Content uploaded by Mohammed Ayoub
Author content
All content in this area was uploaded by Mohammed Ayoub on Nov 19, 2024
Content may be subject to copyright.
1
A Review on Machine Learning Algorithms to Predict Daylighting inside
1
Buildings
2
3
Mohammed Ayoub (Corresponding Author)
4
Professor of Architecture and Environmental Design
5
Arab Academy for Science, Technology and Maritime Transport, Alexandria, Egypt
6
Email: dr.ayoub@aast.edu
7
https://doi.org/10.1016/j.solener.2020.03.104
8
9
Abstract
10
Steep increases in air temperatures and CO2 emissions have been associated with the global demand for energy. This is coupled
11
with population growth and improved living standards that encourages the reliance on mechanical acclimatization. Lighting energy
12
alone is responsible for a large portion of total energy consumption in office buildings; and the demand for artificial light is expected
13
to grow in the next years. One of sustainable approaches to enhance energy-efficiency is to incorporate daylighting strategies,
14
which entail the controlled use of daylight inside buildings. Daylight simulation is an active area of research that offers accurate
15
estimations, yet requires a complex set of inputs. Even with today’s computers, simulations are computationally expensive and
16
time-consuming, hindering to acquire accelerated preliminary approximations in acceptable timeframes, especially for the iterative
17
design alternatives. Alternatively, predictive models that build on machine learning algorithms have granted much interest from
18
the building design community due to their ability to handle such complex non-linear problems, acting as proxies to heavy
19
simulations. This research presents a review on the growing directions that exploit machine learning to rapidly predict daylighting
20
performance inside buildings, putting a particular focus on scopes of prediction, used algorithms, data sources and sizes, besides
21
evaluation metrics. This work should improve architects’ decision-making and increase the applicability to predict daylighting.
22
Another implication is to point towards knowledge gaps and missing opportunities in the related research domain, revealing future
23
trends that allow for such innovative approaches to be exploited more commonly in Architectural practice.
24
25
Abbreviations:
26
ANFIS
Adaptive Neuro-Fuzzy Inference System
ISO
International Standards Organization
ANN
Artificial Neural Network
LS-SVM
Least-Squares Support Vector Machine
AR
Auto Regression
MAD
Mean Absolute Deviation
ARE
Average Relative Error
MAE
Mean Absolute Error
ASE1000/250h
Annual Sunlight Exposure
MBE
Mean Biased Error
ASHRAE
American Society of Heating, Refrigerating and Air-
Conditioning Engineers
MLAs
Machine Learning Algorithms
BPNN
Back-Propagation Neural Network
MLR
Multiple Linear Regression
CART
Classification and Regression Tree
MSE
Mean Square Error
CBDM
Climate-Based Daylight Modelling
NB
Naïve Bayes
CDA
Continuous Daylight Autonomy
PC
Predictive Confidence
CFS
Complex Fenestration Systems
PCA
Principal Component Analysis
CGI
CIE Glare Index
PE
Percentage Error
CHAID
Chi-Square Automatic Interaction Detection
R
Coefficient of Correlation
CIE
Commission Internationale de l'Éclairage
R2
Coefficient of Determination
CNN
Convolutional Neural Networks
RBF
Radial Basis Function
DA
Daylight Autonomy
RBFNN
Radial Basis Function Neural Networks
DC
Daylight Coefficient
RER
Relative Error Rate
DF
Daylight Factor
RF
Random Forest
DGP
Daylight Glare Probability
RMSE
Root Mean Square Error
DT
Decision Tree
RNN
Recurrent Neural Networks
FFNN
Feed-Forward Neural Network
sDA300/50%
Spatial Daylight Autonomy
