................................... 1
1.3 Climate-based daylight modeling: Relevance, Recent Developments and Challenges . 2
1.4 Methodology ................................................................................................................... 5
1.5 Efficiency and Precision: A note on semantics ............................................................... 5
1.6 Dissertation layout .......................................................................................................... 5
Chapter 2. State Of Science ..................................................................................................... 7
2.1 Daylighting simulation software: a topical chronology .................................................... 7
2.2 Annual daylighting simulations ....................................................................................... 7
2.3 Multi-Phase Methods .................................................................................................... 11
2.3.1 Three-Phase Method ............................................................................................. 12
2.3.2 Four-Phase Method ............................................................................................... 12
2.3.3 DDS approach, Five-Phase method and Six-Phase Method ................................ 14
2.4 Daylighting metrics ....................................................................................................... 21
2.5 Relevance of the reviewed research to this dissertation .............................................. 22
Chapter 3. A more accurate approach for Calculating Illuminance with Daylight Coefficients 23
3.1 Background .................................................................................................................. 23
3.2 Aim and Hypotheses .................................................................................................... 24
3.3 Methodology ................................................................................................................. 24
3.3.1 Overview ............................................................................................................... 24
3.3.2 Experimental Design and Simulation Workflows ................................................... 26
3.3.3 Consolidating and visualizing illuminance values from simulations....................... 31
3.4 Results .......................................................................................................................... 31
3.4.1 Summarizing inter-method deviation with image-quality metrics........................... 34
3.4.2 Inter-method deviation for room with skylight ........................................................ 36
3.4.3 Effect of precise-sun positions on Annual Sunlight Exposure (ASE)..................... 37
3.4.4 Simulation runtimes ............................................................................................... 37

iv
3.5 Conclusion .................................................................................................................... 38
Chapter 4. Suitability of Daylight Coefficient-based Imaging for Glare evaluation .................. 39
4.1 Background .................................................................................................................. 39
4.1.1 Suitability of illuminance as a predictor of glare .................................................... 39
4.1.2 Glare evaluations through HDR imaging ............................................................... 39
4.2 Aim, Rationale and Hypotheses ................................................................................... 40
4.3 Methodology ................................................................................................................. 40
4.3.1 Overview ............................................................................................................... 40
4.4 Results .......................................................................................................................... 43
4.5 Conclusion .................................................................................................................... 46
Chapter 5. View-Matrix Subdivisions for Multi-Phase Daylighting Simulations ....................... 49
5.1 Background .................................................................................................................. 49
5.2 Aim and Hypotheses .................................................................................................... 51
5.3 Methodology ................................................................................................................. 52
5.3.1 Benchmark daylight coefficient simulations ........................................................... 54
5.3.2 Three-Phase simulations ....................................................................................... 55
5.4 Results .......................................................................................................................... 57
5.5 Guidelines for assigning view-matrix subdivisions ....................................................... 60
5.5.1 Models with no contextual geometry and no external shading elements .............. 60
5.5.2 Models with no contextual geometry that have external shading elements .......... 60
5.5.3 Models with contextual geometry and external shading ........................................ 61
5.5.4 Models with mullions ............................................................................................. 62
5.6 Conclusion .................................................................................................................... 62
Chapter 6. Applicability of the Four-Phase Method for iterative simulations with non-coplanar
shades 63
6.1 Background .................................................................................................................. 63
6.2 Aim and Hypotheses .................................................................................................... 63
6.3 Methodology ................................................................................................................. 63
6.3.1 Setup ..................................................................................................................... 63
6.3.2 Simulation parameters .......................................................................................... 65
6.4 Results .......................................................................................................................... 66
6.4.1 Comparison of simulation runtimes ....................................................................... 66
6.4.2 Annual illuminance results for different configurations of the shading system ...... 67
6.5 Conclusion .................................................................................................................... 69
Chapter 7. Software Development for Parametric Daylighting Simulations ............................ 70

v
7.1 Background .................................................................................................................. 70
7.2 Motivation and development philosophy for a new parametric daylighting software .... 72
7.3 Compatibility of Radiance-based methods with software development goals .............. 73
7.4 Creation of Honeybee[+]: A high level software development summary ...................... 74
7.5 Public release, complementary activities and response from the AEC community ...... 77
Chapter 8. Discussion and Conclusion ................................................................................... 79
8.1 Impact of discretizing skies to a higher degree of granularity ....................................... 79
8.2 Feasibility of multi-phase methods for complex models ............................................... 79
8.3 The trade-off between computational efficiency and complexity of the Façade matrix 80
8.4 Limitations .................................................................................................................... 80
8.5 Future work ................................................................................................................... 80
Appendix A. Weather data for simulations ............................................................................. 82
A. 1 Locations ...................................................................................................................... 82
A. 2 Dates and times for truth-simulations and benchmark studies ..................................... 84
Appendix B. Detecting inter-method deviation in illuminance-based simulations .................. 87
B.1 Background .................................................................................................................. 87
B.2 Employing image-quality metrics for comparing illuminance values............................. 92
B.3 Relating variation in PSNR to absolute values of illuminance ...................................... 99
Appendix C. Additional data visualizations for Chapter 3 ..................................................... 101
Appendix D. Additional data visualizations for Chapter 4 ..................................................... 104
Appendix E. Additional data visualizations for Chapter 5 ..................................................... 116
BIBLIOGRAPHY ....................................................................................................................... 128

vi
List of Figures

Figure 1. Parametric optimization of perforated shading devices using Octopus and Honeybee
inside Rhino-Grasshopper3D. Honeybee uses Daysim, a Radiance-based simulation engine, for
daylighting calculations (Kim and Lee 2017). ............................................................................... 3

Figure 2. Glazing optimization through Design Explorer for a parametric co-simulation of daylight
and building energy consumption (Thornton Tomasetti 2017). The thumbnail-images on the right
depict design options. The lines and sliders on the top indicate performance parameters and their
numerical values. .......................................................................................................................... 3

Figure 3. A decision-tree for selecting a multi-phase method. The 2 Phase method is the closest
to the original implementation of Daylight Coefficients as proposed by Tregenza. (Credit:
Mostapha Sadeghipour) ................................................................................................................ 4

Figure 4. A schematic representation of daylight coefficients. Credit: (Reinhart 2001) ................ 8

Figure 5. An example of a grid-points assigned to an indoor space for an illuminance-based


daylight simulation. The image on the right shows a data-visualization of point-in-time illuminance
values obtained through the simulation. The letters A, B, C, D and E in both images refer to the
same locations on the measurement grid. .................................................................................... 9

Figure 6. Circular patches proposed by Tregenza for discretizing the hemispherical sky structure.
Credit: (Tregenza 1987) .............................................................................................................. 10

Figure 7. The image (a) shows the Tregenza sky subdivison scheme with 145 sky patches. Image
(b) shows the continuous sky subdivision scheme proposed by Reinhart. Credit: (Bourgeois et al.
2008) ........................................................................................................................................... 10

Figure 8. Fish-eye projections of a continuous sky model (a) and corresponding discretized
versions. Images (b), (c) and (d) contain 145, 580 and 2305 sky patches respectively. As is
apparent in the images above, even with a high degree of discretization, the the size of the sun
is greatly overestimated in discrete sky models. Credit: (McNeil 2013b). ................................... 11

Figure 9. A schematic comparison between the Daylight Coefficient (Two-Phase Method) and the
Three-Phase Method. The matrices invovled in the simulation are depicted through lines and
letters. The Daylight Coefficient Method involves the calculation of the daylight coefficients matrix
and the sky-vector or sky-matrix (and can therefore be thought of as a Two-Phase Method). The
Three-Phase Method involves the calculation of the interior V Matrix, exterior Daylight Matrix and
the sky-vector or sky-matrix. The Transmission(T) Matrix is not calculated as a part of the
simulation and is generated either through LBNL Window 7.4, geBSDF or created from empirically
measured data. ........................................................................................................................... 13

Figure 10. Images generated through Three-Phase Method and other more conventional
methods. The images on top highlight the fact that the external view is obscured in the Three-
Phase Method. Discrete sky patches are visible in the image generated through the Daylight
Coefficient Method. The lack of distinct shadows in the bottom-right image indicates that the
shadows created by shading systems are obscured in the Three-Phase Method. Although the

vii
above observations are made in the context of image-based simulations, they are relevant to
illuminance-based simualtions as well. ....................................................................................... 13

Figure 11. A space that is shaded through adjustable grates that can be adjusted with two degrees
of freedom. .................................................................................................................................. 14

Figure 12. Schematic comparison between the Three-Phase Method and the Four-Phase Method.
.................................................................................................................................................... 14

Figure 13. A comparison of images rendered with conventional ray-tracing and a continuous sky
definition (top row) and images rendered with discretized skies (bottom row). The discretized sky
vector used for the images in the bottom row was generated from the continuous sky definition
employed for the images in the top row. The scattered beam for the images in the bottom row can
be attributed to the use of sky patches. ...................................................................................... 15

Figure 14. Sky and sun patches employed in the Two-Phase DDS, Five-Phase Method and Six-
Phase Method. The image on the left is a standard continuous sky that is used for conventional
ray-tracing simulations. ............................................................................................................... 16

Figure 15. Each of the above images is a fisheye rendering for the solar positions corresponding
to a particular value of MF. The sun-positions so calculated are shown in red. The actual position
of the sun for every hour of the year is shown in white. In the DDS approach, Five-Phase Method
and Six-Phase Method, the position of the sun at any given hour is approximated to one of the
sun-positions in red. As the images indicate, the probability of the actual (white) and assumed
(red) location of the sun coinciding are higher if more sun-positions are considered for the
simulation. ................................................................................................................................... 17

Figure 16. Images rendered through different types of BSDF representations for the same set of
venetian blinds. The image on the right, which features Tensor Tree BSDFs with proxy geometry,
represents the most accurate result as it incorporates both direct and diffuse part of luminous flux
transfer. Credit: (Ward et al. 2012) ............................................................................................. 18

Figure 17. Schematic diagram for the Five-Phase Method. As shown in the images above, this
method seeks to improve upon the results generated through the Three-Phase Method by
incorporating a more accurate calculation for the direct-sun component of the sky. Dark geometry
in two of the above figures indicates non-reflecting surfaces. .................................................... 18

Figure 18. Images rendered through the Five-Phase Method for a space with a daylight redirecting
system (in upper windows) and venetian blinds. The left-most image is the result from a Three-
Phase simualtion (VTDS). The next image is the result from a simulation that only considers the
direct-sun apsect of the Three-Phase Simulation. ...................................................................... 19

Figure 19. Schematic diagram for the Six-Phase Method. The F-aperture employed in the Four-
Phase Method is applicable for calculating the direct-sun component of the Four-Phase Method
as well. The workflow for calculating sun coefficients is identical to that used in the Five-Phase
Method. ....................................................................................................................................... 20

Figure 20. The above images show the images rendered through the different stages of a Six-
Phase Method simulation. The patterns shown on the wall in the images on the right are due to

viii
the grates located outside the window. The ray-tracing calculations for the grates are handled
through the Facade-Matrix. ......................................................................................................... 20

Figure 21. The dotted line represents the actual position of the sun .The crosses denote the 65
representative positions from which the actual sun position is approximated (Reinhart and
Walkenhorst 2001). ..................................................................................................................... 23

Figure 22. The space used for the simulations is a south-facing room with floor dimensions of
20’x20’ and a height of 10’ from ceiling to floor. The south orientation was chosen for simulations
because it is highly susceptible to direct insolation. .................................................................... 25

Figure 23. Location of calculation points considered for simulations. As the numbers in the figure
indicate, each calculation point is spaced 2’ in x and y directions from the other point. All the points
are 2.5’ from the floor level. ........................................................................................................ 25

Figure 24. Room daylit through a single skylight. The recess for the 6’ x’ 6 skylight can be seen
in the figure. ................................................................................................................................ 26

Figure 25. From left to right. Fish-eye representations of the sky models used for diffuse sky
simulation with multiple ambient bounces, direct sky simulation with single ambient bounce and
direct sun simulation. .................................................................................................................. 29

Figure 26. A comparison of fish eye renderings of continuous sky with discretized skies.
Continuous sky is used for conventional raytracing while discretized skies are used matrix-based
methods like DDS, MF1, MF2 etc. .............................................................................................. 29

Figure 27. The sun positions used for DDS method and the improved methods. As the image on
the right indicates, the number of suns is not affected by the multiplying factor for the sky
discretization. .............................................................................................................................. 30

Figure 28. Illuminance calculated through different methods for the TMY weather data of Denver,
CO at 21st June at 10:30 AM. The name of the simulation method is indicated on the top of each
plot. The numbers next to digits L, M and H indicate mean and standard deviation of 0-25th
percentile, 26th-75th percentile and 76th-100th percentile illuminance values respectively. .......... 31

Figure 29. Comparison of illuminance results for Denver on March 20th at 4:30PM ................... 32

Figure 30. Comparison of illuminance results for Fargo on June 21st at 3:30PM ....................... 32

Figure 31. Comparison of illuminance results for Phoenix on September 22nd at 5:30PM ......... 33

Figure 32. Comparison of illuminance results for Seattle on March 20th at 2:30PM ................... 33

Figure 33. Plot of PSNR for Birmingham and Denver. The 40 hours, counted from 0-39 in the x-
axis correspond to 10 hours/day for solstice and equinox dates. ............................................... 35

Figure 34. Plot of PSNR for Fargo and Phoenix. The 40 hours, counted from 0-39 in the x-axis
correspond to 10 hours/day for solstice and equinox dates. ....................................................... 35

Figure 35. Plot of PSNR for Pittsburgh and Seattle. The 40 hours, counted from 0-39 in the x-axis
correspond to 10 hours/day for solstice and equinox dates. ....................................................... 36

ix
Figure 36. Inter-method deviation for the room with skylight for Birmingham and Denver.
Remaining plots are in Appendix C. ............................................................................................ 36

Figure 37. Simulation runtime (for the set of 240 simulations) plotted as a function of sky patches
used in the simulation. The runtimes plotted above were from simulations that were performed
with a Linux operating system on an Intel i7 3770 3.4GHz Desktop computer with 8 virtual cores.
The use of higher resolution skies was accompanied by a corresponding increase in the value of
ambient divisions (-ad) used for the simulation. .......................................................................... 37

Figure 38. The two views considered for the study. .................................................................... 41

Figure 39. Images generated through three stages of the matrix-based process. ..................... 42

Figure 40. A comparison between falsecolor images generated ................................................ 43

Figure 41. Daylight Glare Probability (DGP) for east and south views for Birmingham. ............. 44

Figure 42. DGP for east and south views for Denver. ............................................................... 44

Figure 43. DPG for east and south views for Fargo. ................................................................... 45

Figure 44. Sum of solid angles of glare sources (in steradian) for the east view for Fargo. ....... 45

Figure 45. Glare sources superimposed on the HDR images for 10:30 AM for Fargo. .............. 46

Figure 46. Disk space required to store coefficient-based image files created by MF1 – MF5
simulations. The calculation is based on a file size of 1.4MB. .................................................... 47

Figure 47. Raytracing runtimes for creating images required for MF1 – MF6 simulations. All the
simulations were performed on 30 dedicated cores of an Intel Xeon E5-2680 (v2 @2.80GHz)
processor. ................................................................................................................................... 48

Figure 49. The images on the left and right show the front and back transmission properties for a
BSDF. ......................................................................................................................................... 50

Figure 50. A Three-Phase Method rendering. The patterns on the glazing are indicative of the
Klems-basis employed in the BSDF data-structure. ................................................................... 50

Figure 51. A simulation with multiple view-matrices. As shown by the image on the right, the
results from individual simulations are added arithmetically to derive the final result. ................ 51

Figure 52. Two view-matrix configurations. The wide angle view is shown below the perspective
view as it displays all the view matrix apertures. The red checkered arrangement too is shown for
the sake of clarity. Two apertures of the same color do not indicate that they apply the same view-
matrix. ......................................................................................................................................... 52

Figure 53. Dimensions of the room used for the study. The front glazing faces north. ............... 53

Figure 54. Positioning of the photosensors. 450 photosensors in a 30x15 grid were employed for
the study. The spacing between the photosensors is 2 feet by 2 feet and all photosensors are 2.5
feet above the floor. The red arrows indicate the orientation of the photosensor. ...................... 53

x
Figure 55. The model room without and with the 2 feet overhang. The overhang was used to study
the effect of external façade-based geometry on the results. ..................................................... 54

Figure 56. The model room without and with mullions. ............................................................... 55

Figure 57. The images show the room with and without contextual geometry. The shadow cast
on the room by the contextual geometry can be seen in the image on the right. ....................... 55

Figure 58. View-matrix apertures for the model with mullions. As explained previously, the red
checkered pattern is employed to clarify the number of apertures and does not imply similarity.
For example, the configuration referred to as GM contains a single view matrix while GM2H
contains 12.................................................................................................................................. 56

Figure 59. View-matrix apertures for a model without mullions. ................................................. 56

Figure 60. PSNR plots of different view-matrix configurations for the model with no mullions, no
overhang and no contextual geometry. The data are plotted for the locations of Birmingham and
Seattle. ........................................................................................................................................ 58

Figure 61. PSNR plots of different view-matrix configurations for the model with no mullions, no
contextual geometry with a 2’ feet overhang in the north-facing glazing. The data are plotted for
the locations of Birmingham and Seattle. ................................................................................... 58

Figure 62. PSNR plots of different view-matrix configurations for the model with no mullions,
external contextual geometry and a 2’ feet overhang in the north-facing glazing. The data are
plotted for the locations of Birmingham and Seattle. .................................................................. 59

Figure 63. PSNR plots of different view-matrix configurations for the model with no mullions and
external contextual geometry. The data are plotted for the locations of Birmingham and Seattle.
.................................................................................................................................................... 59

Figure 64. Simulation runtimes for model with no external shades or contextual geometry. The
data-point NM-NG indicates the time taken for the benchmark Daylight Coefficient Method
simulation. The labels (in red) denote the quantity of view-matrices, and therefore also the daylight
matrices, in each simulation method. .......................................................................................... 60

Figure 65. Simulation runtimes for a model with no contextual geometry that has external shading.
The data-point NMO-NG indicates the time taken for the benchmark Daylight Coefficient Method
simulation. ................................................................................................................................... 61

Figure 66. Simulation runtimes for a model with contextual geometry and external shading. The
data-point NMO-WG indicates the time taken for the benchmark Daylight Coefficient Method
simulation. ................................................................................................................................... 61

Figure 67. The software model of the room considered for this study. The louvered overhang is
adjustable through two degrees of freedom. ............................................................................... 64

Figure 68. Dimensions of the room shown in Figure 4. The thickness of the walls is 9” and the
dimensions of the floor are 20’x20’. The dimensions of the glazing are 18’ x 5’. ........................ 64

xi
Figure 69. The images above show various settings for the adjustable shading system considered
for this study. The overhang shown above can be adjusted along two degrees of freedom: along
the local axes of individual fins and along the main axes of the entire overhang. The configurations
considered for the simulation include three settings for fins and two settings for the overhang. The
angles considered for the rotation of fins were 0°, 45° and 90° while that for the entire overhang
were 0° and 30°. From here on, the settings will be referred to by their designation as mentioned
below each image. ...................................................................................................................... 65

Figure 70. The above figure shows the individual runtimes for the Four-Phase Method and
Daylight Coefficient Method while simulating different settings of the overhangs shown in Figure
6. ................................................................................................................................................. 67

Figure 71. Cumulative runtimes for the Four-Phase Method and the Daylight Coefficient Method.
From a purely computational perspective the used Four-Phase Method required 26% less time
than the Daylight Coefficient Method. ......................................................................................... 67

Figure 72. The above figure shows the resultant illuminance grid for the F0S0 setting of the
overhang at 10:30AM on 1st January. The control algorithm for selecting a particular shade setting
for each hour was assigned on the basis of a grid-point located 5’ away from the glazing and 9’
away from the right wall. That grid point is highlighted through a translucent polygon in the above
figure. .......................................................................................................................................... 68

Figure 73. The above bar-plot shows the total number of times each setting of the overhang was
selected by the control algorithm. An hourly annual representation of the above data shown in
Figure 74. .................................................................................................................................... 68

Figure 74. The above grid-plot shows the hour-wise overhang settings for the entire year. The
control algorithm was designed to operate the shade between 8:30 AM and 5:30PM. In the
absence of natural light, the control algorithm reverted to the default F0S0 setting. .................. 69

Figure 75. A screen capture of Su2Rad. The images on the left show a typical GUI for entering
calculation parameters. ............................................................................................................... 71

Figure 76. A screen capture of Adeline. The CAD input needed for Radiance is created in separate
software. ..................................................................................................................................... 71

Figure 77. Agile model of software development. The development cycle is meant to incrementally
improve the software. In the context of this research, this implies improving the software based
on research, testing and changes in functionality of Radiance or modeling interfaces. Image
Credit: (Smartsheet 2018). .......................................................................................................... 73

Figure 78. Schematic diagram for the Daylight Coefficient Method, Three-Phase Method and
Four-Phase Method. The terms on the right detail calculation steps and Radiance-programs
employed for them. The programs are highlighted in red. Programs like rfluxmtx, gendaymtx and
oconv feature in all the simulations, often for accomplishing tasks that are common to the different
methods. ..................................................................................................................................... 74

Figure 79. Contribution graph for Honeybee[+] as of December 2017. Number of “commits” refers
to the individual instances when the code was updated. The green and red numbers next to

xii
commits refer to the lines of codes added and removed. The author’s contributions are credited
as “sariths” Source: (GitHub 2018) ............................................................................................. 75

Figure 80. Logical structure of Honeybee[+]. .............................................................................. 76

Figure 82. The six cities whose TMY data was utilized for simulations in Chapters 3 -5. ........... 82

Figure 83. Plot of Direct-Normal Radiation on Sun-Path diagrams for the six cities. A value of
zero, indicated by blue in the figure, implies an overcast sky for that particular hour. ................ 83

Figure 84. Solar altitudes for 8:30 to 17:30hrs for equinox and solstice dates plotted for the six
locations. ..................................................................................................................................... 84

Figure 85. A typical plot that is representative of the hourly plots used in Chapter 3 – Chapter 5.
The 40 hours, numbered from 0 to 39, in the x-axis correspond to the same serial # for hours in
Table 22 - Table 25. .................................................................................................................... 86

Figure 86. Plot of measured and simulated illuminance values from (Mardaljevic 1995). The study
involved six photosensors. .......................................................................................................... 88

Figure 87. Location of the six photosensors used for measuring and simulating illuminance values
in (Wang et al. 2016). .................................................................................................................. 88

Figure 88. Visualization of annual metrics on a grid where each square represents an illuminance
sensor (DiLaura et al. 2011). The plotted metrics, from left to right, are Daylight Factor, Daylight
Autonomy and Continuous Daylight Autonomy. The glazing is at the bottom. ........................... 89

Figure 89. A comparison of point-in-time illuminance values calculated through seven simulation
methods. The benchmark results are on upper-left. The “DDS” method (second from the left on
top row) indicates an overestimation of illuminance. .................................................................. 89

Figure 90. A comparison of illuminance values for the same model as considered in Figure 89. In
this instance the ‘DDS” method underestimates the illuminance. ............................................... 90

Figure 91. Comparison between point-in-time illuminance values generated through a benchmark
raytrace simulation and the DDS method simulation. Simulation type is mentioned on top of each
image. L, M and H correspond to Low (0-25), Medium (26-75) and High (76-100) percentile
ranges. The numbers within square-brackets indicate the mean illuminance within a particular
range, followed by the standard deviation. Note that the above plots are a subset of the plots
shown in Figure 91. ..................................................................................................................... 91

Figure 92. The mean of three percentile bins described in Figure 91 are plotted in the above plots
for a period of 40 hours. The underestimation by the DDS method in the 76 th-100th percentile bin
is evident in the image on the top. .............................................................................................. 91

Figure 93. A comparison between two calculation methods using binned percentile ranges. The
numerical comparison, as shown by the means above the plots, misleadingly indicates similarity
between the two results. ............................................................................................................. 92

xiii
Figure 94. PSNR values for inter-method illuminance comparison for Birmingham, AL. The
numbers below the plot, which are highlighted through arrows, pertain to hours listed in Table 22
- Table 25. ................................................................................................................................... 93

Figure 96. Illuminance data pertaining to hour #4 from Figure 94. ............................................. 94

Figure 97. Illuminance data pertaining to hour #7 from Figure 94. ............................................. 95

Figure 98. Illuminance data pertaining to hour #14 from Figure 94. ........................................... 95

Figure 99. Illuminance data pertaining to hour #21 from Figure 94. ........................................... 96

Figure 100. Illuminance data pertaining to hour #25 from Figure 94. ......................................... 96

Figure 101. Illuminance data pertaining to hour #27 from Figure 94. ......................................... 97

Figure 102. Illuminance data pertaining to hour #30 from Figure 94. ......................................... 97

Figure 103. Illuminance data pertaining to hour #36 from Figure 94. ......................................... 98

Figure 104. Illuminance data pertaining to hour #39 from Figure 94. ......................................... 98

Figure 105. PSNR values for simulations with different view-factor configurations. The dataset
contains 40 observations corresponding to 40 hours. The reference dataset is a Daylight
Coefficient simulation. The terms “GD”, “G” etc. relate to different view factor configurations. The
individual datasets for hour 16, which is highlighted through a red dotted line, are shown in Figure
106. ............................................................................................................................................. 99

Figure 106. Illuminance plots corresponding to observation 16 for three of the seven
configurations plotted in Figure 105. The corresponding PSNR values are mentioned on the left
side of each plot. The letters in parenthesis below the PSNR values identify the configuration.
The difference between the Reference dataset and “G” is easily discernable. The differences
between the Reference dataset and other two datasets are highlighted through dashed
rectangles. It should be noted that PSNR is expressed in a logarithmic scale to the base of 10.
So, the numerical differences of individual grid-point illuminances between “G” and “G2VH” are
to the order of more than 100 lux. ............................................................................................. 100

xiv
List of Tables

Table 1. The optical properties of different surfaces considered in the simulations. All opaque
surfaces were assumed to be Lambertian. ................................................................................. 25

Table 2. Independent variables for determining the quantity of raytrace simulations. ................ 27

Table 3. Radiance parameters for convergence tests on raytracing simulations. ...................... 27

Table 4. Independent variables for determining the quantity of DDS simulations. ..................... 27

Table 5. Simulation parameters for convergence optimization of Daylight Coefficient simulations.