GA
Genetic Algorithms
SVC
Support Vector Clustering
GP
Gaussian Process
SVM
Support Vector Machine
ID3
Iterative Dichotomiser 3
SVR
Support Vector Regression
IES
Illuminating Engineering Society
UDI
Useful Daylight Illuminance
IESNA
Illuminating Engineering Society of North America
27
2
1. Introduction
28
Over recent years, steep increases in the air temperatures (Santamouris et al., 2015), which cause
29
extreme weather fluctuations and climate changes (NASA, 2019; Pérez-Lombard et al., 2008), have
30
been associated with the growing demand for energy. In the US and other regions around the world,
31
almost 40% of the total energy is consumed by the building sector (Allouhi et al., 2015; USDOE, 2019),
32
accounting for more than 27% of the related CO2 emissions (IEA, 2018); while over 75% of the global
33
electricity is generated from non-renewable resources (IEA, 2019a). This is seemingly clear in cooling-
34
dominant climates, where heat gains due to excessive solar radiation is inevitable (Ayoub, 2019a). With
35
the population growth and the improved living standards, the operation of mechanical acclimatization
36
increased as well, consuming more than three-quarters of the total energy to maintain indoor thermal
37
comfort (Koch-Nielsen, 2013). Even with better energy-efficiency practices (IEA, 2019b), building sector
38
continues to demand more energy (De Cian et al., 2007; Santamouris et al., 2015), supported by the
39
expanded capacity to generate electrical power in many developing countries; despite the resulting
40
environmental impacts, such as pollution, global warming, greenhouse effect and ozone depletion
41
(Pérez-Lombard et al., 2008).
42
43
On the other hand, lighting energy has accounted for almost 19% of total electricity consumption (IEA,
44
2006), constituting over 20% of the consumed energy in office buildings (Ürge-Vorsatz et al., 2012). But
45
with the recent technological progress, and the developed daylight standards and regulations to
46
enhance building performance, such as Commission Internationale de l'Éclairage (CIE) (Wyszecki,
47
1970), International Standards Organization (ISO) (Cianfrani et al., 2009) and Illuminating Engineering
48
Society (IES) (DiLaura et al., 2011), lighting becomes responsible for only 7% of the used energy in
49
buildings (IEA, 2019c). Still, the global demand for artificial light continues to grow by 2.4% annually
50
(IEA, 2006). Moreover, lighting energy represents a main source of heat that has a profound impact on
51
the consumption of cooling energy (Tzempelikos and Athienitis, 2007). Adequate daylighting strategies
52
play an important role in reducing both lighting and cooling energies, especially in harsh climates
53
(Ayoub, 2019a; Iqbal and Al-Homoud, 2007). In fact, a proper consideration of both daylighting and
54
excessive solar penetration can save more than 30% of consumed cooling energy (Wong et al., 2010).
55
Therefore, lighting prediction is imperative to achieve building energy-efficiency (Amasyali and El-
56
Gohary, 2018). One of sustainable approaches to enhance such efficiency (Nasrollahi and Shokri,
57
2016), while promoting visual and thermal comforts as well, is to incorporate daylighting strategies (IEA,
58
2000) that entail the controlled use of natural daylight inside buildings (Reinhart, 2014). Such
59
considerations become essential for Architecture and building design. Yet, even with the constant
60
efforts to integrate those techniques into the design process, the used tools and the applied professional
61
methods to predict daylighting performance in buildings are still impractical (Nault et al., 2017).