.................................................................................................................................................... 28

Table 6. Independent variables for determining the quantity of the simulations through the
improved methods. ..................................................................................................................... 29

Table 7. Quantity of sky patches corresponding to multiplying factors. The formula for calculating
patches is (144 x MF x MF + 1). ................................................................................................. 30

Table 8. Annual Sunlight Exposure values for the six locations considered in the study. This
simulation was performed on a full 8760 hour dataset of TMY3 weather files. ........................... 37

Table 9. Number of conventional raytracing simulations. ........................................................... 41

Table 10. Number of daylight coefficient simulations. ................................................................. 42

Table 11. Root Mean Square Error (RMSE) values for DGP values for MF1 to MF5 approaches
for the south view. The reference data are the DGP values obtained through conventional
raytracing. ................................................................................................................................... 46

Table 12. Root Mean Square Error (RMSE) values for DGP values for MF1 to MF5 approaches
for the east view. The reference data are the DGP values obtained through conventional
raytracing. ................................................................................................................................... 47

13. Optical properties of the surfaces of the room. ..................................................................... 53

Table 14. Independent variables for the daylight coefficient simulations. ................................... 54

Table 15. Independent variables for the Three-Phase simulations. ............................................ 57

Table 16. The optical properties of the diifferent surfaces in the model. .................................... 64

Table 17. Designation of shade setting based on clockwise rotation angles of the fins and the
overhang. Rotation of fins corresponds to each fin on the overhang being rotated about its local
x-axis. Rotation of the overhang corresponds to the rotation of the entire overhang structure about
its local x-axis.............................................................................................................................. 65

Table 18. The table below lists the number of ambient bounces (reflections) considered in each
iteration for different matrices of the Four-Phase Method. As indicated by the left-most column,

xv
the Daylight Coefficient Method involves a one-step simulation involving 12 reflections. The
number of ambient bounces is directly proportional to the amount of time required for computation.
.................................................................................................................................................... 66

Table 19. The number of times each matrix was calculated during the course of the study. ...... 66

Table 20. The table below lists the mean simulation runtime (in minutes) for different aspects of
the six Four-Phase Method and Daylight Coefficient Method simulations. The values in
parenthesis indicates the standard deviation. ............................................................................. 66

Table 21. The latitude and longitude of the six cities used for the simulations. Positive signs are
assigned for latitudes north of the equator and longitudes to the west of prime meridian. ......... 83

Table 22. Direct-Normal and Diffuse-Horizontal radiation values for March 20th. ....................... 85

Table 23. Direct-Normal and Diffuse-Horizontal radiation values for June 21st. ......................... 85

Table 24. Direct-Normal and Diffuse-Horizontal radiation values for September 22nd. ............... 85

Table 25. Direct-Normal and Diffuse-Horizontal radiation values for December 21st.................. 86

xvi
Preface

The review presented in Chapter 2 is an extended version of the literature review detailed in a
tutorial written by the author. This tutorial was sponsored by the Lawrence Berkeley National
Laboratory and the author has been granted permission to adapt that work for this dissertation.
The tutorial is cited as:
Subramaniam S. 2017. Daylighting Simulations with Radiance using Matrix-based
Methods. LBNL.
Versions of chapter 3 and chapter 6 were presented at the Illuminating Engineering Society (IES)
annual conferences in 2017 and 2016 respectively. The authors have the right to include the
presented paper in their dissertation/thesis or publish in LEUKOS, the Journal of the IES. The
citations for these publications are:
Subramaniam S, Mistrick R. 2017 A More Accurate Approach for Calculating Illuminance
with Daylight Coefficients. 2017 Annual IES Conference; Portland, Oregon, USA:
Subramaniam S, Mistrick R. 2016. Incorporating adjustable external shading systems in
annual daylighting simulations: A prototypical study. 2016 Annual IES Conference;
Orlando, Florida, USA
The work presented in other chapters is not published previously.

xvii
Acknowledgements

I would like to thank my advisor Dr. Richard Mistrick for his continuous guidance and support
during my graduate studies over the last six and a half years. The entirety of my academic
endeavors during this time have benefited from his vast experience and insights in the field of
Lighting. I am grateful to him for nudging me towards the specific area of parametric daylighting
simulations and affording me the latitude to work on open source software development in parallel
with my dissertation research.
I am also thankful to the members of the dissertation committee: Dr. Kevin Houser, Dr. Ute
Poerschke, Dr. Donghuyn Rim and Professor Moses Ling for their guidance and constructive
feedback.
Mostapha Sadeghipour Roudsari, the co-founder of Ladybug Tools and my collaborator on
several open-source projects, provided invaluable advice on matters relating to software
development.
Part of this dissertation work was completed during my tenures at the Lawrence Berkeley National
Laboratory (LBNL) in California and HKS Inc. in Texas. I am thankful to Eleanor Lee for inviting
me to LBNL and assigning me a project that was complimentary to my dissertation research.
Through LBNL, I had the opportunity to work with Greg Ward, the creator of the Radiance
rendering system. This research, as well as numerous other such academic endeavors across
the world, would not have been possible without the support from Greg and his tireless
stewardship of Radiance for the past three decades. I would like to thank Heath May, Paul Ferrer
and Tim Logan at HKS Inc. for helping me apply elements of this doctoral research in real-world
projects.
Majority of the simulations performed for this research employed the cloud-based High-
Performance Computing (HPC) system at Penn State. I thank the staff at the Institute of
CyberScience for providing with the access to HPC and also for their guidance on migrating my
software scripts and data to different platforms over the last three years.
I am indebted to Deb Sam, Corey Wilkinson, Richelle Weiger, Chris Decoteau, Nancy Sabol and
Holly Seidel from the Department of Architectural Engineering (AE) for their excellent assistance
on all the administrative and infrastructural matters relating to my graduate studies.
I am grateful to Ling, Nablus, Qi, Tommy, Tony and Reza, my colleagues from the
Lighting/Electrical option in AE, for their advice, humor and companionship.
Finally, I would like to thank my immediate and extended family for their enthusiastic support. I
credit my parents, sister and brother-in-law for keeping me grounded and cheerful during the last
five years. The successful completion of my doctoral education is a testament to their
unconditional support and love.

xviii
Dedication

To Achan and Amma.

xix
Chapter 1. Introduction

1.1 Daylighting
The tenth edition of the Illuminating Engineering Society (IES) Lighting Handbook defines
daylighting as the practice that facilitates the delivery and distribution of light from the sun and
sky to the interior of a building. Ideally, the duration and quantum of this lighting should meet the
visual and biological needs of the occupants of the building (DiLaura et al. 2011)1.
Daylighting has played a crucial part in the selection and design of living spaces since the
beginning of time. Prehistoric cave dwellings were selected and designed to let daylight in through
the mouth of caves as well as punched holes (Kittler et al. 2011). Except for a few decades in the
early 1900s, when electric lighting was being discussed as a more viable or even superior
alternative, daylighting has been universally acknowledged as the preferred source for
illuminating indoor spaces (Phillips 2004). This acknowledgement has its basis in a
comprehensive body of evidence and experience, both anecdotal and scientific, about the health-
related and economic benefits of employing daylight as the primary source of illumination (Aries
et al. 2015; Begemann et al. 1997; Bodart and De Herde 2002; Boubekri et al. 2014; Boyce et al.
2003; Ihm et al. 2009; Mardaljevic et al. 2009; Tregenza and Wilson 2013)2.

1.2 Predicting daylight availability


The annual and diurnal availability of daylight in a building is a function of its geographical location
and climate. The position of the sun with respect to a location also depends on the daily rotation
of the earth around its own axis and the annual revolution of the earth around the sun. The visibility
and luminous intensity of the sun, and the overall luminous distribution of the sky, are affected by
the prevailing climatic conditions. This basic understanding of daylight has been used to design
buildings for several millennia. Numerical efforts to compute the position of the sun in the sky
have been documented as early as 2900 BC. In comparison, the use of climate-based data in
daylighting analysis is a newer practice that is less than a century old. The evolution of daylighting
calculations, from primitive sundials to modern computational methods, is briefly chronicled in
(Kittler et al. 2011).
The practice of predicting daylight availability in the design of buildings, referred to as daylighting
design from hereon, has been heavily gravitating towards computer-based methods since the
early 2000s. This trend can be attributed to the availability of inexpensive personal computers
and advances in the research and development of software tools for daylighting design (Brembilla
2016; Galasiu and Reinhart 2008; Kota and Haberl 2009; Reinhart 2004; Reinhart and Fitz 2006).
Among the multitude of techniques that are available for daylighting simulations, the methods
based on daylight coefficients are by far the most widely employed in academia and industry. The
concept of daylight coefficients was introduced by Tregenza and Waters (1983). This daylight

1
In this dissertation, the term “lighting” implies the process of delivering light required for occupants from natural or artificial means
such as Sun and electric luminaires respectively. A detailed and evocative description about the origins and semantics of lighting can
be found in David DiLaura’s treatise on the subject (DiLaura 2006).
2
Majority of the works cited herein are either reviews or research-based books that summarize several decades worth of studies on
the benefits of Daylighting.

1
calculation approach has been implemented in several popular daylighting software (Brembilla
2016)3.
Daylight coefficients mathematically relate illuminance inside a room to the luminance of the sky
and the geometry and optical properties of the room and its surrounding surfaces. They are
typically used to calculate time-series values of daylight-derived illuminance through workplane-
located sensors inside a building. The luminous contribution from the sun and sky are calculated
based on geographical data and weather-tapes that contain hourly annual values of direct-normal
and diffuse-horizontal radiation. This technique is commonly referred to as Climate-based
Daylight Modeling (CBDM). The most commonly used CBDM tools employ the Radiance
raytracing system, or its derivative, Daysim, as the simulation engine.
This dissertation focuses on the improvement and comparative evaluation of CBDM simulation
techniques that are implemented through Radiance. The following section explains the
significance of this niche within the larger context of parametric whole-building simulations. It also
outlines some of the recent developments and challenges identified in this field.

1.3 Climate-based daylight modeling: Relevance, Recent Developments and


Challenges
The application of weather-data based simulations as a means of guiding building design has
been advocated by building science researchers for many years. The primary impetus for this
advocacy is based on the superior post-occupancy performance of buildings designed through
extensive modelling and analysis (Hensen 2011). In general, the Architecture Engineering and
Construction (AEC) industry has been gradually moving towards the adoption of multi-disciplinary
parametric studies to inform building design (Foliente 2000). This trend has been aided by the
introduction of parametric interfaces such as Grasshopper3D and DynamoBIM. Plugins such as
Honeybee, Vi-Suite and ArchSim, which are built on top of parametric interfaces, link modeling
platforms like Rhino, Blender and Revit to building simulation tools (Dogan 2016; Østergård et al.
2016; Roudsari and Pak 2014; Southall 2016; Southall and Biljecki 2017). The simulations
typically featured in whole building simulations include energy modeling, computational fluid
dynamics (CFD), daylighting and acoustics (Oxman 2017).
The availability of parametric interfaces linking modeling interfaces to simulation engines like
Radiance and EnergyPlus has encouraged the use multi-factor optimization studies involving
thousands of simulations. The simulation runtime for such studies typically spans a few hours to
several weeks. The optimization aspect of the study is usually performed using a plugin like
Octopus for Rhino-Grasshopper3D (Holzer 2016). Figure 1 shows the schematic diagram for one
such study. The large volume of results generated through such studies are typically filtered using
statistical analysis or tools such as Design Explorer that incorporate filtering algorithms and data-
visualization into a single graphical interface. Figure 2 shows a screen capture of Design Explorer.
A recent review-paper by Touloupaki and Theodosiou (2017) has listed several such studies while
also identifying the bottlenecks in their wide-scale adoption. One of the key challenges listed by

3
The justification for what qualifies a daylighting software to be deemed as popular is debatable as there have been limited number
of exhaustive surveys on daylighting software. However, the ubiquity of the Daylight Coefficient Method, as well as the use of the
Radiance ray-tracing engine in its implementation, can be gleaned through the body of research published on this topic. This fact has
also been acknowledged by Ashdown et al. (2017), who are the principal developers of LICASO. LICASO is a recently introduced
daylighting software that, as claimed by the authors, for the first time, employs Radiosity-based techniques for annual daylighting
simulations.

2
Figure 1. Parametric optimization of perforated shading devices using Octopus and Honeybee inside Rhino-
Grasshopper3D. Honeybee uses Daysim, a Radiance-based simulation engine, for daylighting calculations
(Kim and Lee 2017).

Figure 2. Glazing optimization through Design Explorer for a parametric co-simulation of daylight and building
energy consumption (Thornton Tomasetti 2017). The thumbnail-images on the right depict design options. The
lines and sliders on the top indicate performance parameters and their numerical values.

them is the “rapid generation of design alternatives utilizing computer capacity in full”.
Unfortunately, the key to generating a large number of design alternatives is not as straightforward
as maximum utilization of a computer. A commercially available desktop computer, even when
utilized at maximum efficiency, is incapable of performing a large number of simulations in short
time. This is especially the case for simulations involving finite element analysis like daylighting
and CFD where the quantum of computational effort and required disk space is large. Additionally,
while the computational demands of building simulations have gone up, the corresponding

3
performance improvement in conventional desktop computing processors has almost plateaued
(Waldrop 2016).
Within the realm of daylighting simulations, which are mostly implemented through raytracing
algorithms, a few researchers have focused on improving simulation-runtimes through the use of
Graphics Processor Units (GPUs) instead of conventional Central Processing Units (CPUs)
(Jones and Reinhart 2017; Woop et al. 2005; Zuo et al. 2014). The only available tool for
performing GPU-based simulations has been subjected to few empirical validations and is
presently tethered to a GPU manufactured by a single commercial manufacturer.
On the other hand, research conducted primarily by scientists at the Lawrence Berkeley National
Laboratory (LBNL) has focused on evolving the conventional Daylight Coefficient Method into
more efficient, albeit complex, methods. The researchers have proposed advanced methods that
“split” the ray-tracing aspect of the Daylight Coefficient Method into multiple “phases”. The
rationale behind procedurally splitting a simulation in this manner is to isolate and recalculate only
that part of simulation which is being changed parametrically. Figure 3 shows a decision tree for
selecting a particular multi-phase method based on the aspect of the building that is being studied
parametrically. Based on the figure, a parametric simulation involving only dynamic skies will
require a 2-Phase simulation. Similarly, a parametric simulation involving changes in external
shading should be conducted using the 4-Phase method. These methods are explained in further
detail in Chapter 2. It follows that the multi-phase methods are intrinsically suited for parametric
daylighting simulations. Although they were introduced nearly eight years ago, there has been
very limited adoption of these methods by daylighting designers and building simulation
specialists (Brembilla 2016).

Figure 3. A decision-tree for selecting a multi-phase method. The 2 Phase method is the closest to the original
implementation of Daylight Coefficients as proposed by Tregenza. (Credit: Mostapha Sadeghipour)

4
1.4 Methodology
The activities undertaken during the course of this dissertation research can be classified into two
categories:
1. Hypothesis driven research: on daylighting simulation methods that consisted of four
numerical studies. Individually, the intention of each study was to either improve an
existing simulation method, or to investigate the feasibility of using an advanced, and
complex to implement, simulation method. The improvements can relate to precision or
computational effort.
2. Software development: that implemented the simulation methods in an open source library
while leveraging the knowledge gained from the four numerical studies.
The work on the four numerical studies and the open source library started in August 2015 and
concluded in late 2017. The numerical studies were conducted solely by the author under the
supervision of his advisor. The open source library was a collaborative undertaking between the
author and another software developer.

1.5 Efficiency and Precision: A note on semantics


The word “efficiency”, when used in the context of simulation methods discussed in this
dissertation, implies the efficiency in computational effort and runtime. For example, an arbitrary
method “A” is construed to be more efficient than an arbitrary method “B”, if the method “A” takes
less amount of computer runtime to generate same or similar set of results. Furthermore, in such
comparisons, the same computing hardware is used for all the methods being compared.
The word “precision” implies the precision with which a test method can numerically replicate the
results generated by a benchmark method. It should be noted that within the context of this
research, efficiency and precision were found to be interlinked. For example, efficiency was often
obtained at the expense of precision and vice-versa.
The rest of the chapters in this dissertation provide details about these studies and the open
source library. The following section provides a concise summary of the various chapters.

1.6 Dissertation layout


Chapter 2 reviews the scientific research and software development work that provides the
foundation for the numerical studies described in Chapter 3 - 6. The study detailed in Chapter 3
proposes, and demonstrates, improvements to the conventional Daylight Coefficient Method.
Chapter 4 investigates the feasibility of using luminance-based Daylight Coefficient simulations
for generating images that can subsequently be used for computing glare.
Chapter 5 and Chapter 6 focus on the phase-based methods introduced in section 1.3. The
motivation for the studies described in these chapters is to compare the precision and efficiency
of the new, advanced methods against the conventional Daylight Coefficient Method. The broad
hypothesis is that the advanced methods will match the precision of the conventional method
while being more efficacious in terms of computational effort.
Chapter 7 describes the development of an open source software library that was developed to
implement the advanced concepts discussed in Chapter 3 - 6. It also reviews defunct and
currently-active daylighting simulation software and advocates the need for a new approach for
the development of such tools.

5
Chapter 8 summarizes the research and software development activities undertaken during this
dissertation research. It aims to provide a perspective on the relevance of this research to current
daylighting practice, highlight some of its key shortcomings and provide directions for future
research.

6
Chapter 2. State Of Science

The principal research conducted for this dissertation relates to daylighting simulation methods.
The foundational principles and software tools that form the basis of these simulation methods
originated in the early 1980s. An overview of the research pertaining to these principles and tools
is provided in sections 2.1 - 2.3. Section 2.4 reviews the common metrics and statistical methods
employed for summarizing results from annual daylighting simulations. Section 2.5 outlines the
relevance of the reviewed topics to the studies conducted for this research.

2.1 Daylighting simulation software: a topical chronology


One of the earliest scientifically acknowledged attempts towards bringing daylighting calculations
into the realm of software computing was made by DiLaura and Hauser (1978). In their paper
titled “On calculating the effects of daylighting in interior spaces”, they discussed the calculation
of Equivalent Sphere Illumination (ESI) and Visual Comfort Probability (VCP) using a computer
program, and presented a corresponding proof-of-concept. This research drew upon earlier work
by DiLaura and other researchers where computers were employed to perform calculations
involving luminaires as light sources (Blackwell and DiLaura 1973; Blackwell et al. 1973a;
Blackwell et al. 1973b; DiLaura 1975; 1976; DiLaura and Stannard 1978). The ideas proposed by
DiLaura and Hauser were furthered by Harvey Bryan and Michael Modest, who considered
complex geometries and proposed generalized solutions for different sky-types (Bryan 1980;
Bryan and Clear 1980; Modest 1982; 1983). The computational methods used in these works
relied on Radiosity-based calculations4. Radiosity calculations, as applied in the studies
mentioned herein, operate under the assumption that all the participating surfaces in a simulation
are of diffusing nature. Furthermore, at the time, the limitations of Radiosity-based methods for
simulations involving complex geometries had been long acknowledged5. As early as the 1950s,
researchers had begun experimenting with Monte Carlo techniques as a means to solve complex
radiative transfer problems (Howell 1969).
Modest (1982) and Tregenza (1983) discussed the implementation of Monte Carlo Methods for
daylighting simulations by applying the concepts pertaining to thermal radiative transfer within the
context of visible radiation. By the mid-1980s, Monte Carlo methods had also been implemented
in an architectural lighting software being developed at the Lawrence Berkeley National
Laboratory (LBNL). This software, which is named Radiance, has been under continuous
development and validation at LBNL (LBNL 1997; Wang et al. 2016; 2017; Ward 1994b; 2017;
Ward and Rubinstein 1988; Ward et al. 1987; Ward 1994c; Ward et al. 1988). As the following
section will elucidate, over the last three decades the development of Radiance has dovetailed
with the development of annual daylighting simulations.

2.2 Annual daylighting simulations


Daylighting simulations have been an integral part of building design for several years. A survey
of energy consultants, engineers, architects and designers, which was conducted more than a
decade ago, revealed that almost 80% of the respondents used daylighting simulations in their
design process (Reinhart and Fitz 2006). Presently, Climate-based Daylight Modeling (CBDM) is

4
As they are both forms of electromagnetic radiation, the optical and radiative-transfer principles applicable to infrared radiation and
visible radiation are same.
5
Recent developments in Radiosity-based methods have overcome such limitations (Ashdown 2016). These developments, however,
haven’t overly impacted advancements in daylighting software. Therefore, a detailed discussion on this topic is secondary to the scope
and objectives of this research.

7
one of the most popular approaches to daylighting simulations (Brembilla 2016). Such simulations
utilize typical meteorological weather data for an entire year to calculate hourly illuminance values
for daylit spaces. The hourly illuminance values are further condensed into metrics which are
used for performance evaluation of buildings. The use of annual daylighting simulations, and the
metrics generated through them, are mandated in several building guidelines and construction
codes (Hraška 2011; IES 2012; USGBC 2014).
The Daylight Coefficient (DC) method is the most commonly employed simulation method for
annual daylighting simulations in industry and academia. Daylight Coefficients were proposed by
Tregenza and Waters (1983) as a means of efficiently calculating illuminance under varying sky
luminosity conditions. The DC method leverages finite element analysis and matrix-based
operations to calculate time-series illuminance values in a space. The core principle behind this
method is that the daylight directly or indirectly incident on a surface inside a room can be
accounted for by considering two independent factors: luminance of the sky and the geometry
and optical properties of the surrounding surfaces. The illuminance at a measurement point in the
room from a small patch of sky can be calculated as:
∆𝑬 = 𝑫𝜽𝝋 𝑳𝜽𝝋 ∆𝑺𝜽𝝋 …………………….[1]

where 𝐿𝜃𝜑 is the luminance of the sky patch and 𝑆𝜃𝜑 is the angular size of the sky element at an
altitude of θ and azimuth of 𝝋. 𝐷𝜃𝜑 , the Daylight Coefficient, is a factor that depends on the
geometry of the room and the surrounding buildings as well as the reflectances and
transmittances of the surfaces that constitute that geometry. Figure 4 shows a schematic
representation of daylight coefficients.

Figure 4. A schematic representation of daylight coefficients. Credit: (Reinhart 2001)

The equation for the DC method in [1] can be written in terms of matrices as:
E = CdcS ……………………. [2]

where Cdc represents the Daylight Coefficient Matrix and S represents the sky matrix. For example,
assuming a room with 100 illuminance grid-points and a discretized sky with 145 patches, the

8
matrix dimensions of Cdc and S will be [100 x 145] and [145 x 1] respectively. The dimensions of
the resultant illuminance matrix E will be [100 x1], where each data point will represent the
illuminance at a particular grid-point. Figure 5 shows a typical example of grid-points assigned for
an illuminance-based simulation. Expanding this example to an annual simulation, and
considering a sky vector for each of the 8760 (365 x 24) hours in a year, the dimensions of Cdc ,
S and E will be [100 x 145], [145 x 8760] and [100 x 8760] respectively. In this case, the values
in E represent time series illuminance at the 100 grid points for 8760 hours a year.

Figure 5. An example of a grid-points assigned to an indoor space for an illuminance-based daylight


simulation. The image on the right shows a data-visualization of point-in-time illuminance values obtained
through the simulation. The letters A, B, C, D and E in both images refer to the same locations on the
measurement grid.

The luminance values for the skies used in the DC Method are usually derived from Typical
Meteorological Year (TMY) weather data for different geographical locations (National Climatic
Center 1981; Reinhart 2006). The weather data, commonly available in the form of EnergyPlus
Weather (EPW) files, contains hourly values for Direct Normal and Diffuse Horizontal Irradiation.
These irradiance values, along with corresponding month-day-hour data and geographic
coordinates can be utilized to create continuous luminance- or Radiance-based sky definitions
through the Perez Sky Model (Perez et al. 1990; Perez et al. 1993a; 1993b). The continuous sky
models are then discretized to matrix format by approximating the celestial hemisphere to a series
of luminous patches.
The discretized sky model employed for calculating Daylight Coefficients has undergone several
modifications over the years. One of the earliest sky models considered for Daylight Coefficient
calculations was proposed by Tregenza (1987). This model divides the celestial hemisphere into
circular patches and was devised at the time by considering the aperture size of the probe of the
scanning luminance meters used for measuring sky luminance. The angular size of the probe
considered for the patches in Figure 6 is 11.13˚.

9
Figure 6. Circular patches proposed by Tregenza for discretizing the hemispherical sky structure. Credit:
(Tregenza 1987)

Tregenza also discussed variants of this model by using different probe sizes. For example, a
probe size of 10.15˚ will lead to a model with 145 sky patches. One major drawback of models
with circular patches is that a substantial part of the sky is unaccounted for, as the entirety the
celestial hemisphere is not covered by the patches. Mardaljevic (1999) and Reinhart and
Walkenhorst (2001) proposed an improved sky model consisting of ellipsoid sky patches that
cover the entire celestial hemisphere. Figure 7 shows a Tregenza Sky Model consisting of 145
patches as well as a corresponding continuous division sky model. The center of each of the
ellipsoid patches corresponds to the center of the circular sky patches in Tregenza’s model. The
Reinhart sky model, which implies the sky discretization scheme proposed by Reinhart (2001),
has since been further discretized by equally subdividing the original 145 patches.

Figure 7. The image (a) shows the Tregenza sky subdivison scheme with 145 sky patches. Image (b) shows
the continuous sky subdivision scheme proposed by Reinhart. Credit: (Bourgeois et al. 2008)

10
Discretizing the sky to a higher degree, at the expense of simulation runtime and disk memory, is
assumed to improve precision of the simulation. This assumed improvement in precision can
mostly be attributed to a better approximation of the size and luminance of the sun brought about
by the use of smaller patches. As shown in Figure 8, in discretized sky models, the actual position
of the sun in the sky at a given time is approximated to 3-4 sky patches. As is also evident from
Figure 8, this approximation in position is accompanied by an overestimation of the size of the
sun with respect to the sky, especially in the case where a sky with 145 patches is considered.
Measures for correcting the errors arising from this issue are discussed within the context of multi-
phase methods in 2.3.3.

Figure 8. Fish-eye projections of a continuous sky model (a) and corresponding discretized versions. Images
(b), (c) and (d) contain 145, 580 and 2305 sky patches respectively. As is apparent in the images above, even
with a high degree of discretization, the the size of the sun is greatly overestimated in discrete sky models.
Credit: (McNeil 2013b).

Virtually all the popular software implementations of the DC method use Radiance, or tools
derived through Radiance, as their calculation engine (Brembilla 2016; Galasiu and Reinhart
2008; Reinhart and Fitz 2006). The most popular Radiance-based implementation of Daylight
Coefficients is Daysim (2018). Several other software such as DIVA4Rhino (Jakubiec and
Reinhart 2011), SPOT (Rogers 2006) and Ladybug-Honeybee (Roudsari and Pak 2014) employ
Daysim as the ray-trace engine for simulations. Implementations of the DC Method using native
Radiance programs can be found in Groundhog (Molina 2015), OpenStudio (Guglielmetti et al.
2011) and Honeybee[+] (Roudsari and Subramaniam 2016). A Radiosity-based annual
daylighting software was recently introduced by Lighting Analysts Inc. (2017). This software,
which is presently benchmarked against Daysim, has not been subjected to any real-world
empirical studies yet (Ashdown 2016; Ashdown et al. 2017).
Often, the primary objective of daylighting simulations is to parametrically and iteratively evaluate
only a certain aspect of the scene. This is especially the case if multiple daylighting simulations
are performed to evaluate the performance of various types of glazing or shading systems while
keeping every other parameter in the scene constant. In such instances, the DC method, which
involves tracing rays from inside the room to the sky in a single step, becomes prohibitively
expensive. The multi-phase methods discussed in the next section are more suited for such
simulations.

2.3 Multi-Phase Methods


The multi-phase methods, which are all implemented through Radiance-based tools, serve two
broad purposes:

11
Firstly, they allow for a more efficient parametric study of different aspects of a building by isolating
the ray-tracing aspect of the simulation to independent “phases”. For example, the Three-Phase
and Four-Phase methods are meant to make the evaluation of fenestrations and façade systems
(respectively) more efficient.
Secondly, advanced multi-phase methods improve the precision with which the direct-sun
component of the simulation is calculated. The Dynamic Daylighting Simulations (DDS) method,
Five-Phase Method and Six-Phase Method are meant to improve the precision of the calculation
of the direct-sun component. The imprecision in the calculation of the direct-sun component was
addressed by Mardaljevic (1999) and Reinhart and Herkel (2000) in the initial implementation of
the DC Method through Radiance. Bourgeois et al. (2008) proposed an improvement, referred to
as the DDS approach, over Reinhart and Herkel’s method. The Five-Phase Method and Six-
Phase Method provide direct-sun corrections for the Three-Phase Method and Four-Phase
Method respectively (Geisler-Moroder et al. 2017; McNeil 2013b; Wang et al. 2016; 2017).
The following sub-sections provide an overview of the multi-phase methods.

2.3.1 Three-Phase Method


The Three-Phase Method facilitates annual or point-in-time parametric daylighting simulations
with Complex Fenestration Systems (CFS). As depicted in Figure 9, this method builds on the
Daylight Coefficient Method by virtually splitting the flux-transfer path into multiple independent
phases, namely:
1. View (V): Flux transfer from illuminance grid-points or a view specification to glazing or
CFS.
2. Transmission (T): Flux transfer through the glazing or CFS.
3. Daylight (D): Flux transfer from the exterior of glazing or CFS to the sky.
The matrix equation [2] can be adapted to the Three-Phase Method as:
E = VTDS ……………………. [3]
The matrices for View (V) and Daylight (D) “phases” are generated by ray-tracing through
Radiance. The Transmission (T) matrix, which is a Bidirectional Scattering Distribution Function
(BSDF) definition stored in XML format, can be generated either through a Radiance program
called genBSDF or through the LBNL Window software (LBNL 2016; McNeil 2015; McNeil et al.
2013; Mitchell et al. 2008; Molina et al. 2015). The BSDF data structure employs a hemispherical
sampling basis to store information about the optical properties of the glazing or CFS. As indicated
by the ellipsoid patches on the glazing in Figure 10, the use of hemispherical sampling is
noticeable in images generated through the Three-Phase Method. The principles behind the
Three-Phase Method, and a review of the theoretical concepts invoked in it, are presented in
(Ward et al. 2011). A discussion focusing on the parametric capability and data-reuse aspect of
the Three-Phase Method can be found in (Saxena et al. 2010). (Jonsson et al. 2009) and (McNeil
and Lee 2013) contain discussions specifically relevant to BSDFs and their application to the
Three-Phase Method.

2.3.2 Four-Phase Method


In the Three-Phase Method, the luminous flux transfer between the glazing apertures to the sky
is characterized by the Daylight (D) Matrix. Any element of the room geometry that is beyond the
extents of, and also non-coplanar to, the glazing aperture is considered to be a part of the external

12
Figure 9. A schematic comparison between the Daylight Coefficient (Two-Phase Method) and the Three-Phase
Method. The matrices invovled in the simulation are depicted through lines and letters. The Daylight Coefficient
Method involves the calculation of the daylight coefficients matrix and the sky-vector or sky-matrix (and can
therefore be thought of as a Two-Phase Method). The Three-Phase Method involves the calculation of the
interior V Matrix, exterior Daylight Matrix and the sky-vector or sky-matrix. The Transmission(T) Matrix is not
calculated as a part of the simulation and is generated either through LBNL Window 7.4, geBSDF or created
from empirically measured data.

Figure 10. Images generated through Three-Phase Method and other more conventional methods. The images
on top highlight the fact that the external view is obscured in the Three-Phase Method. Discrete sky patches
are visible in the image generated through the Daylight Coefficient Method. The lack of distinct shadows in the
bottom-right image indicates that the shadows created by shading systems are obscured in the Three-Phase
Method. Although the above observations are made in the context of image-based simulations, they are
relevant to illuminance-based simualtions as well.

environment. For example, the grates shown in Figure 11, which serve to shade the interior of the
room, will be considered as a part of the external environment. A parametric study involving the
surface-properties or geometry of the grates will require a recalculation of the D matrix. It follows
that the Three-Phase Method does not have any specific provisions for parametric evaluation of
shading systems that are external and non-coplanar to glazing apertures. The Four-Phase

13
Method, which employs an additional matrix (Facade-Matrix) to account for flux-transfer in the
façade, has been introduced for such cases.
The matrix equation for the Three-Phase Method can be rewritten for the Four-Phase Method as:
E = VTFDS ……………………. [4]
In spaces like the one depicted in Figure 11, this approach is useful in iteratively studying multiple
types of non-coplanar shading systems. Figure 12 provides a schematic comparison between the
Three-Phase Method and the Four-Phase Method.

Figure 11. A space that is shaded through adjustable grates that can be adjusted with two degrees of freedom.

Figure 12. Schematic comparison between the Three-Phase Method and the Four-Phase Method.

After a one-time calculation of all the flux-transfer matrices, subsequent simulations for studying
different non-coplanar shading systems only require the recalculation of the Facade-Matrix. The
Four-Phase method was introduced by Greg Ward at the 14th International Radiance Workshop
(Ward 2015). The empirical validation of the Facade-Matrix approach, on which the Four-Phase
Method is based, is discussed in (Wang et al. 2016; 2017).

2.3.3 DDS approach, Five-Phase method and Six-Phase Method


All the multi-phase simulation methods discussed in the previous sections approximate the
position and shape of the sun to three or four patches in the sky. As discussed previously in 2.2,

14
Figure 13. A comparison of images rendered with conventional ray-tracing and a continuous sky definition
(top row) and images rendered with discretized skies (bottom row). The discretized sky vector used for the
images in the bottom row was generated from the continuous sky definition employed for the images in the
top row. The scattered beam for the images in the bottom row can be attributed to the use of sky patches.

this approximation is detrimental to the precision of results obtained through these methods.
Figure 13 illustrates one such scenario. Mardaljevic (1999; 2000) proposed a solution to this issue
through an approach where the illuminance inside a space was calculated as a sum of four
components: direct and indirect illuminance from the sun, and direct and indirect illuminance from
the sky. Mardaljevic’s method involved sky matrices with 145 patches and sun-coefficients with
5010 patches. The higher number of patches for sun-coefficient calculations were intended to
minimize the error caused due to the angular separation between the actual position of the sun
and that of the assigned patch.
Subsequently, Reinhart and Walkenhorst (2001) proposed a slightly different approach which
employed around 65 patches for direct sun-coefficients. The authors investigated several
strategies to mitigate the errors that might arise from the use of less number of sun-patches.
These strategies involved selecting sun patches based on the Nearest Neighbor, Interpolation
and Shadow Testing. This approach of calculating Daylight Coefficients was implemented in
Daysim (Daysim 2018; Reinhart and Herkel 2000; Reinhart and Walkenhorst 2001). As mentioned
previously, a subsequent proposal by Bourgeois et al. (2008) for the DDS Method, suggested
improvements by using 2305 sun positions instead of the original 65 positions.
The equation for the Daylight Coefficient Method with direct-sun correction can be expressed as:
E = Cdc S - Cdcd Sd + Csun Ssun …………………….[5]
Cdcd and Sd denote direct-sky coefficients and direct-sky matrix respectively. Csun and Ssun denote
direct-sun coefficients and the sun matrix respectively. The relative scale and luminance of the
sky and sun patches in a typical simulation are shown in Figure 14.

15
Figure 14. Sky and sun patches employed in the Two-Phase DDS, Five-Phase Method and Six-Phase Method.
The image on the left is a standard continuous sky that is used for conventional ray-tracing simulations.

In a manner analogous to DDS and conventional Daylight Coefficient calculations, Five-Phase


and Six-Phase methods improve the direct-sun calculations for the Three-Phase and Four-Phase
methods respectively. These methods, which were developed at LBNL, advocate the use of a
sky-model consisting of 5165 sun-patches for improving the precision of the direct-sun
calculations (Geisler-Moroder et al. 2017; McNeil 2013a; 2013b; 2014; Wang et al. 2016; 2017).
The value of 5165 is equal to (144 x 6 x 6+1) and is based on a MF:6 Reinhart subdivision scheme.
The rationale for choosing a higher number of sun-patches is explained through Figure 15.
The Three-Phase Method, described in Equation [3], is extended to the Five-Phase Method as:
E = VTDS − Vd TDd Sd + Cds Ssun ……………………. [6]
where the term Vd TDd S denotes a separate three-phase simulation utilizing the same scene as
the original three-phase calculation denoted by VTDS. Vd TDd S, which stands for the direct-sun
aspect of a Three-Phase Simulation, differs from the VTDS on account of its emphasis on isolating
the direct-sun component of the simulation. The direct-sun component is isolated by performing
a flux-transfer calculation with no ambient bounces and considering a sky-matrix with only the
direct solar component. The term Cds Ssun denotes a more accurate simulation with direct solar
contribution. Cds denotes a coefficient matrix for direct sun that was calculated by incorporating,
wherever possible, a high- resolution Tensor-Tree BSDF along with the physical geometry of the
shading device that was used to create the BSDF. Incorporating the physical geometry of the
shading system in such a way overcomes the issue of obscured shadows shown in Figure 10.
The advantages of using a high-resolution BSDF along with proxy geometry are explained
through Figure 16. Figure 17 shows the schematic diagram for the Five-Phase Method. As the
diagram indicates, the Five-Phase Method involves three independent simulations as well. Figure
18 provides an example of the phase-wise results generated through the Five-Phase Method.
The resultant image features the sharp shadows cast by the shading device, which are otherwise
obscured in the case of a Three-Phase simulation.

16
Figure 15. Each of the above images is a fisheye rendering for the solar positions corresponding to a particular
value of MF. The sun-positions so calculated are shown in red. The actual position of the sun for every hour
of the year is shown in white. In the DDS approach, Five-Phase Method and Six-Phase Method, the position of
the sun at any given hour is approximated to one of the sun-positions in red. As the images indicate, the
probability of the actual (white) and assumed (red) location of the sun coinciding are higher if more sun-
positions are considered for the simulation.

17
Figure 16. Images rendered through different types of BSDF representations for the same set of venetian
blinds. The image on the right, which features Tensor Tree BSDFs with proxy geometry, represents the most
accurate result as it incorporates both direct and diffuse part of luminous flux transfer. Credit: (Ward et al.
2012)

Figure 17. Schematic diagram for the Five-Phase Method. As shown in the images above, this method seeks
to improve upon the results generated through the Three-Phase Method by incorporating a more accurate
calculation for the direct-sun component of the sky. Dark geometry in two of the above figures indicates non-
reflecting surfaces.

18
Figure 18. Images rendered through the Five-Phase Method for a space with a daylight redirecting system (in
upper windows) and venetian blinds. The left-most image is the result from a Three-Phase simualtion (VTDS).
The next image is the result from a simulation that only considers the direct-sun apsect of the Three-Phase
Simulation.

A recently concluded validation study found that further improvements to precision for image-
based Five-Phase simulations can be made by splitting the direct-sun aspect of the simulation
into two different matrices. These matrices address the interior of the space and overall scene
separately (Geisler-Moroder et al. 2017). The Five-Phase equation [5], can be modified for image-
based simulations as:
E = VTDS − Vd TDd Sd + (CR−ds + CF−ds )Ssun ……………………. [7]
The term 𝐶𝑅−𝑑𝑠 refers to the direct sun coefficient matrix for the interior of the space and the term
𝐶𝐹−𝑑𝑠 refers to the direct sun coefficient matrix for the entire scene.
The Six-Phase Method, depicted schematically in Figure 19, is similar to the Five-Phase Method
in its intent and execution. The sole difference between the Five-Phase Method and the Six-Phase
Method relates to the inclusion of the Facade-Matrix. So, the Five-Phase equation can be
extended for the Six-Phase Method as:
E = VTFDS − Vd TFd Dd Sd + Cds Ssun …...………………. [8]
The modifications suggested for the Five-Phase image-based simulations are applicable for the
Six-Phase method as well. So, equation [6] can be rewritten for the Six-Phase Method as
E = VTFDS − Vd TFd Dd Sd + (CR−ds + CF−ds )Ssun ……………………. [9]
Figure 20 provides an example of the phase-wise results generated through the Six-Phase
Method.
The annual simulation methods discussed in Sections 2.2 and 2.3 typically generate a large
volume of illuminance data. An annual hourly simulation for the space depicted in Figure 5 will
generate 876,000 illuminance values (100 values per hour for 365 x 24 hours). So, for datasets
of such magnitudes to be of practical relevance, it is imperative that there exist means to
summarize such volume of data.
The following section discusses metrics that have been devised for annual daylighting
simulations.

19
Figure 19. Schematic diagram for the Six-Phase Method. The F-aperture employed in the Four-Phase Method
is applicable for calculating the direct-sun component of the Four-Phase Method as well. The workflow for
calculating sun coefficients is identical to that used in the Five-Phase Method.

Figure 20. The above images show the images rendered through the different stages of a Six-Phase Method
simulation. The patterns shown on the wall in the images on the right are due to the grates located outside the
window. The ray-tracing calculations for the grates are handled through the Facade-Matrix.

20
2.4 Daylighting metrics
The use of quantitative metrics in Daylighting predates the development of computer-based
simulation methods by several decades. The preliminary development of modern daylight-based
photometry can arguably be attributed to Leonhard Weber (Kittler et al. 2011; Weber 1885). His
invention of the milk-glass photometer and solid angle meter, along with the work on sky
luminance patterns by Kähler (1908) and Schramm (1901), provided the initial impetus for
practical measurement of daylight in buildings. By the late 1930s, a confluence of photometry-
based research in Europe and the United States, as well as standardization efforts by
Commission Internationale de l'Éclairage (CIE), had led to the wide-scale adoption of lighting and
daylighting instrumentation based on the luminous efficiency function (Boynton 1996; Gibson and
Tyndall 1923; Viikari et al. 2005).
For much of the past century, Daylight Factor (DF), which is the ratio of internal illuminance to
external horizontal illuminance under an overcast sky, has been the principal metric for daylighting
simulations and field evaluations (Kittler et al. 2011; Nabil and Mardaljevic 2005; 2006; Reinhart
et al. 2006; Tregenza and Mardaljevic 2018; Tregenza 1980). At least since the early 1950s, a
standardized overcast sky defined by the CIE has been used for calculating DF (Moon and
Spencer 1942; Reinhart et al. 2006). As DF is calculated solely under overcast sky conditions, it
is not useful for summarizing the results from CBDM simulations that feature a variety of sky
conditions based on typical meteorological data.
As discussed in earlier in Section 2.2, the use of CBDM simulations has been continually gaining
traction since the beginning of the 21st century. This, in turn, has fostered the introduction of
several new daylighting metrics that are meant to summarize data generated through such
simulations. The more prominent among these metrics are Useful Daylight Illuminance (UDI),
Daylight Autonomy (DA), Spatial Daylight Autonomy (sDA) and Annual Sunlight Exposure (ASE)
(DiLaura et al. 2011; IES 2012; Nabil and Mardaljevic 2005; 2006; Reinhart et al. 2006). Some of
the more recently introduced metrics include Sunlight Beam Index (SBI) and Total Annual
Illuminance (TAI) (Brembilla et al. 2015b; Mardaljevic and Roy 2016). sDA and ASE have been
recommended by the US Green Building Council (USGBC) as the metrics for evaluating daylight
sufficiency and potential for glare in buildings (USGBC 2014).
All the metrics discussed hitherto in this chapter are meant for summarizing illuminance data
generated through point-in-time or annual simulations. Although metrics like ASE and UDI have
been purported to predict glare potential, the limitations of illuminance-based simulations,
especially those where measurements are made on the work-plane, for predicting visual
discomfort have long been acknowledged (Cantin and Dubois 2011; Goodman 2009; Mardaljevic
et al. 2012). Metrics for quantifying visual discomfort have traditionally been derived using
luminance, vertical luminance, solid-angle of luminous sources in foreground and background
and certain constants based on the human visual system (Bellia et al. 2008; Boyce 2014; Iwata
et al. 1991). Developments in luminance-based data acquisition through High Dynamic Range
Imaging (HDRI) in recent years has led to the introduction of visual discomfort metrics that can
be calculated from such images. HDRI-based images can be captured using commercially
available cameras or through the Radiance-based simulations (Inanici 2010; Reinhard et al.
2010).
Some of the commonly derived visual comfort metrics that are derived from High Dynamic Range
(HDR) images include Daylight Glare Probability (DGP), Discomfort Glare Index (DGI), Unified
Glare Rating (UGR), Visual Comfort Probability (VCP) and Unified Glare Probability (UGP). These

21
metrics can be calculated by analyzing HDR images using Evalglare, a Radiance-based program
(Wienold 2010; Wienold and Christoffersen 2005; Wienold et al. 2004).
Generating images for glare evaluation is a computationally intensive process and it is not feasible
to generate individual images for every hour of the year in a manner similar to illuminance
simulations. Attempts at annual hourly glare studies have typically involved using image-based
simulations in concert with daylight coefficient-generated illuminance values for calculation of
visual comfort metrics (Wienold 2009).

2.5 Relevance of the reviewed research to this dissertation


The studies conducted for this dissertation leverage the principles, findings and techniques
discussed in the previous sections to investigate ways in which the precision and efficiency of
parametric daylighting simulations can be improved. As discussed in Chapter 1, the parameter
for a parametric daylighting simulation can be anything from location, time, geometry or even type
of fenestration. An improvement in efficiency can be construed to be an improvement in
computational effort or disk utilization or the man-hours required by a designer in performing a
simulation. Considering the scope for a wide interpretation of “precision and efficiency” it is
imperative that the scope of the studies, and the relevance of the research reviewed in this chapter
to those studies, be defined with a high degree of specificity.
Chapter 3 describes a study that investigates improvements to the precision of illuminance values
generated through daylight coefficient-based simulations if the sun-patches used in the
calculation of the direct-component were based on the precise location and size of the sun. So,
this study essentially is an attempt to improve the precision of daylight coefficient-based
simulations beyond what is possible through Daysim or DDS-based approaches described in this
chapter.
Chapter 4 presents a study that examines the feasibility of using images generated through
daylight coefficient-based methods for glare evaluations. As discussed in the last section, Wienold
(2009) proposed generating images through a hybrid method involving conventional raytracing
and illuminance-based daylight coefficients. The method employed in Chapter 4, relies on a
hitherto untested method involving luminance-based daylight coefficients.
Chapter 5 describes a study that compares the computational efficiency of Three-Phase Method
simulations with respect to the Daylight Coefficient Method. It also evaluates the improvements
in precision, at the cost of computational effort, obtained through the use of view-matrix
subdivisions. Chapter 6 describes a study that examines the computational efficiency of the Four-
Phase Method, which involves the recently introduced Façade-Matrix approach, with respect to
the Daylight Coefficient Method.
It follows that the studies in Chapter 5 and Chapter 6 are numerical critiques of multi-phase
methods within the context of parametric daylighting simulations. Chapter 4 explores a new way
of performing parametric daylighting simulations with the intention of evaluating it as a faster
approach for glare calculations. The study described in the next chapter examines potential
improvements to the Daylight Coefficient Method for calculating illuminance, which is the most
commonly employed method of performing parametric daylighting simulations.

22
Chapter 3. A more accurate approach for Calculating Illuminance with Daylight
Coefficients

3.1 Background
Section 2.2 in Chapter 2 provides an extensive review of the evolution of the Daylight Coefficient
Method as applied in mainstream CBDM software like Daysim. The implementation of daylight
coefficients in Daysim, which is the industry standard, consists of two major approximations.
Firstly, the position of the sun is always assumed to be in one of 65 representative positions in
the sky. The real-time position and luminance of the sun is calculated by averaging the nearest
four representative positions to the actual position of the sun (Reinhart and Walkenhorst 2001).
Figure 21 shows a typical scenario. A subsequent revision of the initial Daysim method led to the
proposal of a calculation consisting of 2305 suns where the sun positions were approximated to
be at the center of a patch in a sky subdivided according to Reinhart MF:4 subdivisions (Bourgeois
et al. 2008). A visual comparison between actual sun positions and sun positions generated
through the MF:4 subdivision scheme is provided in Figure 15.

Figure 21. The dotted line represents the actual position of the sun .The crosses denote the 65 representative
positions from which the actual sun position is approximated (Reinhart and Walkenhorst 2001).

The second approximation pertains to the use of a 145-patch sky for the diffused and direct-
diffused part of the simulation. A comparison between a 145-patch sky and continuous, non-
discretized sky is provided in Figure 8. It follows that a sky with a greater number of patches is a
more accurate approximation of the continuous sky. A continuous sky with sun is an
approximation of the physical, real-world sky. The developer of Daysim justified these
approximations at the time by relating them to computational efficiency (Reinhart 2001).
The implementation of daylight coefficients in Daysim relies on a modified version of a Radiance
program called “rtrace”. Rtrace is one of the oldest programs in the Radiance toolkit and is not

23
particularly optimized for calculating contribution coefficients. The modification for rtrace for
Daysim essentially involved repeated runs of rtrace to calculate sky-patch contributions and
contribution from representative sun positions.
A relatively newer tool called Radiance tool called “rcontrib” is optimized for the calculation of
contribution coefficients from known light sources like the sun and discretized sky patches
(Jacobs 2010; Ward 2012). The development of rcontrib, coupled with the availability of much
more powerful computing systems than those available in early 2000s, has made it possible to
perform daylight coefficient-based simulations that need not rely on the approximations applied
in Daysim or its revised version known as the Dynamic Daylighting Simulations (DDS). The study
documented in this chapter evaluates the potential improvements in precision obtained through
such simulations.

3.2 Aim and Hypotheses


The aim of this study was to comparatively evaluate the impact of employing precise sun positions
and highly discretized skies on the precision of daylight coefficient simulations. The working
hypotheses for this study are:
1. The precision of the daylight coefficient approach proposed in this study will be greater
than that of the DDS method, especially in instances with direct solar insolation.

2. A progressive improvement in precision will be achieved through the use of skies with
increasing orders of discretization.

3. The values of Annual Sunlight Exposure (ASE) values obtained through the improved
methods will deviate from those obtained through the DDS method.
The following section details the Radiance models, simulation types and overall methodology
employed for testing the above hypotheses.

3.3 Methodology
3.3.1 Overview
The hypotheses outlined in the previous section were tested by comparing the precision of the
simulation results generated through the improved methods with those generated through the
DDS method. Within the context of Radiance, the improved methods and the DDS method are
essentially finite element approximations of conventional raytracing. The benchmark for testing
precision, the so called truth simulations, were individual raytracing calculations.
Simulations were performed for six locations in the United States, namely, Seattle (WA), Fargo
(ND), Pittsburgh (PA), Denver (CO), Birmingham (AL) and Phoenix (AZ). Since conventional
raytracing involves a separate and original calculation for individual timesteps, it is impractical to
perform such a calculation for each of the 8760 hours of the year. So a subset of the TMY3
weather data set was considered for all the comparative studies. The timesteps included in this
subset of data was based on 10 hours per day for solstice and equinox hours, thus amounting to
a total of 40 hours. A detailed description of this weather data, and the justification for choosing
timesteps in this manner, is explained in Appendix A.
Initially, simulations were performed on a simple south-facing room which was lit through
sidelighting by a single aperture. The dimensions of this room are provided in Figure 22 and the
reflectance and transmittance properties assigned to its surfaces are provided in Table 1.

24
Illuminance levels were calculated for 100 photosensors at workplane level inside this room.
These photosensors were placed at a height of 2.5’ from the floor level and were arranged in a 2’
by 2’ configuration. Figure 23 shows the location of the photosensors within the room.

Figure 22. The space used for the simulations is a south-facing room with floor dimensions of 20’x20’ and a
height of 10’ from ceiling to floor. The south orientation was chosen for simulations because it is highly
susceptible to direct insolation.

Table 1. The optical properties of different surfaces considered in the simulations. All opaque surfaces were
assumed to be Lambertian.

Surface Property Value


Ceiling Reflectance 90%
Walls Reflectance 65%
Floor Reflectance 30%
Ground Reflectance 20%
Glazing Transmittance 65%

Figure 23. Location of calculation points considered for simulations. As the numbers in the figure indicate,
each calculation point is spaced 2’ in x and y directions from the other point. All the points are 2.5’ from the
floor level.

25
Subsequent to testing the hypotheses on the model described in Figure 22, the entire set of
simulations and analyses were repeated for an additional model shown in Figure 24. This model
has the same height, width, breadth and surface properties as that of the room shown in Figure
22 and is illuminated solely through a single skylight. The transmittance of the skylight material
was assigned as 0.5. The purpose of this additional study was to ascertain that the precision of
the different methods were not overly dependent on a specific glazing configuration.

Figure 24. Room daylit through a single skylight. The recess for the 6’ x’ 6 skylight can be seen in the figure.

3.3.2 Experimental Design and Simulation Workflows


As discussed earlier in this section, the daylighting simulations performed for this study can be
classified as following:
1. Conventional Raytracing: The benchmark for comparing the precision of the other type of
simulations.
2. DDS: Daylight coefficient simulations based on the DDS Method proposed by (Bourgeois
et al. 2008).
3. MF1, MF2 …MF6: The methods proposed through this research as improvements to the
DDS method.
The following subsections describe the experimental design, high-level workflows and
convergence-testing processes for each of the above simulation methods.
3.3.2.1 Conventional Raytracing
A total of 240 raytracing simulations were performed. The independent variables for determining
the number of simulations are described through Table 2. A high-level description of the
procedure followed for each of the simulations is described through the following steps. Radiance
programs, if used for a particular step, are listed in square brackets at the end of each step:

• Step1: Create octree for the scene comprised of the sky file and room model [oconv].
• Step2: Perform ray-tracing by specifying a points file through standard-input and then
derive illuminance based on the Photopic Efficiency Function [rtrace, rcalc]
• Step3: Repeat Step2 by increasing the precision settings for ambient calculations till
convergence is achieved. Convergence in this context was defined as the state where the

26
difference between the illuminance results from two progressively accurate simulations
was less than 2%. The ambient settings chosen for successive iterations are shown in
Table 3.
Of the 240 benchmark simulations, convergence for all but eight instances occurred within the
calculation parameters of (-ab 6, -ad 2048, -ar 3072 and -aa 0.05). For the eight outlying
instances, convergence occurred when the ambient precision was set to 0.01, while the other
parameter values were retained. The simulation parameters for direct calculation were set at
maximum levels as (-dc 1, -dt 0, -dj 1).
Table 2. Independent variables for determining the quantity of raytrace simulations.

Independent Variable Levels Quantity


Birmingham (AL)
Denver (CO)
Fargo (ND)
Location 6
Pittsburgh (PA)
Phoenix (AZ)
Seattle (WA)
Hourly Sky files
Weather Data (10 hours/day for 40
Solstice and Equinox)

Total number of simulations 6*40=240

Table 3. Radiance parameters for convergence tests on raytracing simulations.

Parameter Options
Ambient Bounces (-ab) 5,6,7,8
Ambient Divisions (-ad) 1024, 2048, 3072, 4096
Ambient Resolution (-ar) 1024, 2048, 3072, 4096
Ambient Accuracy (-aa) 0.5, 0.1, 0.05, 0.01

3.3.2.2 DDS simulations


As the DDS approach is based on Daylight Coefficients, and therefore involves matrix
calculations, a different set of Radiance programs was used for these simulations. The
independent variables for the study are shown in Table 4.
Table 4. Independent variables for determining the quantity of DDS simulations.

The steps involved in each simulation are summarized below:

27
• Step1: Create an octree of the room model [oconv].
• Step2: Create a daylight coefficient file with 145 sky patches by tracing rays to
hemispherical representations of the celestial sky and ground [rfluxmtx]
• Step3: Create a 145 patch sky-matrix from the weather data file containing radiation
values for solstice and equinox hours [epw2wea, gendaymtx].
• Step4: Derive illuminance by multiplying the daylight coefficients with the sky-matrix
[dctimestep, rmtxop].
• Step5: Repeat Step2 and Step4 with progressively higher options for Step2 till
convergence is achieved. Convergence here is defined in the same manner as it was for
Raytracing simulations. The parameters for successive iterations are shown in Table 5.
• Step6: Create an octree by making all surfaces except the glazing as non-reflective and
non-transmissive [xform, oconv].
• Step7: Create a direct-only daylight coefficient file by using a single ambient bounce and
the rest of the parameters as those derived through convergence testing in Step 5.
• Step8: Create a 145 patch direct sky-matrix (by using the –d option of gendaymtx) from
the weather data file containing radiation values for solstice and equinox hours [epw2wea,
gendaymtx].
• Step9: Derive illuminance by multiplying the direct daylight coefficients with the direct sky-
matrix [dctimestep, rmtxop].
• Step 10: Generate a sun coefficients file for 2305 suns by employing the same octree
used in Step 6 [rcalc, rcontrib]
• Step 11: Generate a sun-matrix for 2305 suns by using the –d and -5 option in gendaymtx
[gendaymtx].
• Step 12: Derive illuminance by multiplying the sun coefficients with the sun-matrix
[dctimestep, rmtxop].
• Step 13: Combine the results from Step 5, Step 9 and Step 12 through matrix addition and
subtractions. First, subtract results of Step 9 from results of Step 5 to remove the
inaccurate direct-sun component of the calculation. To the result so obtained, add the
results from Step 12 to incorporate a more accurate direct sun component [rmtxop].
Point-in-time fish-eye renderings of the sky-matrices and sun-matrix referred to in Step 3, Step 8
and Step 10 are shown in Figure 25.
Table 5. Simulation parameters for convergence optimization of Daylight Coefficient simulations.

Parameter Options
Ambient Divisions 10E4, 20E4, 30E4, 40E4, 50E4, 60E4, 70E4, 80E4
Limit Weight Reciprocal of value for Ambient Divisions
Ambient Bounces 5, 6, 7, 8

3.3.2.3 Improved Methods: MF1 to MF6


The intent of these methods, as stated earlier, is to improve the precision of the Daylight
Coefficient-based calculations by incorporating precise sun positions and highly discretized skies.
Besides these two considerations, and the changes affected due to them in the workflow, the
methodology for the improved methods differs only slightly from the DDS method. The key
changes to the workflow are affected by the choice of settings and input files used for the
simulations. Independent variables for these simulations are shown in Table 6. The number in the

28
terms MF1, MF2, MF6 etc. refers to the multiplying factor used to discretize the sky. The number
of sky-patches corresponding to different multiplying factors is shown in Table 7

Figure 25. From left to right. Fish-eye representations of the sky models used for diffuse sky simulation with
multiple ambient bounces, direct sky simulation with single ambient bounce and direct sun simulation.

Table 6. Independent variables for determining the quantity of the simulations through the improved methods.

Independent Variable Levels Quantity


Birmingham (AL)
Denver (CO)
Fargo (ND)
Location 6
Pittsburgh (PA)
Phoenix (AZ)
Seattle (WA)
Weather Data Sky Matrix File 1
MF1, MF2, MF3,
Sky discretization 6
MF4, MF5, MF6
Total number of simulations 6*1*6=36

Figure 26. A comparison of fish eye renderings of continuous sky with discretized skies. Continuous sky is
used for conventional raytracing while discretized skies are used matrix-based methods like DDS, MF1, MF2
etc.

29
Table 7. Quantity of sky patches corresponding to multiplying factors. The formula for calculating patches is
(144 x MF x MF + 1).

The steps involved in the simulations are explained in relation to the steps for the DDS method
as follows:

• Step 1: Same as DDS method.


• Step 2, Step 3: Same as DDS except for one critical change. The number of sky patches
for daylight coefficients and sky matrix will be determined as per the multiplying factor. So,
as listed in Table 7, for MF1, it will be 145 patches while for MF2 it will 577 patches and
so on.
• Step 4, Step 5, Step 6: Same as DDS Method.
• Step 7, Step 8: Same as DDS method except that sky patches will be changed as per the
multiplying factor.
• Step 9: Same as DDS Method.
• Step 10: Compile Radiance definitions of suns as per their precise location and generate
a corresponding sky-matrix. This functionality is not available in Radiance and therefore a
tool with this functionality had to be written specifically for this research. This tool was
written in the Python language and is now a part of the Honeybee[+] open source library
(Roudsari and Subramaniam 2016). Casey and Mistrick (2015), had described a similar
tool written in the C++ language.
• Step 11: Generate a sky-coefficients file as per the quantity and positions of suns derived
through step 10.
• Step 12, Step 13: Same as DDS method.
Figure 27 provides a comparison of sun models used for the DDS Method and the improved
methods. The sun positions used for the improved methods are specific to locations and timesteps
involved in the simulation.

Figure 27. The sun positions used for DDS method and the improved methods. As the image on the right
indicates, the number of suns is not affected by the multiplying factor for the sky discretization.

30
3.3.3 Consolidating and visualizing illuminance values from simulations
As indicated by the independent variables in Table 2, Table 4 and Table 6, the number of actual
simulations involved in the Raytracing method, DDS Method and improved methods were widely
different. However, the cumulative end result of each method was hourly values of illuminance for
240 instances (10 hours x 4 days x 6 locations). Inter-method analysis and data visualizations
were performed for the results from each of the 240 instances. Figure 28 shows one such
visualization of illuminance data. The results from all the simulations are summarized in the next
section.

Figure 28. Illuminance calculated through different methods for the TMY weather data of Denver, CO at 21st
June at 10:30 AM. The name of the simulation method is indicated on the top of each plot. The numbers next
to digits L, M and H indicate mean and standard deviation of 0-25th percentile, 26th-75th percentile and 76th-100th
percentile illuminance values respectively.

3.4 Results
Comparison of results for individual time-steps, as shown in Figure 28, indicate that for the
majority of the simulations there was numerical conformity between the results from conventional
ray-tracing and those from the matrix-based methods. Four such instances for Denver, Fargo,
Phoenix and Seattle are visualized through Figure 29, Figure 30, Figure 31 and Figure 32
respectively. It should be noted that, due to Photopic adaptation, a difference of a few 100 lux is
not discernable in a room where the average illuminance is in thousands of lux. For the dates and
times considered in the simulations, inter-method deviations were observed in limited cases.

31
Figure 29. Comparison of illuminance results for Denver on March 20th at 4:30PM

Figure 30. Comparison of illuminance results for Fargo on June 21st at 3:30PM

32
Figure 31. Comparison of illuminance results for Phoenix on September 22nd at 5:30PM

Figure 32. Comparison of illuminance results for Seattle on March 20th at 2:30PM

33
It might already be apparent to the reader that discussing individual results for 240 such
observations is impractical. Evaluating results from the different simulation types essentially
involves comparing timeseries values of illuminance from 100 photosensor points in a space. The
illuminance values can vary from 0 lux to as high as 20,000 lux. As discussed in Chapter 2,
traditional daylighting metrics are more suited for summarizing such results through conditional-
filtering and simple arithmetic only on an annual basis. Appendix B provides a detailed review of
such traditional metrics and their shortcomings with respect to comparing a large number of
photosensor measurements for point-in-time simulations.
Section B.2 of Appendix B discusses the applicability of image-quality or distortion based metrics
for comparing illuminance results. The following sub-section summarizes the results of the
individual simulations through such a metric.

3.4.1 Summarizing inter-method deviation with image-quality metrics


3.4.1.1 Selection and interpretation of the metric
Figure 33 shows a plot of the Peak Signal to Noise Ratio (PSNR) values for comparisons between
illuminance values calculated through raytracing and daylight coefficient based methods. PNSR
is a distortion metric that can be used to measure the similarity between two datasets. It is affected
by Mean Square Error (MSE), the maximum value of the dataset, and structural conformity of the
test data set to the reference data set. It is calculated on a logarithmic decibel scale and its
maximum value is infinity, which indicates a perfect match between the reference and test data
sets. Defining what constitutes an acceptable value of PSNR depends purely on the context in
which they are applied (Hore and Ziou 2010; Huynh-Thu and Ghanbari 2008).
Daylighting simulations performed in this research were based on stochastic Monte-Carlo
techniques. So a perfect convergence between results from two simulations was practically
impossible. Based on observation of over 240 datasets, a PSNR value of around 50 dB was found
to be indicative of acceptable conformity between reference and test datasets. A value in the
vicinity of 20dB, or lower, was found to be indicative of significant deviation in the test dataset. A
series of illustrative examples are shown in Appendix B. Simulations for hours with no daylight
resulted in illuminance with zero values irrespective of the simulation method employed. Naturally,
this resulted in perfect conformity and a PSNR of infinity. For the purposes of visualizing the entire
dataset, the value of infinity was replaced with 200dB in the plots.
3.4.1.2 Inter-method comparisons
Figure 33 shows the PSNR plot for Birmingham and Denver. The values for Birmingham indicate
that, except for four instances, the performance of the DDS method was on par with the improved
methods. More instances of distortion were observed for Denver.
As the plots indicate, there were no consistent patterns of performance improvement or
deterioration among the improved methods, referred to as MF1 – MF6 in the legend, as a function
of sky discretization. The trends observed in the PSNR plots for Fargo and Phoenix (Figure 34)
and Pittsburgh and Seattle (Figure 35) are consistent with those observed for Birmingham and
Denver. The positive spikes observed towards the end of the x-axes in all the plots are due to
sunset at or before 17:30 hours.

34
Figure 33. Plot of PSNR for Birmingham and Denver. The 40 hours, counted from 0-39 in the x-axis correspond
to 10 hours/day for solstice and equinox dates.

Figure 34. Plot of PSNR for Fargo and Phoenix. The 40 hours, counted from 0-39 in the x-axis correspond to
10 hours/day for solstice and equinox dates.

35
Figure 35. Plot of PSNR for Pittsburgh and Seattle. The 40 hours, counted from 0-39 in the x-axis correspond
to 10 hours/day for solstice and equinox dates.

3.4.2 Inter-method deviation for room with skylight


The results and analyses discussed thus far are based on simulations conducted on a model
shown in Figure 22. Subsequently, the entire set of simulations and analyses were repeated on
a second model shown in Figure 24. As discussed previously, the intention of doing so was to
ascertain that the performance of the simulation methods based on sky discretization and precise
sun location were not a function of the glazing configuration. Figure 36 shows the corresponding
PSNR plots for Birmingham and Denver. The trends in PSNR values observed for the simulations
and analysis for the second model were consistent with those observed for the first model. Sharp
deviations from the benchmark result were observed for the DDS method while no consistent
improvements or deteriorations in performance were noticed as a function of the sky
discretization.

Figure 36. Inter-method deviation for the room with skylight for Birmingham and Denver. Remaining plots are
in Appendix C.

36
3.4.3 Effect of precise-sun positions on Annual Sunlight Exposure (ASE)
One of the logical drawbacks of using imprecise or approximated positions and radiation values
of sun is its effect on the precision of any calculation that seeks to quantify direct solar insolation.
ASE is one such metric. ASE, as defined by LM 83-12, is the “percent of sensors in the analysis
area, using a maximum of 2’ spacing between points, that are found to be exposed to more than
1000 lux of direct sunlight for more than 250 hours per year” (IES 2012).
A comparison of direct-only simulations based on the sun position and sky-matrices for the DDS
method and the improved method is shown in Table 8. It should be noted that in the case of the
improved methods proposed in this study, the sun positions employed in the simulation are not
dependent on the sky discretization. So, all the improved methods will yield the same value for
ASE. ASE calculated through the DDS method is based on approximated sun positions. So, the
values in Table 8 are essentially a measure of the error in the estimation of ASE through the DDS
method.
Table 8. Annual Sunlight Exposure values for the six locations considered in the study. This simulation was
performed on a full 8760 hour dataset of TMY3 weather files.

Location DDS Method Improved (MF1) Model


Birmingham, AL 0.32 0.32
Denver, CO 0.47 0.49
Fargo, ND 0.46 0.47
Phoenix, AZ 0.38 0.4
Pittsburgh, PA 0.28 0.29
Seattle, WA 0.28 0.28

3.4.4 Simulation runtimes


Figure 37 shows the mean simulation runtime for the 240 simulations that were run for the main
model. Considering the lack of any definitive advantages observed for employing a higher level
of sky discretization, at least on the basis the model considered, a MF1 simulation will be the most
optimal approach.

Figure 37. Simulation runtime (for the set of 240 simulations) plotted as a function of sky patches used in the
simulation. The runtimes plotted above were from simulations that were performed with a Linux operating
system on an Intel i7 3770 3.4GHz Desktop computer with 8 virtual cores. The use of higher resolution skies
was accompanied by a corresponding increase in the value of ambient divisions (-ad) used for the simulation.

37
3.5 Conclusion
The aim of this study was to investigate means by which the efficiency of daylight coefficient
simulations could be improved. Based on the assumptions made in the development of the current
industry standard method for daylight coefficients, the two potential sources of error were
identified as low resolution skies and approximated sun positions and radiation values. At the
onset of this study, the influence of either of these factors on the magnitude of error was unknown.
Based on the simulations and analyses performed in this study, approximations relating to sun
position and radiation were found to have a much larger impact on the precision of the simulations.
The absence of any definitive impact of sky discretization negates the initial hypothesis about the
advantages of employing highly discretized skies.
The deviations between the PNSR values measured for the DDS method and the improved
method ranged from 0dB to 30dB, with the improved methods consistently outperforming the DDS
method.
Based on the analysis and conclusions drawn from this study, a daylight coefficient simulation
with 145 sky patches and precise sun positions can be recommended as the optimal solution for
illuminance based simulations.

38
Chapter 4. Suitability of Daylight Coefficient-based Imaging for Glare evaluation

4.1 Background
4.1.1 Suitability of illuminance as a predictor of glare
The previous chapter proposed improvements to existing methods of calculating illuminance.
Illuminance-based simulations are the most commonly employed type of daylighting analysis for
buildings. The metrics derived from hourly annual illuminance simulations are the most commonly
used daylighting metrics. Some of the illuminance-derived metrics such as ASE and UDI are even
regarded as credible indicators of the potential for glare (Mardaljevic et al. 2012; Reinhart 2015).
Consequently, such metrics are employed for estimating glare in mainstream daylighting design
practice and have been used in academic studies as well.
However, lighting researchers have long acknowledged the shortcomings of any illuminance-only
evaluation of glare. Illuminance is a measure of luminous flux incident on a surface. Although it is
useful for quantifying the amount of visible light falling on a surface, it is not always appropriate
for quantifying the visual perception of that surface by an observer. Based on its optical properties,
the surface might appear as bright, dull or even shiny to the observer. Illuminance is also not
suited for quantifying the perception of discomfort or disability glare experienced by an observer
when there is a direct or reflected source of light in the field of view. The shortcomings of
illuminance for glare evaluation are especially compounded in the case of daylighting because
illuminance measurements are typically performed on the workplane only. Workplane illuminance
does not often correlate to the glare experienced at eye level (Mardaljevic et al. 2012).
Most of the mathematical models developed specifically for glare evaluation require several other
factors than illuminance. Moreover, illuminance when featured in a glare calculation model, is
usually measured in the vertical plane rather than the horizontal work plane. Some of the basic
factors common to most glare models are light source luminance, size and location of the light
sources and the adaptation state of the observer (Clear 2013). It follows that numerical
quantification of glare, whether real-world or through simulation, is a lot more complex and
involved than measuring illuminance.

4.1.2 Glare evaluations through HDR imaging


Traditionally, glare measurements have entailed sampling luminance and vertical illuminance
values for different parts of a potential visual field through point-by-point measurements. These
measurements are time-consuming and also prone to errors in the case of transient and
unpredictable sources like daylight. Since the early 2000s researchers have been evaluating High
Dynamic Range (HDR) photography as a means of acquiring data that could be used for glare
evaluation. HDR images captured through high-quality commercial cameras can be used for
capturing luminances that span a wide range of values (Inanici 2006; Inanici and Galvin 2004;
Reinhard et al. 2010). Such images can subsequently be used for evaluating glare through
software tools that were specifically designed for this purpose (Kumaragurubaran and Inanici
2013; Wienold et al. 2004).
HDR images generated through Radiance can be used as tools for preliminary, early-stage glare
evaluation (Jones and Reinhart 2017; Reinhart and Wienold 2011; Wienold 2009). Generating
HDR images through Radiance is far more computationally intensive than calculating illuminance.
Wienold (2009) proposed a hybrid solution by leveraging Daysim for calculating vertical
illuminance and using the rpict program from Radiance to generate low-quality HDR files. This

39
approach, in addition to being complicated, has not been successfully implemented in any
mainstream daylighting software.
Since the introduction of the rcontrib tool in Radiance, it has been possible to generate daylight
coefficient-based HDR images (McNeil 2013a; 2013b; 2013c; 2014). However, there have been
no documented efforts to evaluate this capability as a means of generating such images for glare
evaluation. The study presented in this chapter is an attempt in that direction.

4.2 Aim, Rationale and Hypotheses


This study seeks to determine if HDR images generated through daylight-coefficient based
simulations can be used as a means of evaluating glare. The rationale for doing such an
investigation relates to the savings in computational effort and runtime that can be achieved by
generating multiple images through matrix operations (as is the case with daylight coefficients).
After a one-time computation of the coefficient matrices, a time series of images can be generated
by multiplying the coefficient matrices with a sky-matrix.
An hourly annual glare calculation involves the generation of 3650 images based on a 10 hour/day
occupancy schedule. Considering that each conventional ray-tracing simulation is an independent
simulation, a daylight coefficient-based approach for generating a large number images can
provide significant savings in time and effort.
The standard approach of generating such images involves shadow testing, conventional Monte-
Carlo raytracing and ambient caching algorithms through a Radiance tool called rpict. Rpict is the
image-centric counterpart of rtrace, the tool that was used to generate illuminance through
conventional raytracing in Chapter 3. The images generated through the daylight-coefficient
based approach rely on shadow testing and Monte-Carlo approach through a tool called rcontrib.
So, within the context of Radiance, the aim of this study is to test the suitability of rcontrib-
generated images for glare evaluation.
The following are the hypotheses for this study:
1. HDR images generated through the Daylight-Coefficient Method, which also incorporate
precise solar position and luminance in the simulation, can be a reliable alternative to
conventionally generated Radiance HDR images for the purposes of glare evaluation.
2. Increasing the level of sky discretization for daylight-coefficient calculations will yield more
accurate images. The precision implied in this case relates to comparing the glare metrics
derived through such images with the glare metrics derived from images generated
through conventional raytracing.

4.3 Methodology
4.3.1 Overview
Two sets of image-based simulations were performed for this study. The first set of images, which
served as the benchmark, were generated through conventional raytracing with rpict. The second
set of images, which were to be tested as alternatives, were generated through daylight coefficient
simulations. The primary Radiance programs used for image-based daylight coefficient
simulations are rcontrib (for raytracing), dctimestep (for matrix-multiplication) and pcomb (for
image addition and subtraction).
The Radiance model used for the simulations was the same one used for preliminary study in
Chapter 3. The dimensions of this model are shown in Figure 22 and its surface properties are

40
listed in Table 1. The weather data used for simulations was the same as that used in Chapter 3
as well. So, a total of 240 sky conditions were considered for the simulation (based on 10 hours
x 4 days x 6 locations). Two views were chosen for the analysis. As shown in Figure 38, one of
the views faced east and the other south. Therefore the total number of individual images was
480. The number of independent raytracing simulations are expressed as a function of
independent variables in Table 9. The number of daylight coefficient simulations corresponding
to these settings are shown in Table 10.

Figure 38. The two views considered for the study.

Table 9. Number of conventional raytracing simulations.

Independent Variable Levels Quantity


Birmingham (AL)
Denver (CO)
Fargo (ND)
Location 6
Pittsburgh (PA)
Phoenix (AZ)
Seattle (WA)
Hourly Sky files
Weather Data (10 hours/day for 40
Solstice and Equinox)

Views East and South 2


Total number of simulations 6*40*2=480

41
Table 10. Number of daylight coefficient simulations.

Independent Variable Levels Quantity


Birmingham (AL)
Denver (CO)
Fargo (ND)
Location 6
Pittsburgh (PA)
Phoenix (AZ)
Seattle (WA)
Weather Data Sky Matrix File 1
MF1, MF2, MF3,
Sky discretization 5
MF4, MF5
Views East and South 2
Total number of simulations 6*1*5*2=60

In the context of the value of 60 derived in Table 10, it should be noted that each daylight
coefficient simulation was actually a combination of a diffused simualtion, a direct-diffused
simulation and direct-only simulation. These three types of simulations are depicted through an
example in Figure 39. The results from each of these stages is arithmetically combined as per
R,G and B channels of the HDR image into a single resultant image. As hypothesized in section
4.2, this resultant image is expected to be a reliable alternative to conventionally raytraced
images. The direct-only simulation employed the same custom script that was used for the direct-
sun component for illuminance-based simulations. So, the position of the sun was exepcted to be
in the correct location in the resultant image.
Figure 40 shows a comparison between resultant images so generated for skies with varying
levels of discretization. Comparing the falsecolor images from MF1 to MF5 with the benchmark
image generated through conventional raytracing, it can be inferred that the level of sky
discretization does have an impact on at least how the sky and circumsolar region are perceived
by the observer.

Figure 39. Images generated through three stages of the matrix-based process.

42
Figure 40. A comparison between falsecolor images generated

4.4 Results
Figure 41, Figure 42 and Figure 43 show the plots of Daylight Glare Probability (DGP) values for
east and south views for Birmingham, Denver and Fargo. The values of DGP were derived
through the Radiance-based evalglare program. As the plots indicate, for most of the instances,
the values of DGP calculated through conventional raytracing and daylight coefficients were
nearly the same. Deviations in results between the conventionally raytraced images and daylight
coefficient-based images also indicate a deviation based on sky discretization values. One clear
indicator of the effect of sky discretization on results is the sum of solid angles of glare sources
estimated by evalglare. Figure 44 shows the sum of solid angles for Fargo, ND for the same hours
for which the DGP data was plotted.
The impact of sky discretization on results, especially when the sun is in the field of view, can be
explained through Figure 45. It is evident that a higher level of discretization leads to a better
estimate of the location and magnitude of potential glare sources. This, in turn, allows for a more
accurate calculation of the position index for glare studies.

43
Figure 41. Daylight Glare Probability (DGP) for east and south views for Birmingham.

Figure 42. DGP for east and south views for Denver.

44
Figure 43. DPG for east and south views for Fargo.

Figure 44. Sum of solid angles of glare sources (in steradian) for the east view for Fargo.

45
Figure 45. Glare sources superimposed on the HDR images for 10:30 AM for Fargo.

4.5 Conclusion
Unlike the results from the previous chapter, the simulation results generated in this study indicate
that employing skies with high levels of discretization might lead to more accurate simulation
results. As values in both Table 11 and Table 12 show, there is a consistent increase in the
precision of the simulation with the increase in the level of sky discretization.
Table 11. Root Mean Square Error (RMSE) values for DGP values for MF1 to MF5 approaches for the south
view. The reference data are the DGP values obtained through conventional raytracing.

Location MF:1 MF:2 MF:3 MF:4 MF:5


Birmingham 0.0306 0.0302 0.0301 0.03 0.0298
Denver 0.0637 0.0304 0.0131 0.0099 0.0065
Fargo 0.134 0.1177 0.1142 0.0813 0.0813
Phoenix 0.0803 0.0517 0.0483 0.0446 0.0429
Pittsburgh 0.0368 0.012 0.0107 0.0105 0.0105
Seattle 0.0283 0.0272 0.0257 0.0229 0.0205

46
Table 12. Root Mean Square Error (RMSE) values for DGP values for MF1 to MF5 approaches for the east
view. The reference data are the DGP values obtained through conventional raytracing.

Location MF:1 MF:2 MF:3 MF:4 MF:5


Birmingham 0.0147 0.0143 0.0129 0.0123 0.0125
Denver 0.1495 0.145 0.1265 0.1188 0.0925
Fargo 0.0907 0.0817 0.0759 0.0679 0.0644
Phoenix 0.026 0.0184 0.0158 0.0165 0.0161
Pittsburgh 0.0523 0.0563 0.0552 0.0531 0.0492
Seattle 0.0733 0.0634 0.0585 0.0551 0.0515

The improvement in results, however, is not always commensurate with the amount of
computational effort required. An MF1 simulation requires only 145 HDR files to be created during
the simulation. The number of files for a MF5 simulation requires 5185 HDR files to be created.
The implications of such a high volume of files being generated are evident from Figure 46, which
shows the disk space required for storing the files generated through MF1 – MF5 simulations.
Files created through MF4 and MF5 simulations require approximately 16 and 25 times more
space than an MF1 simulation.

Figure 46. Disk space required to store coefficient-based image files created by MF1 – MF5 simulations. The
calculation is based on a file size of 1.4MB.

The other consideration with respect to using simulations with highly discrete skies relates to the
runtime for raytracing calculations. Rcontrib, the Radiance program employed for raytracing
calculations, generates the images for coefficients simultaneously. For example, in the case of an
MF3 simulation this implies generating 1297 (144*3*3+1) images at the same time. The runtimes
for the simulations performed in this study are shown in Figure 47. So, based on RMSE values
as well as disk space utilization and simulation runtimes, it can be concluded that the while higher
sky discretization might lead to a more precise simulation, its application is only justifiable in cases
where the need for higher precision over-rides any other computational performance-related
considerations.

47
Figure 47. Raytracing runtimes for creating images required for MF1 – MF6 simulations. All the simulations
were performed on 30 dedicated cores of an Intel Xeon E5-2680 (v2 @2.80GHz) processor.

48
Chapter 5. View-Matrix Subdivisions for Multi-Phase Daylighting Simulations

5.1 Background
The motivation behind the research and development of the multi-phase methods is discussed in
Section 2.3 of Chapter 2. In each of these methods, the flux-transfer path between an indoor
space and the sky is split into multiple phases. The ray-tracing result from each phase is stored
in a matrix. After creating all the matrices, the illuminance or luminance is calculated by matrix
multiplication, which in some cases is followed by matrix addition. Figure 48 shows a schematic
diagram for the Three-Phase Method. As indicated by the figure, the view matrix accounts for
flux transfer from the glazing to the interior of the room. The flux transfer within the glazing, or any
fenestration system that is present within the glazing aperture, is accounted for by the
transmission matrix.

Figure 48. A schematic diagram for the Three-Phase Method.

The transmission matrix, which is either created from empirically measured data from a glazing
or shade sample, or through a Radiance program called genBSDF, is a data-structure for a Bi-
directional Scattering Distribution Function (BSDF). The transmission matrix relates incident flux
directions to exitant flux distributions for glazing or complex fenestration systems. The BSDF
format compatible with Radiance is based on the WINDOW 6 XML format which contains a Bi-
directional Transmission Distribution Function (BTDF) and Bidirectional Reflectance Distribution
Function (BRDF). The WINDOW 6 format employs a Klems angle basis to allocate radiation
values for incoming and outgoing directions. A full Klems basis consists of 145 incoming and
outgoing directions. Figure 49 depicts a typical distribution pattern for a BSDF with full Klems
basis.
The experimental setup for validating the Three-Phase Method consisted of a simple room with a
single glazing (McNeil and Lee 2012). Subsequent research on the Three-Phase Method also
has relied on simple experimental setups for field studies as well as simulation–based studies.

49
Figure 49. The images on the left and right show the front and back transmission properties for a BSDF.

A Three-Phase simulation for a single space can employ one or more view matrices. The current
recommendation, and prevalent practice, is to assign one view matrix per glazing aperture. A
simple example is shown in Figure 50. It is apparent from Figure 50 that assigning a single view
matrix to an aperture results in the glazing or fenestration system present in that aperture acting
as a singular entity whose optical properties are a function of 145 incident and exitant directions.
This simplistic assumption, while appropriate for the kind of spaces considered in validating the
Three-Phase Method, is not appropriate for complex spaces.
The limitations of employing a single view matrix for complex spaces has been acknowledged by
the researchers who developed the Three-Phase Method. The remedial measure recommended
by them involves subdividing the glazing aperture(s) and then generating view matrices
corresponding to individual subdivisions (McNeil 2013c). While increasing the granularity of the
glazing might result in more accurate simulations, this increase is also directly proportional to the
computational effort required in the simulation. For example, the Three-Phase simulation with six
view matrices as shown in Figure 51 is actually six independent three-phase simulations whose
results are arithmetically combined.

Figure 50. A Three-Phase Method rendering. The patterns on the glazing are indicative of the Klems-basis
employed in the BSDF data-structure.

50
Figure 51. A simulation with multiple view-matrices. As shown by the image on the right, the results from
individual simulations are added arithmetically to derive the final result.

There have been no documented efforts to numerically evaluate the impact of view-matrix
subdivisions on the precision and computational efficiency of the Three-Phase Method6. However,
gaining a definite understanding of this aspect of the Three-Phase Method is essential in
determining whether it can be of practical relevance to daylighting simulations involving models
that are more complex than simple rooms with unidirectional sidelighting.

5.2 Aim and Hypotheses


The principal goal of this study is to quantitatively evaluate the impact of sub-dividing view-
matrices on the precision and computational efficiency of the Three-Phase Method. The relevance
of the specificity of this study can be explained by considering a scenario shown in Figure 52.
One of the configurations in that figure contains 24 view matrices while the other contains 2. Since
each view matrix entails a separate calculation, it follows that the calculation with 24 view matrices
will be far more computationally expensive. There are presently no information resources that can
provide guidance on when a higher number of view matrices might be appropriate and if so, how
their arrangements should be configured. The simulations and the analyses performed in this
study are meant to provide a better understanding about the criticality of view-matrix subdivisions
to the Three-Phase Method and all the further methods that are derived from it.
The working hypotheses for this study are the following:
1. A more discretized glazing system is likely to be accurate in instances where the building
is surrounded by external obstructions or external shading devices.
2. Using a single view matrix for a glazing system with windows in multiple directions is
expected to be detrimental to the precision of the simulations.
3. Horizontal subdivision-schemes are likely to be more effective in the presence of external
shading devices such as overhangs.

6
A tutorial on the Three-Phase Method by McNeil (2013c) states that a “simulation may need to be split into many parts with more
than one D (daylight) matrix to reproduce non-uniform shading effects by external obstructions”. However since that document is
essentially an instructional manual, it does not numerically justify the advantages of using subdivisions or investigate which type of
subdivisions (such as vertical horizontal etc.) for a particular space.

51
Figure 52. Two view-matrix configurations. The wide angle view is shown below the perspective view as it
displays all the view matrix apertures. The red checkered arrangement too is shown for the sake of clarity. Two
apertures of the same color do not indicate that they apply the same view-matrix.

5.3 Methodology
Like the studies described in Chapters 3 and 4, this study involves a series of comparative
evaluations between a benchmark method and a test method. However, unlike the previous
studies, the test method in this case is not a new method proposed by the author. The Three-
Phase Method is being employed as the test method to test the performance of different view-
matrix configurations.
The model used for the study is shown in Figure 53 . The rationale for choosing a north-facing
room was to avoid any confounding effects due to the impact of direct or direct-diffuse radiation
through the BSDF layers that constitute the transmission matrix. The optical properties of the
room are shown in Figure 53. All the surfaces for the Three-Phase and benchmark daylight
coefficient simulations are the same except for the transmission matrix in the Three-Phase
Method. The BSDF file for the transmission matrix was created through the Radiance genBSDF
program. A total of 450 sensors were employed for this study. The rationale of using a larger room
for this study, as compared to the ones used previously, was to allow the results to be influenced
by the surrounding geometry and also to create a demarcation between the influence of the
glazing on the two sides. The positioning of photosensors is shown in Figure 54.

52
Figure 53. Dimensions of the room used for the study. The front glazing faces north.

13. Optical properties of the surfaces of the room.

Figure 54. Positioning of the photosensors. 450 photosensors in a 30x15 grid were employed for the study.
The spacing between the photosensors is 2 feet by 2 feet and all photosensors are 2.5 feet above the floor.
The red arrows indicate the orientation of the photosensor.

53
5.3.1 Benchmark daylight coefficient simulations
The weather data locations and timestamps employed for this study were the same as those
employed in the previous chapters (i.e. six locations across the United States and 10 hours/day
for two days of solstice and equinox). Further details about the weather data can be found in
Appendix A. Based on the intention of analyzing the suitability of different view-matrix
configurations, the benchmark simulations were assigned on three criteria:
1. Presence of contextual geometry.
2. Presence of external shading.
3. A structural division in the glazing due to the presence of mullions.
The independent variables for the study are summarized in Table 14. The options relating to the
contextual geometry, external overhang and mullions are shown through renderings of the model
in Figure 55, Figure 56 and Figure 57. The daylight coefficient simulations were iterated through
convergence in a manner similar to that described in Chapter 3.
Table 14. Independent variables for the daylight coefficient simulations.

Independent Variable Levels Quantity


Birmingham (AL)
Denver (CO)
Fargo (ND)
Location 6
Pittsburgh (PA)
Phoenix (AZ)
Seattle (WA)
Present,
Contextual Geometry 2
Not Present
Present,
External Overhang 2
Not Present
Present,
Mullions present 2
Not Present
Total number of simulations 6*2*2*2=48

Figure 55. The model room without and with the 2 feet overhang. The overhang was used to study the effect of
external façade-based geometry on the results.

54
Figure 56. The model room without and with mullions.

Figure 57. The images show the room with and without contextual geometry. The shadow cast on the room by
the contextual geometry can be seen in the image on the right.

5.3.2 Three-Phase simulations


Based on the presence or absence of mullions across the glazing area, two sets of view-matrix
configurations were conceptualized for this study. Figure 58 shows the view matrix configuration
for glazing with mullions and Figure 59 shows the view matrix configuration for glazing without
mullions. Since the mullions act as a natural divider of the glazing, the vertical subdivisions
employed in the simulations with mullions were less than the ones in the simulations without
mullions. Based on the view-matrix configurations and the 3 factorial design for the geometric
elements described in Table 14, a total of 312 Three-Phase simulations were performed. It should
be noted, however, that the actual ray-tracing calculations in these 312 simulations varied widely.
For example for option GM shown in Figure 58, only a single view-matrix is employed. On the
other hand, 24 view-matrices are employed in the option GM2VH, thus necessitating 24 individual
Three-Phase simulations that are then combined arithmetically to generate a single illuminance
result.

55
Figure 58. View-matrix apertures for the model with mullions. As explained previously, the red checkered
pattern is employed to clarify the number of apertures and does not imply similarity. For example, the
configuration referred to as GM contains a single view matrix while GM2H contains 12.

Figure 59. View-matrix apertures for a model without mullions.

56
Table 15. Independent variables for the Three-Phase simulations.

Independent Variable Levels Quantity


Birmingham (AL)
Denver (CO)
Fargo (ND)
Location 6
Pittsburgh (PA)
Phoenix (AZ)
Seattle (WA)
Present,
Contextual Geometry 2
Not Present
Present,
External Overhang 2
Not Present
Present,
Mullions present 2
Not Present
7 with No Mullions
View Matrices 7+6
6 with Mullions
6*2*2*1*7
Total number of simulations +
6*2*2*1*6=312

The entire set of simulations for both the Daylight Coefficient Method and the Three-Phase
Method involved only stochastic sampling as the sky model used for the simulations was based
on 145 Tregenza sky patches. Direct sun corrections like the ones performed for the simulations
in Chapters 3 - 4 was not performed in this study, as the motivation of the study was to evaluate
the performance of the view-matrix subdivisions alone.

5.4 Results
The Peak Signal to Noise Ratio (PSNR) plots for four representative cases are shown in Figure
60 - Figure 63. When compared with the results from Chapter 3, the general trend of PSNR values
indicates a greater amount of distortion. This is expected and mostly attributable to the loss of
precision due to the use of a Klems-basis in the BSDFs that intercept flux from the external part
of the scene to the room.
The comparison between the results for the different models indicate that higher number of
subdivisions usually leads to less distortion in results. The plots indicate a consistently better
performance for the model G4VH (containing 24 view-matrices as per Figure 59) in instances
where the space is shaded through external context or the overhang. However, as indicated in
Figure 60 and Figure 61, except for the configuration with a single view matrix, there are no view-
matrix subdivision schemes that offer consistently poor performance. As shown in Figure 62, the
separation between the distortion values is more aparent in the case of the model with overhang
and external contextual geometry. No significant deviations in the results were obtained due to
the presence of mullions in the simulations. This can be attributed to the fact that the glazing
subdivision obtained through mullions were also factored into the simulations without them
through the use of extended vertical subdivisions as shown in options GM2VH and G4VH in
Figure 58 and Figure 59 respectively.
The full set of PSNR plots for all the simulations from this study are provided in Appendix E. The
following section provides a general summary of the results with the intention of outlining best-
practice approaches for Three-Phase simulations with different configurations of shading and
contextual geometry.

57
Figure 60. PSNR plots of different view-matrix configurations for the model with no mullions, no overhang and
no contextual geometry. The data are plotted for the locations of Birmingham and Seattle.

Figure 61. PSNR plots of different view-matrix configurations for the model with no mullions, no contextual
geometry with a 2’ feet overhang in the north-facing glazing. The data are plotted for the locations of
Birmingham and Seattle.

58
Figure 62. PSNR plots of different view-matrix configurations for the model with no mullions, external
contextual geometry and a 2’ feet overhang in the north-facing glazing. The data are plotted for the locations
of Birmingham and Seattle.

Figure 63. PSNR plots of different view-matrix configurations for the model with no mullions and external
contextual geometry. The data are plotted for the locations of Birmingham and Seattle.

59
5.5 Guidelines for assigning view-matrix subdivisions
The following subsections provide guidelines for general simulation models that can be classified
according to the three-point criteria listed in 5.3.1.

5.5.1 Models with no contextual geometry and no external shading elements


For spaces with simple glazing, the PSNR plots indicates that single view-matrices for windows
in multiple directions distinctively underperforms compared to all the other view-matrix
configurations. Considering the simulation times shown in Figure 64, it is apparent that higher
number of matrices correlate to higher simulation runtimes. So, for spaces with simple glazings,
any matrix configuration that assigns separate view-matrices as per the directional normal of the
glazing should suffice as a view-matrix-configuration.

Figure 64. Simulation runtimes for model with no external shades or contextual geometry. The data-point
NM-NG indicates the time taken for the benchmark Daylight Coefficient Method simulation. The labels (in red)
denote the quantity of view-matrices, and therefore also the daylight matrices, in each simulation method.

5.5.2 Models with no contextual geometry that have external shading elements
The external shading element for the model considered in this study consisted of a 2-foot
overhang. For the simulations involving the overhang, view-matrix configurations featuring vertical
subdivisions were more accurate than the others. This can be attributed to the fact that the
shading provided by the overhang influences only a certain portion of the glazing height. View-
matrix configurations with vertical subdivisions are identified as GH, G2VH and G4VH in Figure
59 as well as the corresponding PSNR plots and simulation runtime plot. The simulation runtime
plot is shown in Figure 65. As indicated by the figure, simulations involving G2VH and G4VH
required nearly 200% and 400% more runtime compared to GH. Further, the PSNR plots indicate
that the horizontal subdivisions did not result in any consistent improvements in precision. So, for
spaces with external shading elements, vertical subdivisions alone would suffice to improve the
precision of the simulation.

60
Figure 65. Simulation runtimes for a model with no contextual geometry that has external shading. The data-
point NMO-NG indicates the time taken for the benchmark Daylight Coefficient Method simulation.

5.5.3 Models with contextual geometry and external shading


Simulations involving contextual geometry and external shading were the ones where the results
were most influenced by the type of view-matrix subdivision employed. Based on the PSNR plots,
G2VH and G4VH were the most accurate configurations. However, as Figure 66 highlights, the
runtimes for G2VH and G4VH are 500% and 1100% higher than that for a conventional simulation
with daylight coefficients. So, in such instances, if the quantity of glazing options being iterated
through are few, it might be more efficacious to perform multiple runs of daylight coefficient-based

Figure 66. Simulation runtimes for a model with contextual geometry and external shading. The data-point
NMO-WG indicates the time taken for the benchmark Daylight Coefficient Method simulation.

61
simulations rather than performing a Three-Phase simulation with multiple view-matrices.

5.5.4 Models with mullions


The purpose of incorporating mullions in the study was to ascertain if the presence of physical
obstructions in the glazing had any bearing on the performance of a particular view-matrix
configuration. However, no performance trends specifically attributable to mullions were
observed. The PSNR plots and runtime plots for simulations with mullions are provided in
Appendix E.

5.6 Conclusion
This study was conceptualized to provide preliminary guidance for selecting view-matrix
configurations based on the specificity of the model being simulated through a through Three-
Phase Method simulation. The guidelines provided in the previous section indicate that a
simplistic, single view-matrix configuration is not suited for spaces that are more complex than
rooms with unidirectional sidelighting
The results also indicate that the Three-Phase Method might not be the right approach for models
that have multiple windows and are shaded, intentionally or unintentionally, through multiple
geometric features.

62
Chapter 6. Applicability of the Four-Phase Method for iterative simulations with
non-coplanar shades

6.1 Background
Shading devices play a critical role in daylighting design. A well-designed shading system will
allow sufficient diffused daylight into a space while blocking or curtailing glare-inducing direct
sunlight. The use of interior shading devices such as blinds and fabric shades is now
recommended in design guides as well as building codes. For example, IES LM 83-12 and LEED
v4 mandate the application of shades in the analysis of Spatial Daylight Autonomy (sDA) (IES
2012; USGBC 2014). There is considerable amount of anecdotal as well as empirical support in
favor of external shading devices. Devices such as overhangs, mechanically adjustable vertical
shading-fins and awnings have been shown to be beneficial in controlling daylight glare as well
as improving the thermal environment of indoor spaces(Gugliermetti and Bisegna 2006; Hien and
Istiadji 2003; Jeong Tai Kim and Kim 2010; Kim et al. 2012) .
As discussed in Chapter 2, the Four-Phase method was introduced for the purposes of enabling
parametric simulations with external shading systems. However, there have been no documented
efforts yet to quantify the benefits of employing the Four-Phase Method over the conventional
Daylight Coefficient Method. The preliminary research on the Façade Matrix, which is the
additional matrix involved in the Four-Phase Method, has focused instead on empirical validation
of the accuracy of this method. Any potential applications of this method to shading system
optimization are also yet to be studied.

6.2 Aim and Hypotheses


This chapter describes a two-part study on the Four-Phase Method. The first part quantitatively
evaluates the implicit benefit of using the Four-Phase Method over the Daylight Coefficient
Method from a computational-runtime perspective. The second part of the study demonstrates a
way to utilize the illuminance data generated through the Four-Phase simulations to derive a
control algorithm for the shading systems evaluated in such a manner.
The primary working hypothesis for this study is that for the entire set of shading configurations
considered for this study, the cumulative runtime for the Four-Phase Method will be lower than
that for the conventional Daylight Coefficient Method.

6.3 Methodology
6.3.1 Setup
The model considered for this study consists of a simple room with a single south-facing glazing
system. The room is shaded from direct sun solely through the use of an external adjustable
louvered overhang. The room and the overhang are shown in Figure 67. Figure 68 provides the
dimensions of the room. The optical properties of the different surfaces are listed in Table 16. The
dimensions of this model are mostly the same as those considered for the preliminary model in
Chapter 3. The positioning of the 100 photosensors considered for this study are the same as
depicted in Figure 23.
The overhang considered for this study can be rotated along two degrees of freedom as shown
in Figure 69. Each setting of the overhang requires an individual simulation. In the case of the
Daylight Coefficient Method, individually simulating the overhang will require six separate
simulations. In the case of the Four-Phase Method, however, the scope of each iteration can be

63
confined to simulating the Façade-Matrix part of the simulation alone as the change in each
iteration is restricted to the façade part of the simulation.

Figure 67. The software model of the room considered for this study. The louvered overhang is adjustable
through two degrees of freedom.

Figure 68. Dimensions of the room shown in Figure 4. The thickness of the walls is 9” and the dimensions of
the floor are 20’x20’. The dimensions of the glazing are 18’ x 5’.

Table 16. The optical properties of the diifferent surfaces in the model.

64
Figure 69. The images above show various settings for the adjustable shading system considered for this
study. The overhang shown above can be adjusted along two degrees of freedom: along the local axes of
individual fins and along the main axes of the entire overhang. The configurations considered for the
simulation include three settings for fins and two settings for the overhang. The angles considered for the
rotation of fins were 0°, 45° and 90° while that for the entire overhang were 0° and 30°. From here on, the
settings will be referred to by their designation as mentioned below each image.

Table 17. Designation of shade setting based on clockwise rotation angles of the fins and the overhang.
Rotation of fins corresponds to each fin on the overhang being rotated about its local x-axis. Rotation of the
overhang corresponds to the rotation of the entire overhang structure about its local x-axis.

Designation Clock-Wise Rotation


Fins Overhang
F0S0 0° 0°
F0S30 0° 30°
F45S0 45° 0°
F45S30 45° 30°
F90S0 90° 0°
F90S30 90° 30°

6.3.2 Simulation parameters


Table 18 lists the number of ambient bounces, which is directly related to simulation-runtime,
considered for the different matrices of the Four-Phase Method. Table 19 details the number of
times a particular matrix was calculated in this study. In summary, the total number of ambient
bounces considered for the Daylight Coefficient Method was (6 x 12) while that for the Four-Phase
Method was (1 x 4 + 6 x 12 + 1 x 4).

65
Table 18. The table below lists the number of ambient bounces (reflections) considered in each iteration for
different matrices of the Four-Phase Method. As indicated by the left-most column, the Daylight Coefficient
Method involves a one-step simulation involving 12 reflections. The number of ambient bounces is directly
proportional to the amount of time required for computation.

F-Matrix Method Daylight Coefficient Method


V Matrix 4
F Matrix 9 12
D Matrix 4

Table 19. The number of times each matrix was calculated during the course of the study.

F-Matrix Method Daylight Coefficient Method


V Matrix 1
F Matrix 6 6
D Matrix 1

The matrix calculations listed in Table 19 were performed during the course of six separate annual
daylighting simulations. The weather data chosen for the simulation was a TMY3 dataset for State
College, PA. Each simulation was intended to evaluate the performance of a particular setting of
the overhang described in Figure 69. The illuminance data generated through the simulations was
used to select an appropriate setting of the overhang for each hour of the year. The simulation-
runtimes for the Four-Phase Method simulations were compared with that of the conventional
Daylight Coefficient Method to quantify the advantage of using the Four-Phase Method from a
computational perspective.

6.4 Results
6.4.1 Comparison of simulation runtimes
Table 20 provides the average simulation time for Facade-Matrix and Daylight Coefficient Method
according to different phases. The View (V) and Daylight (D) matrix were only required to be
calculated once during the study as the change in the model from one simulation to the other was
restricted to the façade part of the simulation. The advantage of this phase-based approach is
evident in the data visualized in Figure 70.
The first Facade-Matrix simulation required a greater simulation-runtime than the Daylight
Coefficient Method as the matrices for the V and D phases were calculated in that simulation. The
Facade-Matrix simulations performed thereafter were able to utilize the results of the V and D
phases from the first simulation. As Figure 70 indicates, the runtimes for the subsequent
simulations were nearly 50% lower than the first simulation. Figure 71 provides a cumulative
comparsion between the runtimes for the Four-Phase Method and the Daylight Coefficient
Method. The total runtime for simulations using the Four-Phase Method was 26% less than that
for the Daylight Coefficient Method.
Table 20. The table below lists the mean simulation runtime (in minutes) for different aspects of the six Four-
Phase Method and Daylight Coefficient Method simulations. The values in parenthesis indicates the standard
deviation.

F-Matrix Method Daylight Coefficient Method


V Matrix 37
F Matrix 76 (3.87) 122 (3.39)
D Matrix 48

66
Figure 70. The above figure shows the individual runtimes for the Four-Phase Method and Daylight Coefficient
Method while simulating different settings of the overhangs shown in Figure 6.

Figure 71. Cumulative runtimes for the Four-Phase Method and the Daylight Coefficient Method. From a purely
computational perspective the used Four-Phase Method required 26% less time than the Daylight Coefficient
Method.

6.4.2 Annual illuminance results for different configurations of the shading system
The annual simulations for the six settings of the overhang, as described previously in Figure 69,
generated six sets of time-series illuminance values for the entire year. A real-world shading
system like the one described in this study will ideally be adjusted such that the illuminance levels
inside the space are maintained at a visually comfortable level. The control algorithm used in this
research was designed to select a configuration of the shading system, based on the settings
shown in Figure 69, such that the illuminance on the measurement point 9 feet from the side wall
and 5 feet from the glazing did not exceed 3000 lux. The measurement point is shown in Figure
72 and the shade settings selected based on the control algorithm shown in Figure 73. An annual
hour-based representation of the shade-settings is provided in Figure 74.

67
Figure 72. The above figure shows the resultant illuminance grid for the F0S0 setting of the overhang at
10:30AM on 1st January. The control algorithm for selecting a particular shade setting for each hour was
assigned on the basis of a grid-point located 5’ away from the glazing and 9’ away from the right wall. That grid
point is highlighted through a translucent polygon in the above figure.

Figure 73. The above bar-plot shows the total number of times each setting of the overhang was selected by
the control algorithm. An hourly annual representation of the above data shown in Figure 74.

68
Figure 74. The above grid-plot shows the hour-wise overhang settings for the entire year. The control algorithm
was designed to operate the shade between 8:30 AM and 5:30PM. In the absence of natural light, the control
algorithm reverted to the default F0S0 setting.

6.5 Conclusion
As shown in Figure 71, simulations performed with the Four-Phase Method resulted in runtimes
that were around 26% lower than that for the Daylight Coefficient Method simulations. The data
visualized in Figure 71, however, does not account for the amount of time and effort required for
setting up the Four-Phase simulations. As the Four-Phase Method minimizes recalculation of
certain aspects of a simulation through data-reuse, the trade-off between runtimes and
computational effort is likely to be higher in cases where a greater number of iterations are
involved. Although an effort was made to perform both the Facade-Matrix and daylight coefficient
simulations such that their results converged to approximately the same values, significant
differences were observed especially in cases where the direct-normal radiation of daylight was
high.
The control algorithm used to select different settings of the overhang was designed to monitor a
single point at workplane height and select the shade setting such that the illuminance at that
point was closest to 3000 lux. A secondary logic in the algorithm was incorporated to minimize
the number of total grid points in the room above 3000 lux once the primary condition was met for
an overhang setting. The results generated through the application of the algorithm, as shown in
Figure 73 and Figure 74, indicate that the prevailing illuminance levels in the room did not require
the entire overhang structure to be tilted for shading for nearly 80% ((262+285+2370)/3650) of
the operating time. This suggests that the shading in the case of the room considered in this study
can be accomplished mostly by manipulating the fins of a stationary overhang structure. Analyses
of this nature can be employed in real-world projects to provide a more informed opinion on the
use of overhangs or similar external shading structures, and their preferred settings.
The scope and intention of this study was to explore the potential of the Four-Phase Method as a
viable means to evaluate adjustable external shading systems. The results from the analysis
indicate that employing the Four-Phase Method can be advantageous in parametric simulations
with multiple factors. The control algorithm and approach to selecting the overhang settings
discussed in this study demonstrate that results generated from such simulations can be
leveraged to improve the selection-process for external shading systems.

69
Chapter 7. Software Development for Parametric Daylighting Simulations

This chapter describes the motivation and conceptual methodology underlying the development
of an open source software platform that facilitates parametric daylighting simulations. The
parametric simulations addressed by this software primarily relate to the studies described in
Chapters 3 – 6. This chapter also provides a high level overview of the software development
process that entailed the creation of the software platform.

7.1 Background
Respondents in a recent survey on Climate-based Daylight Modeling (CBDM) named 15
daylighting software that were commonly used in academia and industry. 11 out of those 15
software were derived from Radiance (Brembilla 2016). As the review in Chapter 2 shows,
Radiance and Radiance-based tools have been under consistent use and development for nearly
three decades. Furthermore, as the body of research cited for the studies in Chapters 3 - 6
demonstrates, the prominence of Radiance in daylighting practice and research can be expected
to continue in the foreseeable future.
Radiance is a collection of command-line programs developed primarily by Greg Ward over the
last three decades (Ward 1994a; 1994b; 2014; 2017; Ward and Rubinstein 1988; Ward et al.
1987; Ward 1994c; Ward et al. 1988). For performing a daylighting simulation directly with
Radiance, one needs to be familiar with the specificity of the tasks addressed by its various
programs, the esoteric vocabulary of commands and parameters understood by those programs,
and the format and structure of the datasets generated by the programs (Ward et al. 1998). The
steep learning entailed in gaining a proficiency of Radiance has been repeatedly acknowledged
over the years (Ashdown 2016; Estes Jr et al. 2004; Jacobs 2012; 2014; Ubbelohde and Humann
1998). Consequently, the developer of Radiance, as well as the organizations that have funded
its development over the past three decades, have encouraged the development of third-party
tools that conceal the complexity of Radiance through Graphical User Interfaces (GUIs). Since
the early 1990s, at least three dozen such tools have been developed (Brembilla 2016; Roy 2000).
Among these tools, some like Autodesk Ecotect have had highly sophisticated features and been
widely used, while others like HDRscope have addressed a very specific set of features of
Radiance (Autodesk Inc. 2010; Kumaragurubaran and Inanici 2013). From the perspective of this
research, the GUI-based tools developed for Radiance can be classified into two categories:
1. Tools that are tethered to popular modelling software: Such tools are typically
developed as plugins for software like AutoCAD, Rhino and Sketchup. They leverage the
inbuilt functionality of the modeling software for user interaction as well as data
visualization. Some of the popular tools of this type are DIVA4Rhino, Su2rad, Groundhog
and Honeybee (legacy version). A screen capture of Su2rad is shown in Figure 75.
2. Standalone tools that are modeling platform agnostic: These tools usually have their
own GUI and 3D modelling interface. The 3D modelling usually serves to display the
imported Radiance model from software like AutoCAD. Some of the popular examples are
Autodesk Ecotect, Adeline and Daysim. A screen capture of Adeline is shown in Figure
76.

70
Figure 75. A screen capture of Su2Rad. The dialog window shows a typical GUI for entering calculation
parameters.

Figure 76. A screen capture of Adeline. The CAD input needed for Radiance is created in separate software.

71
Daysim, a tool initially developed by Christoph Reinhart (2001), and subsequently by researchers
at Penn State University and National Research Council Canada (NRC), is a standalone program
which can also be deployed as the calculation engine for plugins described above (Daysim 2018;
Mistrick 2013). The tools shown in Figure 75 and Figure 76 are now defunct. This issue of such
tools becoming obsolete or defunct is emblematic of Radiance-based tools in general. The
following reasons can be attributed to this:
1. Differing development cycles: The development of tools tethered to popular modeling
platforms like AutoCAD should ideally be synchronized with the development cycle of the
modeling platform. However, this has rarely happened with Radiance-based tools due to
lack of funding or availability of developers.
2. Choice of development language: Most of the tools have been written in compiled
object-oriented languages like Java, C++ or C#. Deploying such tools on 32-bit and 64-bit
systems requires independent compiling and testing. As a result, any changes to the
source require considerable coordination between the core-developers of the tools.
3. Lag in development with respect to advances in research: At the time of writing this
document, most of the Radiance-based tools have not implemented any of the new
features introduced in the last eight years or so. Daysim, by far the most popular CDBM
tool still operates on the calculation engine that was written in the early 2000s. In the case
of Rayfront and Su2Rad, which are two Radiance-based tools with extensive sets of
features, this issue has led to a gradual decline in their user-base (Bleicher 2008; Mischler
2004).
At the time of conceptualizing this dissertation research, none of the mainstream Radiance-based
tools were found to have the functionality needed for studying or implementing the advanced
methods discussed in Chapters 3 - 6. So, a new Radiance-based tool was developed by the
author in collaboration with another developer. The following section details the background,
motivation and philosophy underlying the development of this software.

7.2 Motivation and development philosophy for a new parametric daylighting


software
From the perspective of this dissertation research, the software development was meant to serve
two broad purposes. Firstly, it was meant to facilitate the simulations and subsequent numerical
computation required for the studies described in Chapters 3 - 6. Secondly, it was intended to be
an Open Source platform that could be used either directly via-command line, or form the back-
end for plugins for popular platforms like Revit and Rhino.
The structure of the object-oriented framework designed for the software tool was expected to be
conducive to the Agile Development paradigm (Martin 2002; Warsta and Abrahamsson 2003).
This was to enable a continuous cycle of development, debugging, and feature-addition to the
software. Such a functionality would also allow for optimizing the workflows based on the
knowledge gained from studies like those performed in Chapters 3 - 6. A schematic diagram for
the Agile model of software development is shown in Figure 77.
The next section provides an overview of the way in which workflows for different simulation
methods like the Daylight Coefficient Method and the Three-Phase Method are typically realized
through native Radiance programs. It also details the attributes of such workflows that are
compatible to the goals described above.

72
Figure 77. Agile model of software development. The development cycle is meant to incrementally improve the
software. In the context of this research, this implies improving the software based on research, testing and
changes in functionality of Radiance or modeling interfaces. Image Credit: (Smartsheet 2018).

7.3 Compatibility of Radiance-based methods with software development goals


As described earlier, Radiance is a collection of several command-line programs. It was, and is
still being, designed according to the Unix Philosophy (Jacobs 2012; Ward et al. 1998).
This philosophy was summarized by Peter Salus (1994), a prominent computer historian, as:
1. Write programs that do one thing and do it well.

2. Write programs to work together.

3. Write programs to handle text streams, because that is a universal interface.


#1 and #2 from above are especially relevant to the way advanced matrix-based methods are
implemented through Radiance. Figure 78 shows the schematic diagrams for the Daylight
Coefficient Method, the Three-Phase Method and Four-Phase Method. As evidenced from the
repetition in the names and sequence of the Radiance programs used in both the figures, there
are several syntactical and procedural overlaps between the two methods. These aspects of
Radiance allow for the development of an external software tool that can rely on code-reuse and
also be scaled according to the complexity of different methods. The development of such a tool,
named Honeybee[+], began during the course of this dissertation research. Plugins developed on
top of Honeybee[+] now enable users to perform advanced multi-phase simulations through
Rhino, Revit and cloud-based systems (Ladybug Tools 2017b).
Honeybee[+] is one among several open source libraries developed collaboratively by a team of
building scientists and engineers. This collaborative effort operates under the umbrella of a project
called Ladybug Tools (Ladybug Tools 2017c). The author is an acknowledged contributor to two
open source libraries within Ladybug Tools, the primary one among them being
Honeybee[+](GitHub 2017a; 2017b). All the open source libraries developed through Ladybug
Tools to date have been distributed under the GPL-3.0 license.

73
Figure 78. Schematic diagram for the Daylight Coefficient Method, Three-Phase Method and Four-Phase
Method. The terms on the right detail calculation steps and Radiance-programs employed for them. The
programs are highlighted in red. Programs like rfluxmtx, gendaymtx and oconv feature in all the simulations,
often for accomplishing tasks that are common to the different methods.

7.4 Creation of Honeybee[+]: A high level software development summary


Honeybee[+] is eventually intended to be a whole-building simulation platform that links building
simulation tools like Radiance, EnergyPlus and OpenStudio to modeling platforms like Revit,

74
Rhino and Blender. The simulation features within Honeybee[+], as of January 2018, are primarily
focused on daylighting simulations. At present, it links Radiance to Revit and Rhino via
DynamoBIM and Grasshopper3D respectively. DynamoBIM and Grasshopper3D are visual
programming interfaces that facilitate parametric workflows for building design and simulation.
The development for Honeybee[+] commenced in December 2015 and the functionality relevant
to this research was incorporated into it by early 2017.
Thus far, Mostapha Sadeghipour Roudsari and the author are the only developers of this
software. The contribution details are shown in Figure 79. Roudsari is also the core-developer for
Ladybug, Honeybee and Butterfly, software tools which are meant for weather data visualization,
daylight and energy simulations, and computational fluid dynamics respectively. Honeybee, which
was released in 2014, is a Grasshopper3D-based plugin which relies on Daysim to perform
annual daylighting simulations. At the time of writing this dissertation, these tools had been
downloaded over 145,000 times (McNeel Europe S.L. 2018).

Figure 79. Contribution graph for Honeybee[+] as of December 2017. Number of “commits” refers to the
individual instances when the code was updated. The green and red numbers next to commits refer to the lines
of codes added and removed. The author’s contributions are credited as “sariths” Source: (GitHub 2018)

As is partially evident from Figure 79, Honeybee[+] consists of thousands of lines of software
code. This entire code was written in Python 2.x. The primary motivation for choosing Python was
its compatibility with popular modeling platforms like Rhino and Revit. Additionally, Python is a
cross-platform language that is interpreted and not compiled. As software written in interpreted
languages can be updated or debugged very quickly, they are conducive to the Agile paradigm
discussed in Section 7.2. The cross-platform functionality makes it possible to run Honeybee[+]
on Windows, Linux and OS-X operating systems seamlessly.
Considering the volume of the code written for Honeybee [+], a detailed explanation of the Python-
based code is impractical within this dissertation. Additionally, from the perspective of this
research, since the software was written to facilitate the studies conducted in Chapters 3-6, a

75
detailed exposition of the code by itself is secondary to the intent of this document. A more
software-centric overview of Honeybee[+] can be found in (Roudsari and Subramaniam 2016).
The logical design and functionality of Honeybee [+] is summarized in Figure 80. As the figure
shows, Honeybee[+] consists of a singular low-level core-library of classes and functions that
evoke native Radiance programs at runtime. The parameters required for the native Radiance
programs are translated by the core-library from inputs provided by higher level plugin-libraries.
The plugin-libraries are specific to individual modeling platforms. For example, the Rhino and
Revit plugin-libraries are separately written and leverage the same set of Python classes and
methods to run the simulations. Once a simulation has been completed by the native Radiance
programs, the results are retrieved back to the modeling interface through the plugin. The plugin
also leverages the native functionality of the modeling interface to visualize the results.

Figure 80. Logical structure of Honeybee[+].

A tiered approach of this nature has multiple advantages:


1. Data reuse: By encapsulating most of the Radiance-specific code in the core-library, it is
possible to create high-level plugins for different modeling platforms with minimal effort.
While the development for the core-library took nearly fourteen months, the plugins for
both Rhino and Revit were created in three weeks.
2. Better code synchronization: Any modifications to the functionality of Radiance by its
developer(s) can be addressed by affecting a corresponding change in the core-library.
Similarly, any change to a platform like Rhino can be addressed by modifying its plugin.
This implies that the stability of the majority of the code is impervious to changes within a
single high-level or low-level platform like Rhino or Radiance respectively.
3. Ability to port individual simulations across multiple platforms: As the core-library is
modeling platform-independent, it is possible to generate Radiance inputs using a
modeling platform like Rhino and then perform the actual simulation on a different

76
machine. This feature is especially useful for running remote cloud-based simulations and
was heavily leveraged during this research.
Since Honeybee[+] was developed as an open source software, the documentation of the code
by itself is intended to foster further contributions by other software developers. The extensive
use of Python docstrings enabled the code to be self-documenting and compatible with
Application Programming Interface (API) documentation platforms like Sphinx as shown in Figure
81.

Figure 81. API documentation for Honeybee[+] using the Sphinx platform. The API is auto-generated based on
the document strings created for different components of the code (Ladybug Tools 2018a).

7.5 Public release, complementary activities and response from the AEC
community
The software-development methodology and motivation for Honeybee[+] were discussed by the
developers during the 15th International Radiance workshop in Padova, Italy through a remote-
presentation (Roudsari and Subramaniam 2016). Honeybee[+] was publicly released on 22nd
August 2017 in the following formats:
1. A python-based command-line library. This is referred to as the “core-library” in previous
section.
2. Plugin for Grasshopper3D. Grasshopper3D is a parametric interface for Rhino.
3. Plugin for DynamoBIM. DynamoBIM is a parametric interface for Revit.
The functionality implemented in the first release drew upon the methods discussed in Chapter 3,
Chapter 4 and Chapter 5. These were referred to as “Daylight Coefficients (grid-based)”, “Daylight
Coefficients (image-based)” and “Three-Phase Method (grid-based)” respectively. The Four-
Phase Method, while studied in this dissertation research, has not been implemented as on

77
January 2018. A bullet-point summary of the development effort attributable to the author is
provided below:
1. Object-oriented interface: Wrote several base-classes for the wrappers that link
Radiance-binaries to Grasshopper3D and Rhino.
2. Wrapper-classes for Radiance binaries: Incorporated the functionality for Radiance
binaries like rcontrib, rfluxmtx, dctimestep, rmtxop and pcomb.
3. Script for improved Daylight Coefficient Method: Wrote the primary script that was
eventually condensed into a series of classes and functions in the deployed version of the
library. Also tested and validated several examples against raytracing benchmarks
[Chapter 3].
4. Script for Three-Phase Method and Five-Phase Method: Wrote the primary script.
Added functionality for cumulative addition of matrix results [Chapter 6].
5. Image-based daylight coefficients: Wrote the primary script. Implemented cross-
platform changes so that the script could be run seamlessly on Windows ® and Unix-
based systems [Chapter 4].
6. Documentation and API: Created Python-based docstrings for several classes that form
the automated API documentation for the library.
7. Cross-platform testing: Tested the library on Linux Ubuntu, Redhat, FreeBSD and Mac
OS X systems to ensure cross-platform conformity.
A question-answer based forum to support the potential users of Honeybee[+] was started by
Roudsari and curated by him, the author and other domain-specialists on the forum (Ladybug
Tools 2018b). Since its release, Honeybee[+] has been employed in two new projects that aim to
facilitate web-based simulations and docker-based simulations (Dao 2017; Ladybug Tools
2017a). The development work on Honeybee[+] was also recently acknowledged in a journal
paper by scientists at LBNL and Greg Ward. LBNL and Ward are the principal funding agency
and the author of Radiance respectively (Lee et al. 2018).

78
Chapter 8. Discussion and Conclusion

As articulated by its title, this dissertation research was conceptualized to study parametric
strategies for efficient annual analysis of daylight in buildings. The body of work carried out during
the course of this research can be summarized through the following points:
1. A comprehensive review of the current state of science with respect to parametric
daylighting simulations.
2. A series of empirical studies that evaluated means of optimizing the precision and runtime
of parametric daylighting simulations.
3. Open source software development to initially facilitate the empirical studies mentioned in
#2 and to eventually assimilate the functionality and learnings from the research to
mainstream daylighting simulation software.
The following sections discuss the findings from the numerical studies conducted research in the
context of certain recurring themes in this dissertation.

8.1 Impact of discretizing skies to a higher degree of granularity


It was hypothesized in Chapters 3 and 4 that a higher level of sky discretization will lead to more
accurate simulations. However this assumption was only partially realized, only in the case of the
image-based study in Chapter 4. For illuminance-based simulations, the precision of the
simulation was found to be heavily dependent on the direct-sun part of the simulation. Of the six
improved approaches with increasing order of sky discretization that were evaluated in the study,
no clear trends were attributable to sky discretization. Peak Signal to Noise Ratio (PSNR), the
distortion metric employed for the study, was consistently measured at a value of 50dB for all the
simulation methods. PSNR in the vicinity of 50dB correlated with numerical conformity between
the results from the different approaches.
Image-based simulations were found to be more affected by sky discretization, especially in cases
where the sun and the circumsolar region was in the field of view. Simulations with highly
discretized skies registered lower RMSE values in comparison with benchmark simulation.
However, this too was not a definitive trend. For example, in the case of south-facing views for
Fargo and Pittsburgh the RMSE values for the MF4 and MF5 solutions were identical. Considering
that the difference in computational effort between the two simulations is approximately 31%, in
locations with more overcast sky conditions it would be more appropriate to select a lower level
of sky discretization for the simulations.

8.2 Feasibility of multi-phase methods for complex models


The view-matrix subdivisions studied in Chapter 5 are a common element in Three-Phase, Four-
Phase, Five-Phase and Six-Phase simulations. The comparative study in Chapter 5 was able to
discern that a simple, un-divided view matrix was prone to errors, irrespective of the type of
simulation. Conversely, a very high level of view-matrix subdivision yield limited improvements in
the precision of the simulations. The model with the most subdivisions contained 24 individual
view matrices and when compared with the models with 4 or 8 subdivisions, yield very minor
improvements in precision.
The distortions encountered in the simulations in this study was higher than that observed for the
daylight coefficient study in Chapter 3.

79
8.3 The trade-off between computational efficiency and complexity of the
Façade matrix
The study in Chapter 6 demonstrated definitive advantages in employing the Façade matrix to
study simulations involving non-coplanar shading. However, the biggest drawback of the Four-
Phase Method, at present, is the complexity in setting up such a simulation and knowing the
correct use-cases.
For single simulations, the daylight coefficient approach was found to be 26% more efficient.
However, for studies involving parametric analysis, the ability to repeat a simulation by only
recalculating certain parts of it implies that the façade matrix will be more efficient in such
purposes.

8.4 Limitations
The following are the major limitations of this research:
1. Weather data for solstice hours: The weather data for the majority of the simulations
conducted for this research were based on a subset of TMY3 weather data for six locations
in the United States. The times and hours chosen for the simulations were based on
solstice and equinox hours. This selection, however, was found not to be representative
of the general weather conditions of that location. This was especially true in the case of
image-based simulations in Chapter 4.

2. Absence of complex models for simulations in Chapters 3 - 4: Simulations in Chapters 3 -


4 were essentially pilot studies meant to evaluate new methods of analyzing daylight. So,
the focus of these studies was primarily on identifying differences in the methods,
themselves, rather than any specific trends from the results generated through them.

3. Lack of direct-sun component in simulations for Chapter 5: The intention of choosing a


north-facing room was to focus purely on the view-matrix aspect of the simulations. So,
the effect of direct solar component was minimized by choosing a north facing room.
Simulations focusing on the direct-solar component typically tend to incorporate additional
steps for improving the precision of the direct sun-component.

4. Reliance on a non-BSDF approach for incorporating complex fenestrations in Chapter 6:


The study in Chapter 6 evaluated the possible advantages of employing the Façade-matrix
for parameterizing the analysis of external shading systems. An assumption was made
that such shades are represented through their native geometry in Radiance simulations.
However, often the shading systems are pre-processed for a simulation by calculating
their Tensor-Tree distributions in advance and incorporating those values into the scene.
Following this methodology in the simulation can reduce the advantages offered by the
Façade-Matrix method over the conventional Daylight Coefficient Method.

8.5 Future work


As summarized in Chapter 7, the research on the numerical studies conducted in this research
was paralleled by a complementary software development effort. Beyond the purposes of this
dissertation, that open source software is expected to be continually used by daylighting
designers and architects. Since there exists a gap between the concepts discussed in this

80
dissertation and the functionality present in that software, further development has been planned
to upgrade the features in the software library to enable more advanced simulations like the ones
discussed in Chapter 7.
As discussed in the previous section, the weather data set chosen for this research was not
entirely representative of the climatic conditions of some of the weather stations for which the
data was recorded. A more comprehensive analysis, considering performance across an entire
year, and at different locations, is recommended to corroborate the findings from the image-based
study in Chapter 4.

81
Appendix A. Weather data for simulations

A. 1 Locations
The daylighting simulations conducted in this research rely on Typical Meteorological Year (TMY)
data for calculating the luminous intensity of the sun and sky (National Climatic Center 1981;
Wilcox and Marion 2008). The data was obtained from the EnergyPlus website hosted by the US
Department of Energy and was in the TMY3-EPW format (US DOE 2012). The subset of TMY
data that is specifically relevant to daylighting simulations includes the following:
1. Geographical information such as longitude, latitude, meridian and daylight savings.
2. Time-series values of Direct-Normal Radiation in Wh/m2.
3. Time-series values of Diffuse-Horizontal Radiation in Wh/m2.
The studies conducted in Chapters 3 - 5 utilized TMY data for six cities in the United States. Those
six cities are Seattle (WA), Fargo (ND), Pittsburgh (PA), Denver (CO), Birmingham (AL) and
Phoenix (AZ). Their approximate locations on the United States map are plotted in Figure 82.and
the longitude and latitude of each city are listed in Table 21. The paired locations of Seattle-Fargo,
Denver-Pittsburgh and Phoenix-Birmingham were chosen on the basis of similar latitudes but
varying sky conditions. Figure 83 shows the plots of direct-normal radiation values overlaid on
sunpath diagrams for each location. The primary goal of choosing multiple paired locations was
to identify any specific trends in precision of the newer simulations that might be predisposed to
a certain type of geographical location or climatic conditions.

Figure 82. The six cities whose TMY data was utilized for simulations in Chapters 3 -5.

82
Table 21. The latitude and longitude of the six cities used for the simulations. Positive signs are assigned for
latitudes north of the equator and longitudes to the west of prime meridian.

Figure 83. Plot of Direct-Normal Radiation on Sun-Path diagrams for the six cities. A value of zero, indicated
by blue in the figure, implies an overcast sky for that particular hour.

83
A. 2 Dates and times for truth-simulations and benchmark studies
The comparative analyses performed in the studies detailed in Chapters 3 - 5 involve comparisons
between results generated for individual points in time. Since a TMY dataset contains 8760 (365
x 24) hours of data, it was impractical to perform the simulations for every hour of the year. Even
considering just the standard occupancy of 10 hours a day would have resulted in 3650
simulations for every inter-method or inter-model comparison.
Therefore, for the purposes of comparative analysis, a subset of the TMY data set for each
location was considered instead. The dates, selected on basis of equinox and solstice for 2017,
were: March 20th, June 21st, September 22nd and December 21st. The rationale for this selection
was to enable the testing of the different methods across a variety of sun and sky conditions. For
each of the dates, the simulations were performed for hourly timesteps starting from 8:30 AM to
5:30 PM (both inclusive). The solar altitudes corresponding to these dates and hours are plotted
chronologically in Figure 84. Assuming clear sky conditions, solar altitude is inversely related to
the amount of solar penetration inside a space for south-facing facades. The figure shows that
the selected dates and times had the potential to allow for a wide variety of direct solar insolation
in the space.

Figure 84. Solar altitudes for 8:30 to 17:30hrs for equinox and solstice dates plotted for the six locations.

84
The direct-normal and diffuse horizontal radiations for the four dates are listed in Table 22 - Table
25. The names of the six locations, which are the same as discussed in the previous section,
have been shortened to their first three letters. “Bir” implies Birmingham, “Pho” implies Phoenix
and so on. The direct-normal and diffuse-horizontal radiation, expressed in Wh/m2 in the tables,
together determine the sky condition as per the Perez Sky Model. For example, a very low value
of direct-normal radiation and high-value of diffuse horizontal radiation will result in a bright
overcast sky. The sky will be completely dark if both values are zero.
The left-most column in each of the tables below indicates the serial numbers of the hours. This
sequence of serial numbers for particular hours has been followed for all the comparative plots in
this Chapters 3-5. Figure 85 shows one such plot.
Table 22. Direct-Normal and Diffuse-Horizontal radiation values for March 20th.

Direct Normal Radiation Diffuse Horizontal Radiation


# Time
Bir Den Far Pho Pit Sea Bir Den Far Pho Pit Sea
0 8.5 8 566 716 742 51 3 169 146 241 324 214 138
1 9.5 12 640 816 802 13 59 229 194 233 352 213 116
2 10.5 14 718 854 824 7 216 270 184 211 337 207 267
3 11.5 268 419 868 829 4 490 356 161 176 295 173 209
4 12.5 439 952 866 827 2 398 292 133 133 242 122 282
5 13.5 428 878 853 816 6 169 268 115 85 180 198 315
6 14.5 623 987 801 789 24 498 154 60 47 104 227 200
7 15.5 684 934 633 699 55 404 94 50 47 53 199 149
8 16.5 453 825 421 383 31 129 93 36 46 89 117 79
9 17.5 151 543 112 45 8 58 50 21 32 53 35 46

Table 23. Direct-Normal and Diffuse-Horizontal radiation values for June 21st.

Direct Normal Radiation Diffuse Horizontal Radiation


# Time
Bir Den Far Pho Pit Sea Bir Den Far Pho Pit Sea
10 8.5 237 755 31 850 644 593 287 119 241 285 253 196
11 9.5 368 669 33 892 683 560 319 159 285 275 254 263
12 10.5 431 733 34 906 694 600 320 207 315 244 244 278
13 11.5 477 805 4 913 689 567 307 180 163 206 226 310
14 12.5 548 276 315 906 683 595 264 460 345 157 199 289
15 13.5 576 298 315 894 660 508 230 298 317 109 171 372
16 14.5 586 137 156 866 620 358 194 317 308 59 145 289
17 15.5 544 217 289 820 528 495 172 254 232 15 140 257
18 16.5 374 367 111 688 403 315 175 144 206 10 135 211
19 17.5 178 0 395 452 224 210 130 117 84 27 115 171

Table 24. Direct-Normal and Diffuse-Horizontal radiation values for September 22nd.

Direct Normal Radiation Diffuse Horizontal Radiation


# Time
Bir Den Far Pho Pit Sea Bir Den Far Pho Pit Sea
20 8.5 623 616 379 514 320 86 153 172 198 305 214 176
21 9.5 744 474 21 665 11 427 154 239 200 311 203 200
22 10.5 801 344 23 717 755 486 149 322 235 297 182 218
23 11.5 822 751 24 671 724 742 142 194 253 299 168 161
24 12.5 833 647 24 589 506 797 131 228 251 305 207 121
25 13.5 813 125 23 514 217 805 123 330 231 292 281 112
26 14.5 757 12 19 400 177 763 118 196 194 272 248 102
27 15.5 629 0 74 483 220 683 116 80 175 83 171 99
28 16.5 362 0 239 87 155 466 114 38 84 159 114 100
29 17.5 17 0 20 66 3 0 47 5 39 19 27 0

85
Table 25. Direct-Normal and Diffuse-Horizontal radiation values for December 21st.

Direct Normal Radiation Diffuse Horizontal Radiation


# Time
Bir Den Far Pho Pit Sea Bir Den Far Pho Pit Sea
30 8.5 6 634 0 215 1 286 73 32 52 155 20 21
31 9.5 3 807 80 37 2 383 60 44 103 182 39 43
32 10.5 8 881 34 67 4 247 162 53 120 230 77 120
33 11.5 8 910 7 38 6 615 183 57 105 244 116 100
34 12.5 8 908 43 246 8 368 180 57 140 251 143 94
35 13.5 8 876 149 69 54 107 159 52 123 238 166 163
36 14.5 49 795 251 118 171 48 165 43 70 134 117 111
37 15.5 5 587 63 57 209 141 69 32 38 84 54 52
38 16.5 3 87 0 1 0 53 18 9 0 29 2 1
39 17.5 0 0 0 0 0 0 0 0 0 0 0 0

Figure 85. A typical plot that is representative of the hourly plots used in Chapter 3 – Chapter 5. The 40 hours,
numbered from 0 to 39, in the x-axis correspond to the same serial # for hours in Table 22 - Table 25.

86
Appendix B. Detecting inter-method deviation in illuminance-based simulations

Three out of the four numerical studies conducted for this research are concerned with the
calculation of illuminance through daylighting simulations. Each of those studies identifies a
conventional simulation method as benchmark and then compares the results generated through
the benchmark method against newer simulation methods. The study discussed in Chapter 3
involves the comparison of results generated through conventional raytracing simulations against
those generated through the DDS Method, and the improved method proposed in this research.
The study in Chapter 5 compares benchmark Daylight Coefficient-based simulations against
simulations generated through the Three-Phase Method. Collectively, the intention of these
studies is to assess the newer simulation methods as a more accurate and/or faster means of
performing illuminance based simulations.
All of the aforementioned studies involve inter-method comparisons of several thousand point-in-
time illuminance values. Conventional data-visualizations or daylighting metrics are not suitable
for such a high volume of comparisons. Section B.1 reviews, and critiques, the currently
prevalent means of comparing illuminance values within the context of this research. Section B.2
proposes and demonstrates the applicability of image-quality metrics for comparing the deviation
of illuminance values.

B.1 Background
Comparison of computer-simulated illuminance values against real-world photosensor-derived
illuminance values has been a mainstay in research on lighting simulations for several decades
(Grynberg 1989; Mardaljevic 1995; Mardaljevic 2001; McNeil and Lee 2012; Reinhart and
Walkenhorst 2001; Wang et al. 2016; 2017). Typically, studies featuring empirical validation of
illuminances have involved less than ten photosensors, thereby making the process of
comparison between simulated and measured values relatively straightforward. Examples of the
sensor-configurations employed in two such studies are provided in Figure 86 and Figure 87.
Simulation-only studies with a large number of photosensors are typically employed for
performance evaluation of certain aspects of a building such as glazing, surface materials or
orientation. These studies rely on a combination of data-visualization and annual metrics for
assessing the magnitude and distribution of illuminance (DiLaura et al. 2011; Konis and Selkowitz
2017)7. Figure 88 shows one such example where annual metrics are visualized on a grid that
represents illuminance sensors on the workplane.
Research featuring inter-method illuminance comparisons have been rarely conducted in the
past. This can arguably be attributed to the fact that methods such as the Three-Phase, Four-
Phase, Five-Phase and Six-Phase Methods have only recently been introduced and their use in
research has mostly been confined to researchers who were responsible for developing these
methods. Two recent studies that involved inter-method illuminance comparisons employed
conventional metrics such as Useful Daylight Illuminance (UDI) and Daylight Autonomy (DA)
(Brembilla et al. 2015a; Brembilla et al. 2015b).

7
As there are far too many such studies, it is difficult to cite a few of them as authoritative references in this appendix. Therefore two
general references, that describe the trends in daylighting simulations, have been cited instead. For typical examples one may further
refer (Tzempelikos and Shen 2013), (Chen et al. 2014), (Wagdy and Fathy 2015) or (Manzan and Padovan 2015).

87
Figure 86. Plot of measured and simulated illuminance values from (Mardaljevic 1995). The study involved six
photosensors.

Figure 87. Location of the six photosensors used for measuring and simulating illuminance values in (Wang
et al. 2016).

88
Figure 88. Visualization of annual metrics on a grid where each square represents an illuminance sensor
(DiLaura et al. 2011). The plotted metrics, from left to right, are Daylight Factor, Daylight Autonomy and
Continuous Daylight Autonomy. The glazing is at the bottom.

One of these studies suggested the use of Total Annual Illuminance (TAI) as a means of
comparing the deviation in results generated through different simulation methods (Brembilla et
al. 2015b). TAI is the sum of all the illuminance values for the occupied hours across a year. The
authors rationalized their preference for TAI by acknowledging that conventional daylighting
metrics like UDI, DA and sDA were not suited for inter-method comparison as these methods bin
absolute values of illuminances into percentages. However, TAI too is an imperfect metric as any
cumulative method is prone to average out positive and negative deviations in results over a
prolonged interval. This issue is explained through Figure 89 and Figure 90.

Figure 89. A comparison of point-in-time illuminance values calculated through seven simulation methods.
The benchmark results are on upper-left. The “DDS” method (second from the left on top row) indicates an
overestimation of illuminance.

89
Figure 90. A comparison of illuminance values for the same model as considered in Figure 89. In this instance
the ‘DDS” method underestimates the illuminance.

The images show that the “DDS” method overestimates illuminance in the first instance and
underestimates illuminance in the subsequent instance. Adding the illuminances for all the
simulation types in both instances essentially conceals the errors encountered in the DDS method
and puts it on par with the other simulation methods. It follows that any cumulative metric which
ignores the point-in-time granularity of annual simulations is likely to misrepresent the precision
of the simulation methods.
In the current research, an initial attempt at identifying illuminance-based deviations between
different simulation methods entailed binning illuminance values into three percentile ranges and
then comparing the means of each of these percentile bins. This approach is described through
Figure 91 and Figure 92 and was employed in a conference paper that presented the preliminary
results from the study conducted in Chapter 3 (Subramaniam and Mistrick 2017). When compared
with the traditional metrics discussed in the previous section, this approach was found to be better
suited for identifying deviations between different methods. However, as indicated by Figure 92,
the binning methodology employed in this case is not capable of numerically discriminating
between any pattern-based shifts in illuminance results. One such scenario is illustrated in Figure
93.
Since none of the numerical comparison methods discussed above were found to be appropriate
for inter-method comparisons, the applicability of non-traditional metrics for comparisons was
explored. The following section details the pattern-recognition and image-quality metrics that were
eventually chosen for comparing deviations in illuminance results.

90
Figure 91. Comparison between point-in-time illuminance values generated through a benchmark raytrace
simulation and the DDS method simulation. Simulation type is mentioned on top of each image. L, M and H
correspond to Low (0-25), Medium (26-75) and High (76-100) percentile ranges. The numbers within square-
brackets indicate the mean illuminance within a particular range, followed by the standard deviation. Note that
the above plots are a subset of the plots shown in Figure 91.

Figure 92. The mean of three percentile bins described in Figure 91 are plotted in the above plots for a period
of 40 hours. The underestimation by the DDS method in the 76th-100th percentile bin is evident in the image on
the top.

91
Figure 93. A comparison between two calculation methods using binned percentile ranges. The numerical
comparison, as shown by the means above the plots, misleadingly indicates similarity between the two results.

B.2 Employing image-quality metrics for comparing illuminance values.


For a new simulation method to be considered accurate, it should generate illuminance values
that are numerically as close as possible to that generated through the benchmark simulation.
Additionally, this numerical similarity should extend to individual photosensor measurements as
well. These two aspects of comparison are easily discernable through visual inspection of
illuminance grids as shown in Figure 89 - Figure 93. As evidenced by the discussion in the
previous section, conventional metrics are not suited for numerically expressing such a
comparison. So, for the purposes of such comparisons, image-quality metrics were considered.
Image-quality metrics were initially developed to provide quantitative assessments of the outputs
of image-compression and image-format-conversion software. Such software typically accept an
uncompressed image as input and then generate compressed images in various formats. In
recent years the scope of such metrics has expanded to several applications such as digital signal
processing, pattern-recognition, computer vision, medical imaging and statistical learning. The
wide applicability of these metrics can be attributed to the fact that they are essentially means of
comparing 2-Dimensional or 3-Dimensional data-structures that are stored in matrix-format (Bovik
2010; Wang et al. 2004).
The three metrics considered for illuminance-comparison in this research are Mean Squared Error
(MSE), Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). The
data-structures being compared by each of these metrics are the two-dimensional illuminance
arrays whose dimensions are as per the photosensor grid used for simulations. So, in the case of
the simulations performed in Chapter 3, the dimensions of the array were (10, 10). The rationale
for the choice of these metrics is explained below:
MSE: This metric quantifies the global difference between the benchmark data and test data.
Major deviations in illuminance can be discerned through the use of MSE alone. However, MSE
ignores localized errors and therefore is not suitable for detecting any pattern-based shifts.
Additionally, MSE is not suitable for comparing multiple instances of deviation as its value in each
instance depends on the minima and maxima of data being considered (Dosselmann and Yang
2005; Veldhuizen 2010).

92
PSNR: PSNR is essentially a metric that is derived from MSE. It scales the MSE value as per the
maximum value of the data within the specified matrix. It is expressed in a logarithmic scale to
the base of 10. When performing inter-method comparisons, a higher value of PSNR indicates
greater similarity with the benchmark dataset. A value of infinity indicates a perfect match with the
benchmark dataset (Huynh-Thu and Ghanbari 2008).
SSIM: SSIM, as the name suggests, is a measure of structural similarity. This metric was
employed to identify any noise-based deviations in illuminance values across different simulation
methods. The range of SSIM is between 0-1 where 1 indicates a perfect match. Since the
parameters for all the simulations in this research were iteratively improved till convergence,
noise-based errors were not encountered. All the SSIM values calculated for the simulation results
were above 0.95.
Based on the above factors, PSNR was chosen as the metric for evaluating inter-method
illuminance deviations. The suitability of employing this metric is explained through a series of
plots. Figure 94 contains the plot of PNSR values for an inter-method illuminance comparison
from the study in Chapter 3. Figure 95 – Figure 104 show the raw-data used to derive the PNSR
values for hours highlighted by arrows. In the case of low PNSR values, the erroneous data are
highlighted with ellipses.

Figure 94. PSNR values for inter-method illuminance comparison for Birmingham, AL. The numbers below the
plot, which are highlighted through arrows, pertain to hours listed in Table 22 - Table 25.

93
Figure 95. Illuminance data pertaining to hour #1 from Figure 94.

Figure 96. Illuminance data pertaining to hour #4 from Figure 94.

94
Figure 97. Illuminance data pertaining to hour #7 from Figure 94.

Figure 98. Illuminance data pertaining to hour #14 from Figure 94.

95
Figure 99. Illuminance data pertaining to hour #21 from Figure 94.

Figure 100. Illuminance data pertaining to hour #25 from Figure 94.

96
Figure 101. Illuminance data pertaining to hour #27 from Figure 94.

Figure 102. Illuminance data pertaining to hour #30 from Figure 94.

97
Figure 103. Illuminance data pertaining to hour #36 from Figure 94.

Figure 104. Illuminance data pertaining to hour #39 from Figure 94.

In the above figures, the inter-method deviation in PSNR values, when present, very distinctly
identifies the imprecise method. For the study conducted in Chapter 5, PSNR was used to discern
the improvement in precision based on the progressive-complexity of the modelling approach
used for a simulation. So, for several simulation timesteps, PSNR values spanned a wide range

98
of values. The following section provides an example of the variation of PSNR values as a function
of absolute illuminances in such cases.

B.3 Relating variation in PSNR to absolute values of illuminance


Figure 105 shows a plot of PSNR values that compares the precision of illuminance results
obtained through Three-Phase Method simulations for different view-factor configurations, when
compared with a reference Daylight Coefficient Method simulation. For the hour highlighted
through red dotted line in the figure, it follows that configurations “G2VH” and “G4VH” are
numerically closest to the reference dataset. “G2V” is less precise than “G2VH” and “G4VH” while
“G” is the least precise. Figure 106, provides a comparison of absolute values of illuminance for
the same hour. The plots of illuminance values conform to the inferences made through the PSNR
values in Figure 105.
Comparing absolute values of illuminance for each dataset, especially for hypothetical models
like the ones considered in this research, does not serve any useful purpose in the context of this
research as the objective of the simulations was solely to identify the precision of individual
approaches. As stated previously, the objective of employing a metric like PSNR was to compare
the precision of multitudes of datasets like the ones shown in Figure 106.

Figure 105. PSNR values for simulations with different view-factor configurations. The dataset contains 40
observations corresponding to 40 hours. The reference dataset is a Daylight Coefficient simulation. The terms
“GD”, “G” etc. relate to different view factor configurations. The individual datasets for hour 16, which is
highlighted through a red dotted line, are shown in Figure 106.

99
Figure 106. Illuminance plots corresponding to observation 16 for three of the seven configurations plotted in
Figure 105. The corresponding PSNR values are mentioned on the left side of each plot. The letters in
parenthesis below the PSNR values identify the configuration. The difference between the Reference dataset
and “G” is easily discernable. The differences between the Reference dataset and other two datasets are
highlighted through dashed rectangles. It should be noted that PSNR is expressed in a logarithmic scale to
the base of 10. So, the numerical differences of individual grid-point illuminances between “G” and “G2VH”
are to the order of more than 100 lux.

100
Appendix C. Additional data visualizations for Chapter 3

The PSNR plots for the study involving the room with skylight, which is shown in Figure 24 and
discussed in Section 3.4.2 are provided below. The name of the location is mentioned in the title
of each plot.

101
102
103
Appendix D. Additional data visualizations for Chapter 4
The entire set of DGP plots and Sum of Glare sources plots for all the locations discussed in
Chapter 4 are provided below:

104
105
106
107
108
109
110
111
112
113
114
115
Appendix E. Additional data visualizations for Chapter 5

PSNR plots for the model with glazing only.

116
PSNR plots for the model with glazing and overhang.

117
PSNR plots for the model with glazing, overhang and external geometry.

118
PSNR plots for the model with glazing and external geometry.

119
PSNR plots for the model with glazing and mullions.

120
PSNR plots for the model with glazing, mullions and overhang.

121
PSNR plots for the model with glazing, mullions, overhang and external geometry.

122
PSNR plots for the model with glazing, mullions and external geometry.

123
124
125
126
127
BIBLIOGRAPHY

Aries M, Aarts M, van Hoof J. 2015. Daylight and health: A review of the evidence and
consequences for the built environment. Lighting Research and Technology 47(1):6-27
Ashdown I. 2016. Climate-based daylight modeling: From Theory to practice [Internet].
Acessed: 12/01/2017. Available from: agi32.com/blog/2016/03/26/climate-based-
daylight-modeling/
Ashdown I, Jackson C, Spahn J, Saemisch T. 2017. LICASO and DAYSIM [Internet]. Acessed:
12/01/2017. Available from: agi32.com/blog/2017/02/11/licaso-and-daysim/
Autodesk Inc. 2010. Ecotect Analysis [Internet]. Acessed: 01/01/2018. Available from:
autodesk.com/education/free-software/ecotect-analysis
Begemann SHA, van den Beld GJ, Tenner AD. 1997. Daylight, artificial light and people in an
office environment, overview of visual and biological responses. International Journal of
Industrial Ergonomics 20(3):231-239
Bellia L, Cesarano A, Iuliano GF, Spada G. 2008. Daylight glare: a review of discomfort indexes.
Napoli, Italy. Università degli Studi di Napoli Federico II.
Blackwell HR, DiLaura DL. 1973. IERI. Application Procedures for Evaluation of Veiling
Reflections in Terms of ESI: II. Gonio Data for the Standard Pencil Task. Journal of the
Illuminating Engineering Society 2(3):254-283
Blackwell HR, DiLaura DL, Helms RN. 1973a. IERI. Application Procedures for Evaluation of
Veiling Reflections in Terms of ESI: IV. Final Validation of Measurement and
Predetermination Methods for Luminaire Installations. Journal of the Illuminating
Engineering Society 2(3):299-327
Blackwell HR, Helms RN, DiLaura DL. 1973b. IERI. Application Procedures for Evaluation of
Veiling Reflections in Terms of ESI: III Validation of a Predetermination Method for
Luminaire Installations. Journal of the Illuminating Engineering Society 2(3):284-298
Bleicher T. 2008. su2rad - Radiance exporter for SketchUp. 7th International Radiance
Workshop. Fribourg, Germany.
Bodart M, De Herde A. 2002. Global energy savings in offices buildings by the use of
daylighting. Energy and Buildings 34(5):421-429
Boubekri M, Cheung IN, Reid KJ, Wang C-H, Zee PC. 2014. Impact of windows and daylight
exposure on overall health and sleep quality of office workers: a case-control pilot study.
Journal of clinical sleep medicine: JCSM: official publication of the American Academy of
Sleep Medicine 10(6):603
Bourgeois D, Reinhart CF, Ward G. 2008. Standard daylight coefficient model for dynamic
daylighting simulations. Building Research & Information 36(1):68-82
Bovik AC. 2010. Handbook of image and video processing. Massachusetts, USA: Academic
press. ISBN 0080533612
Boyce P, Hunter C, Howlett O. 2003. The benefits of daylight through windows. Rensselaer
Polytechnic Institute.

128
Boyce PR. 2014. Human factors in lighting. Florida, USA: CRC Press. ISBN 0203426347
Boynton RM. 1996. History and current status of a physiologically based system of photometry
and colorimetry. Journal of the Optical Society of America A 13(8):1609-1621
Brembilla E. 2016. Survey on Climate-Based Daylight Modelling workflows. 15th International
Radiance Workshop. Padova, Italy
Brembilla E, Mardaljevic J, Anselmo F. 2015a. The effect of the analysis grid on daylight
simulations with climate-based daylight modelling. The 28th Session of the International
Commission on Illumination. Manchester, UK.
Brembilla E, Mardaljevic J, Hopfe CJ. 2015b. Sensitivity Analysis studying the impact of
reflectance values assigned in Climate-Based Daylight Modelling. The 14th International
Conference of the International Building Performance Simulation Association.
Hyderabad, India
Bryan HJ. 1980. A Simplified Procedure for Calculating the Effects of Daylight from Clear Skies.
Journal of the Illuminating Engineering Society 9(3):142-151
Bryan HJ, Clear RD. 1980. A Procedure for Calculating Interior Daylight Ilumination. 5th
National Passive Solar Conference. Massachusetts, USA
Cantin F, Dubois M. 2011. Daylighting metrics based on illuminance, distribution, glare and
directivity. Lighting Research and Technology 43(3):291-307
Casey CA, Mistrick RG. 2015. Simulation Tools for Architectural Daylighting and Integrated
Controls (STADIC) - Utilities. 14th International Radiance Workshop. Philadelphia, USA.
Chen Y, Liu J, Pei J, Cao X, Chen Q, Jiang Y. 2014. Experimental and simulation study on the
performance of daylighting in an industrial building and its energy saving potential.
Energy and Buildings 73(0):184-191
Clear R. 2013. Discomfort glare: What do we actually know? Lighting Research and Technology
45(2):141-158
Dao A. 2017. Radhoneywhale [Internet]. Acessed: 01/01/2018. Available from:
hub.docker.com/r/antoinedao/radhoneywhale/
Daysim. 2018. Daysim: Advanced daylight simulation software [Internet]. Acessed: 01/10/2018.
Available from: daysim.ning.com/
DiLaura D. 2006. A history of light and lighting. New York, USA: IES. ISBN 9780879952099
DiLaura DL. 1975. On the Computation of Equivalent Sphere Illumination. Journal of the
Illuminating Engineering Society 4(2):129-149
DiLaura DL. 1976. On the Computation of Visual Comfort Probability. Journal of the Illuminating
Engineering Society 5(4):207-217
DiLaura DL, Hauser GA. 1978. On Calculating the Effects of Daylighting in Interior Spaces.
Journal of the Illuminating Engineering Society 8(1):2-14
DiLaura DL, Mistrick RG, Houser KH, Steffy G. 2011. The lighting handbook: reference &
applications. New York, USA: IES. ISBN 0879952415, 9780879952419

129
DiLaura DL, Stannard SM. 1978. IES Transaction: An Instrument for the Measurement of
Equivalent Sphere Illumination. Journal of the Illuminating Engineering Society 7(3):183-
189
Dogan T. 2016. Archsim: Energy modeling tools for Grasshopper [Internet]. Acessed:
03/01/2018. Available from: archsim.com
Dosselmann R, Yang XD. 2005. Existing and emerging image quality metrics. Canadian
Conference on Electrical and Computer Engineering. Saskatchewan, Canada
IEEE:1906-1913
Estes Jr J, Schreppler S, Newsom T. 2004. Daylighting prediction software: comparative
analysis and application. Fourteenth Symposium on Improving Building Systems in Hot
and Humid Climates. Texas, USA.
Foliente GC. 2000. Developments in performance-based building codes and standards. Forest
Products Journal 50(7/8):12
Galasiu AD, Reinhart CF. 2008. Current daylighting design practice: a survey. Building
Research & Information 36(2):159-174
Geisler-Moroder D, Lee ES, Ward G. 2017. Validation of the Five-Phase Method for Simulating
Complex Fenestration Systems with Radiance against Field Measurements. 15th
International Conference of the International Building Performance Simulation
Association. California, USA
Gibson KS, Tyndall EPT. 1923. Visibility of radiant energy. Scientific Papers of the Bureau of
Standards. Washington, USA: United States Department of Commerce:131-191
GitHub. 2017a. Honeybee - List of contributors [Internet]. Acessed: 01/01/2018. Available from:
github.com/mostaphaRoudsari/honeybee/graphs/contributors
GitHub. 2017b. Honeybee[+] - list of contributors [Internet]. Acessed: 01/01/2018. Available
from: github.com/ladybug-tools/honeybee/graphs/contributors
GitHub. 2018. Ladybug-Tools/Honeybee- Contributions to master [Internet]. Acessed:
01/01/2018. Available from: github.com/ladybug-tools/honeybee/graphs/contributors
Goodman T. 2009. Measurement and specification of lighting: A look at the future. Lighting
Research and Technology 41(3):229-243
Grynberg A. 1989. Validation of Radiance. California, USA. LBNL.
Guglielmetti R, Macumber D, Long N. 2011. OpenStudio: An Open Source Integrated Analysis
Platform. 12th Conference of International Building Performance Simulation Association.
Sydney, Australia.
Gugliermetti F, Bisegna F. 2006. Daylighting with external shading devices: design and
simulation algorithms. Building and Environment 41(2):136-149
Hensen JL. 2011. Building performance simulation for design and operation. Oxford, USA:
Routledge. p. 235-276. ISBN 0415474140
Hien WN, Istiadji AD. 2003. Effects of external shading devices on daylighting and natural
ventilation. 11th Conference of International Building Performance Simulation
Association. Eindhoven, The Netherlands:475-482

130
Holzer D. 2016. Design exploration supported by digital tool ecologies. Automation in
Construction 72:3-8
Hore A, Ziou D. 2010. Image quality metrics: PSNR vs. SSIM. 2010 International Conference on
Pattern Recognition. Istanbul, Turkey IEEE:2366-2369
Howell JR. 1969. Application of Monte Carlo to Heat Transfer Problems. In: Irvine TF, Hartnett
JP, editors. Advances in Heat Transfer. New York, USA: Elsevier:1-54
Hraška J. 2011. Daylight requirements in sustainable building rating systems. Ingineria
iluminatului 13(2):5-11
Huynh-Thu Q, Ghanbari M. 2008. Scope of validity of PSNR in image/video quality assessment.
Electronics letters 44(13):800-801
IES. 2012. LM-83-12 IES Spatial Daylight Autonomy (sDA) and Annual Sunlight Exposure
(ASE). New York, USA: IES. ISBN 9780879952723
Ihm P, Nemri A, Krarti M. 2009. Estimation of lighting energy savings from daylighting. Building
and Environment 44:509-514
Inanici M. 2006. Evaluation of high dynamic range photography as a luminance data acquisition
system. Lighting Research and Technology 38(2):123-134
Inanici M. 2010. Evalution of High Dynamic Range Image-Based Sky Models in Lighting
Simulation. LEUKOS 7(2):69-84
Inanici M, Galvin J. 2004. Evaluation of High Dynamic Range Photography as a Luminance
Mapping Technique. California, USA. LBNL.
Iwata T, Kimura K-I, Shukuya M, Takano K. 1991. Discomfort caused by wide-source glare.
Energy and Buildings 15(3):391-398
Jacobs A. 2010. Understanding rtcontrib [Internet]. Acessed: 01/01/2018. Available from:
jaloxa.eu/resources/radiance/documentation/docs/rtcontrib_lesson.pdf
Jacobs A. 2012. Radiance Tutorial [Internet]. Available from:
jaloxa.eu/resources/radiance/documentation/docs/radiance_tutorial.pdf
Jacobs A. 2014. Radiance Cookbook [Internet]. Acessed: 01/01/2018. Available from:
jaloxa.eu/resources/radiance/documentation/docs/radiance_cookbook.pdf
Jakubiec JA, Reinhart CF. 2011. DIVA 2.0: Integrating daylight and thermal simulations using
Rhinoceros 3D, Daysim and EnergyPlus. The 12th International Conference of the
International Building Performance Simulation Association. Sydney, Australia.
Jeong Tai Kim, Kim G. 2010. Advanced External Shading Device to Maximize Visual and View
Performance. Indoor and Built Environment 19(1):65-72
Jones NL, Reinhart CF. 2017. Speedup Potential of Climate-Based Daylight Modelling on
GPUs. 15th International Conference of the International Building Performance
Simulation Association. California, USA
Jonsson JC, Rubin MD, Nilsson AM, Jonsson A, Roos A. 2009. Optical characterization of
fritted glass for architectural applications. Optical Materials 31(6):949-958

131
Kähler K. 1908. Flächenhelligkeit des Himmels und Beleuchtungsstärke in Räumen. Meteorol.
Zeitschrift 25(2):52-57
Kim B, Lee Y. 2017. Genetic algorithms for balancing multiple variables in design practice.
Advances in Computational Design 2(3):225-240
Kim G, Lim HS, Lim TS, Schaefer L, Kim JT. 2012. Comparative advantage of an exterior
shading device in thermal performance for residential buildings. Energy and Buildings
46(0):105-111
Kittler R, Kocifaj M, Darula S. 2011. Daylight science and daylighting technology. New York,
USA: Springer. ISBN 9781441988164
Konis K, Selkowitz S. 2017. Effective Daylighting with High-Performance Facades: Emerging
Design Practices. Switzerland: Springer. ISBN 3319394630
Kota S, Haberl JS. 2009. Historical Survey of Daylighting Calculations Methods and Their Use
in Energy Performance Simulations. Ninth International Conference for Enhanced
Building Operations. Texas, USA.
Kumaragurubaran V, Inanici M. 2013. Hdrscope: high dynamic range image processing toolkit
for lighting simulations and analysis. 13th Conference of International Building
Performance Simulation Association. Chambéry, France:25-28
Ladybug Tools. 2017a. Honeybee Server [Internet]. Acessed: 01/01/2018. Available from:
github.com/ladybug-tools/honeybee-server-daylight
Ladybug Tools. 2017b. Honeybee[+] 0.0.3 for Grasshopper and Dynamo Release [Internet].
Acessed: 01/01/2018. Available from: discourse.ladybug.tools/t/honeybee-0-0-3-for-
grasshopper-and-dynamo-release/2024
Ladybug Tools. 2017c. Ladybug Tools website [Internet]. Acessed: 01/01/2018. Available from:
ladybug.tools/about.html
Ladybug Tools. 2018a. Honeybee documentation [Internet]. Acessed: 01/01/2018. Available
from: ladybug.tools/apidoc/honeybee/
Ladybug Tools. 2018b. Latest Honeybee-Plus Topics [Internet]. Acessed: 01/15/2018. Available
from: discourse.ladybug.tools/c/grasshopper/honeybee-plus
LBNL. 1997. The RADIANCE Lighting Simulation and Rendering System [Internet]. Acessed:
02/11/2017. Available from: radsite.lbl.gov/radiance/framew.html
LBNL. 2016. Genbsdf [Internet]. Acessed: 01/01/2018. Available from: radiance-
online.org/learning/documentation/manual-pages/pdfs/genBSDF.pdf
Lee ES, Geisler-Moroder D, Ward G. 2018. Modeling the direct sun component in buildings
using matrix algebraic approaches: Methods and validation. Solar Energy 160:380-395
Lighting Analysts Inc. 2017. Licaso: Annual daylight simulation software for AGI32 and
ElumTools [Internet]. Acessed: 12/01/2017. Available from: licaso.com
Manzan M, Padovan R. 2015. Multi-criteria energy and daylighting optimization for an office with
fixed and moveable shading devices. Advances in Building Energy Research 9(2):238-
252

132
Mardaljevic J. 1995. Validation of a lighting simulation program under real sky conditions.
Lighting Research and Technology 27(4):181-188
Mardaljevic J. 1999. Daylight simulation: validation, sky models and daylight coefficients.
[Leicester, UK]. De Montfort University
Mardaljevic J. 2000. Simulation of annual daylighting profiles for internal illuminance. Lighting
Research and Technology 32(3):111-118
Mardaljevic J. 2001. The BRE-IDMP dataset: a new benchmark for the validation of illuminance
prediction techniques. Lighting Research and Technology 33(2):117-134
Mardaljevic J, Andersen M, Roy N, Christoffersen J. 2012. Daylighting Metrics: Is There a
Relation between Useful Daylight Illuminance and Daylight Glare Probability? Building
Simulation and Optimization Conference (BSO12) Loughborough, UK.
Mardaljevic J, Heschong L, Lee E. 2009. Daylight metrics and energy savings. Lighting
Research and Technology 41(3):261-283
Mardaljevic J, Roy N. 2016. The sunlight beam index. Lighting Research and Technology
48(1):55-69
Martin RC. 2002. Agile software development: principles, patterns, and practices. New Jersey,
USA: Prentice Hall. ISBN 0135974445
McNeel Europe S.L. 2018. Ladybug Tools Downloads [Internet]. Acessed: 01/01/2018.
Available from: food4rhino.com/app/ladybug-tools
McNeil A. 2013a. The 5-phase method. 12th International Radiance Workshop. Colorado, USA.
McNeil A. 2013b. The Five-Phase Method for Simulating Complex Fenestration with Radiance.
LBNL.
McNeil A. 2013c. The Three-Phase Method for Simulating Complex Fenestration with Radiance.
LBNL.
McNeil A. 2014. BSDFs, Matrices and Phases. 13th International Radiance Workshop. London,
UK.
McNeil A. 2015. genBSDF Tutorial. LBNL.
McNeil A, Jonsson CJ, Appelfeld D, Ward G, Lee ES. 2013. A validation of a ray-tracing tool
used to generate bi-directional scattering distribution functions for complex fenestration
systems. Solar Energy 98(C)
McNeil A, Lee E. 2013. A validation of the Radiance three-phase simulation method for
modelling annual daylight performance of optically complex fenestration systems.
Journal of Building Performance Simulation 6(1):24-37
McNeil A, Lee ES. 2012. A validation of the Radiance three-phase simulation method for
modeling annual daylight performance of optically-complex fenestration systems.
Journal of Building Performance Simulation April 2012
Mischler G. 2004. Lighting Simulation with Rayfront. 3rd International Radiance Workshop.
Fribourg, Switzerland

133
Mistrick RG. 2013. DAYSIM Development. 12th International Radiance Workshop. Colorado,
USA
Mitchell R, Christian K, Joseph HK, Michael DR, Dariush KA, Charlie H, Tiefeng Y, Dragan CC.
2008. WINDOW 6.2/THERM 6.2 Research Version User Manual. California, USA. LBNL.
Modest MF. 1982. A general model for the calculation of daylighting in interior spaces. Energy
and Buildings 5(1):69-79
Modest MF. 1983. Daylighting Calculations for Non-Rectangular Interior Spaces with Shading
Devices. Journal of the Illuminating Engineering Society 12(4):226-241
Molina G. 2015. Groundhog, A sketchup plugin for Radiance analysis. 14th International
Radiance Workshop. Philadelphia, PA, USA
Molina G, Bustamante W, Rao J, Fazio P, Vera S. 2015. Evaluation of radiance's genBSDF
capability to assess solar bidirectional properties of complex fenestration systems.
Journal of Building Performance Simulation 8(4):216-225
Moon P, Spencer DE. 1942. Illumination from a non-uniform sky. Illuminating Engineering
37(10):707-726
Nabil A, Mardaljevic J. 2005. Useful daylight illuminance: a new paradigm for assessing daylight
in buildings. Lighting Research and Technology 37(1):41-57
Nabil A, Mardaljevic J. 2006. Useful daylight illuminances: A replacement for daylight factors.
Energy and Buildings 38(7):905-913
National Climatic Center. 1981. TD-9734:Typical Meteorological Year User's Manual. Asheville,
North Carolina, USA. National Climatic Center.
Østergård T, Jensen RL, Maagaard SE. 2016. Building simulations supporting decision making
in early design–A review. Renewable and Sustainable Energy Reviews 61:187-201
Oxman R. 2017. Thinking difference: Theories and models of parametric design thinking.
Design Studies 52:4-39
Perez R, Ineichen P, Seals R, Michalsky J, Stewart R. 1990. Modeling daylight availability and
irradiance components from direct and global irradiance. Solar energy 44(5):271-289
Perez R, Seals R, Michalsky J. 1993a. All-weather model for sky luminance distribution—
preliminary configuration and validation. Solar energy 50(3):235-245
Perez R, Seals R, Michalsky J. 1993b. To all-weather model for sky luminance distribution—
preliminary configuration and validation. Solar Energy 51(5):423
Phillips D. 2004. Daylighting: natural light in architecture. Oxford, UK: Routledge. ISBN
0750663235
Reinhard E, Heidrich W, Debevec P, Pattanaik S, Ward G, Myszkowski K. 2010. High dynamic
range imaging: acquisition, display, and image-based lighting. Massachusetts, USA:
Morgan Kaufmann. ISBN 0080957110
Reinhart CF. 2001. Daylight availability and manual lighting control in office buildings:
Simulation studies and analysis of measurement. [Stuttgart, Germany]. Fraunhofer-IRB-
Verlag

134
Reinhart CF. 2004. Key findings from a online survey on the use of daylight simulation
programs. ESIM 2004 Conference. Vancouver, Canada.
Reinhart CF. 2006. Tutorial on the use of daysim simulations for sustainable design [Internet].
2006
Reinhart CF. 2015. Opinion: Climate-based daylighting metrics in LEEDv4 – A fragile progress.
Lighting Research and Technology 47(4):388
Reinhart CF, Fitz A. 2006. Findings from a survey on the current use of daylight simulations in
building design. Energy and Buildings 38(7):824-835
Reinhart CF, Herkel S. 2000. The simulation of annual daylight illuminance distributions — a
state-of-the-art comparison of six RADIANCE-based methods. Energy and Buildings
32:167-187
Reinhart CF, Mardaljevic J, Rogers Z. 2006. Dynamic Daylight Performance Metrics for
Sustainable Building Design. LEUKOS 3(1):7
Reinhart CF, Walkenhorst O. 2001. Validation of dynamic RADIANCE-based daylight
simulations for a test office with external blinds. Energy & Buildings 33(7):683-697
Reinhart CF, Wienold J. 2011. The daylighting dashboard – A simulation-based design analysis
for daylit spaces. Building and Environment 46:386-396
Rogers Z. 2006. Daylighting metric development using daylight autonomy calculations in the
sensor placement optimization tool. Boulder, CO: Architectural Energy Corporation
Roudsari M, Pak M. 2014. Ladybug: A Parametric Environmental Plugin for Grasshopper to
Help Designers Create an Environmentally-Conscious Design. 13th Conference of
International Building Performance Simulation Association. Chambery, France.
Roudsari M, Subramaniam S. 2016. Automating radiance workflows using Python. 15th
International Radiance Workshop. Padova, Italy
Roy GG. 2000. A comparative study of lighting simulation packages suitable for use in
architectural design. School of Engineering, Murdoch University, Perth, Australia
Salus PH. 1994. A quarter century of UNIX. Massachusetts, USA: Addison-Wesley Reading
ISBN 0201547775
Saxena M, Ward G, Perry T, Heschong L, Higa R. 2010. Dynamic Radiance–Predicting annual
daylighting with variable fenestration optics using BSDFs. Proceedings of SimBuild:11-
13
Schramm W. 1901. Über die Verteilung des Lichtes in der Atmosphäre. [Kiel, Germany].
Christian-Albrechts-Universität zu Kiel
Smartsheet. 2018. Agile Project Management [Internet]. Acessed: 01/01/2018. Available from:
smartsheet.com/everything-you-need-to-know-about-agile-project-management
Southall R. 2016. VI-Suite, a building environmental performance addon for Blender [Internet].
Acessed: 01/01/2018. Available from: github.com/rgsouthall/vi-suite04
Southall R, Biljecki F. 2017. The VI-Suite: a set of environmental analysis tools with geospatial
data applications. Open Geospatial Data, Software and Standards 2(1):23

135
Subramaniam S, Mistrick R. 2017. A More Accurate Approach for Calculating Illuminance with
Daylight Coefficients. 2017 Annual IES Conference. Oregon, USA IES.
Thornton Tomasetti. 2017. Design Explorer [Internet]. Acessed: 01/01/2018. Available from: tt-
acm.github.io/DesignExplorer/
Touloupaki E, Theodosiou T. 2017. Performance Simulation Integrated in Parametric 3D
Modeling as a Method for Early Stage Design Optimization—A Review. Energies
10(5):637
Tregenza P. 1983. The Monte Carlo method in lighting calculations. Lighting Research &
Technology 15(4):163-170
Tregenza P, Mardaljevic J. 2018. Daylighting buildings: Standards and the needs of the
designer. Lighting Research & Technology 50(1):63-79
Tregenza P, Wilson M. 2013. Daylighting: architecture and lighting design. Oxford, UK:
Routledge. ISBN 1135882967
Tregenza PR. 1980. The daylight factor and actual illuminance ratios. Lighting Research and
Technology 12(2):64-68
Tregenza PR. 1987. Subdivision of the sky hemisphere for luminance measurements. Lighting
Research and Technology 19(1):13-14
Tregenza PR, Waters IM. 1983. Daylight coefficients. Lighting Research and Technology
15(2):65-71
Tzempelikos A, Shen H. 2013. Comparative control strategies for roller shades with respect to
daylighting and energy performance. Building and Environment 67(0):179-192
Ubbelohde MS, Humann C. 1998. Comparative Evaluation of Four Daylighting Software
Programs. ACEE Summer Study on Energy Efficiency in Buildings. CA, USA.
US DOE. 2012. Weather Data [Internet]. Acessed: 01/01/2018. Available from:
energyplus.net/weather
USGBC. 2014. LEED V4 for building design and construction. Washington, DC, USA: US Green
Building Council. ISBN 9781932444148
Veldhuizen T. 2010. Measures of image quality. CVonline: The Evolving, Distributed, Non-
Proprietary, On-Line Compendium of Computer Vision
Viikari M, Eloholma M, Halonen L. 2005. 80 years of V (λ) use: a review. Light & Engineering
13(4):24-36
Wagdy A, Fathy F. 2015. A parametric approach for achieving optimum daylighting performance
through solar screens in desert climates. Journal of Building Engineering 3:155-170
Waldrop MM. 2016. The chips are down for Moore's law. Nature 530(7589):144-147
Wang T, Ward G, Lee ES. 2016. Validation of F-matrix and six-phase method 15th International
Radiance Workshop. Padova, Italy.
Wang T, Ward G, Lee ES. 2017. Validating Radiance methods for parametric analysis of non-
coplanar shading system — an update. 16th International Radiance Workshop. Portland,
Oregon, USA

136
Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. 2004. Image quality assessment: from error
visibility to structural similarity. IEEE transactions on image processing 13(4):600-612
Ward G. 1994a. A contrast-based scalefactor for luminance display. Graphics gems IV:415-421
Ward G. 1994b. The RADIANCE lighting simulation and rendering system. 21st annual
conference on Computer graphics and interactive techniques. Florida, USA. p. 459-472
Ward G. 2012. The Radiance rtcontrib Program. 11th International Radiance Workshop.
Copenhagen, Denmark.
Ward G. 2014. Radiance 4.2 Changes for 2014. 13th International Radiance Workshop.
London, UK.
Ward G. 2015. Annual Simulation for Out-of-Plane Shading Systems. 14th International
Radiance Workshop. Philadelphia, PA, USA.
Ward G. 2017. What’s New in Radiance for 2017. 16th International Radiance Workshop.
Portland, Oregon, USA.
Ward G, Kurt M, Bonneel N. 2012. A Practical Framework for Sharing and Rendering Real-
World Bidirectional Scattering Distribution Functions. DOE/LBNL.
Ward G, R Mistrick P, Lee ES, McNeil A, J Jonsson P. 2011. Simulating the Daylight
Performance of Complex Fenestration Systems Using Bidirectional Scattering
Distribution Functions within Radiance. LEUKOS 7(4):241
Ward G, Rubinstein FM. 1988. A new technique for computer simulation of illuminated spaces.
Journal of the Illuminating Engineering Society 17(1):80-91
Ward G, Rubinstein FM, Grynberg A. 1987. Luminance in computer-aided lighting design.
California, USA. LBNL.
Ward G, Shakespeare R, Ehrlich C, Mardaljevic J, Phillips E, Apian-Bennewitz P. 1998.
Rendering with radiance: the art and science of lighting visualization. San Francisco, CA,
USA: Morgan Kaufmann ISBN 9780974538105
Ward GJ. 1994c. Adaptive shadow testing for ray tracing. Photorealistic Rendering in Computer
Graphics. Springer:11-20
Ward GJ, Rubinstein FM, Clear RD. 1988. A ray tracing solution for diffuse interreflection. ACM
SIGGRAPH Computer Graphics 22(4):85-92
Warsta J, Abrahamsson P. 2003. Is open source software development essentially an agile
method. 3rd Workshop on Open Source Software Engineering. Portland, Oregon,
USA:143-147
Weber L. 1885. Intensitätsmessungen des diffusen Tageslichtes. Annalen der Physik
262(11):374-389
Wienold J. 2009. Dynamic daylight glare evaluation. 11th Conference of International Building
Performance Simulation Association. Glasgow, Scotland. p. 27-30
Wienold J. 2010. Daylight glare in offices. [Stuttgart, Germany]. Fraunhofer-Verlag
Wienold J, Christoffersen J. 2005. Towards a new daylight glare rating. Lux Europa. Berlin,
Germany:157-161

137
Wienold J, Reetz C, Kuhn T. 2004. Evalglare: a new RADIANCE-based tool to evaluate glare in
office spaces. 3rd International Radiance Workshop. Fribourg, Switzerland.
Wilcox S, Marion W. 2008. Users manual for TMY3 data sets. Colorado, USA. NREL.
Woop S, Schmittler J, Slusallek P. 2005. RPU: a programmable ray processing unit for realtime
ray tracing. ACM Transactions on Graphics (TOG) 24(3):434-444
Zuo W, McNeil A, Wetter M, Lee ES. 2014. Acceleration of the matrix multiplication of Radiance
three phase daylighting simulations with parallel computing on heterogeneous hardware
of personal computer. Journal of Building Performance Simulation 7(2):152-163

138
VITA – Sarith Subramaniam

EDUCATION______________________________________________________
The Pennsylvania State University University Park, PA, USA
Master of Science in Architectural Engineering 2013
Pune University Maharashtra, India
Bachelor of Electrical Engineering 2006

PROFESSIONAL EXPERIENCE______________________________________
HKS Inc. Dallas, TX, USA
Simulation Specialist 2017
Lawrence Berkeley National Laboratory Berkeley, CA, USA
Graduate Researcher 2016
The Pennsylvania State University University Park, PA, USA
Graduate Research Assistant 2011-2017
BMTC Dubai, UAE
Lighting and Lighting Controls Engineer 2007-2010
Wipro Ltd. Chennai, India
Lighting and Lighting Controls Engineer 2006-2007

You might also like