62
63
Early years of daylighting prediction have witnessed several attempts to quantify internal daylighting
64
conditions via simple methods, including diagrams (Millet et al., 1980; Waldram and Waldram, 1923),
65
protractors (Dufton, 1946; Hopkinson et al., 1966), scale models (Aizlewood and Littlefair, 1996),
66
mathematical formulas (Athienitis and Tzempelikos, 2002; Copping, 1987) and rule-of-thumbs (Reinhart
67
3
and LoVerso, 2010). Supported by the progress of computer processing power (Koomey et al., 2011;
68
Parkhurst et al., 2006), newly developed methods, known as white-box or engineering, extended such
69
conventional practices to simulation-derived approaches (Ward, 1994) that involve Daylight Coefficient
70
(DC) (Mardaljevic, 2000a; Reinhart and Walkenhorst, 2001) and Climate-Based Daylight Modelling
71
(CBDM) (Mardaljevic, 2000b; Reinhart and Herkel, 2000), which offer more accurate predictions. While
72
daylight simulation methodologies vary in scale and complexity from evaluating simple internal spaces,
73
complex façade fenestrations, to urban contexts, they rely on a common set of steps (Reinhart et al.,
74
2006; Rogers and Goldman, 2006). They involve using physically-based simulation tools to quantify the
75
spatiotemporal luminous conditions at given sensor grid-points within a built environment, accounting
76
for material properties and fluctuations of sky luminance distribution as derived from representative
77
weather data of the study location at designated temporal resolutions (Ayoub, 2019b; Mardaljevic,
78
2000b; Reinhart and Herkel, 2000). For additional details on different approaches of CBDM, such as 4-
79
Component Method, DAYSIM, 3-, 5-, 4- and 6-Phase Methods, the readers are referred to (Ayoub,
80
2019b; McNeil, 2014, 2013).
81
82
Still, the building design community is confronted by several challenges: (i) Daylight simulations demand
83
a complex set of inputs (Ayoub, 2019b), such as weather data (ASHRAE, 2019; NREL, 1995; NSRD,
84
2019), sky model (CIE, 2003; Perez et al., 1993), space geometry, location, surrounding context
85
(Ayoub, 2019a), opening configurations (Reinhart et al., 2013) and material properties (Ward, 1992).
86
(ii) Along with potential sources of errors (Labayrade and Fontoynont, 2009; Lim et al., 2010; Ochoa et
87
al., 2012), daylight simulations inherit many difficulties that stem from uncertainties related to complex
88
geometries and material definition that cannot accurately match the reality (Inanici, 2013). (iii) Also, the
89
output of daylight simulations typically yields a huge amount of data that need to be reduced into
90
simplified metrics (Reinhart et al., 2006; Reinhart and Wienold, 2010). Most of daylight metrics, such
91
as Daylight Autonomy (DA) (Reinhart and Walkenhorst, 2001) and Useful Daylight Illuminance (UDI)
92
(Nabil and Mardaljevic, 2006), report the temporal occurrences of visual comfort, but without spatial
93
considerations. Although the approved method of daylight metrics IES LM-83-12 (IES, 2012) expanded
94
such metrics with spatial consideration, and introduced Spatial Daylight Autonomy (sDA300/50%) and
95
Annual Sunlight Exposure (ASE1000/250h), they do not capture both spatial and temporal occurrences all
96
together (Ayoub, 2019b). (iv) Even with today’s computers, daylight simulations are time-consuming
97
and computationally expensive. These limitations pushed many researchers to develop more efficient
98
approaches, such as 3- and 5-phase CBDM (Lee et al., 2018; Saxena et al., 2010), along with real-time
99
prediction of building performance (Jones and Reinhart, 2015; McNeil and Lee, 2012; Schardl,
100
2016). (v) Yet again, architects that lack simulation skills become unexcited to include what they
101
consider complex methods into their practices (Mardaljevic, 2015). Even for experienced practitioners,
102
annual simulations can be tedious to acquire multiple design alternatives in acceptable timeframes. (vi)
103
This bottlenecks the design process of iterative nature, and hinders rapid preliminary approximations
104
during the early design stages (Ibarra and Reinhart, 2009), where many vital decisions are made.
105
106
107
4
Today, Artificial Intelligence applications are achieving exceptional successes in many vital fields. They
108
are being exploited not only to analyze data, but also in applications of creativity, including image
109
recognition, language processing and others. Worthwhile predictive models that build on Machine
110
Learning Algorithms (MLAs), called black-box since the building and its components are unknown, have
111
been and continues to receive much recognition from the building design community due to their ability
112
to handle complex non-linear problems. MLAs act as proxies to heavy simulations, as they learn from
113
relevant data to construct mathematically-fit models. Useful information is then extracted before
114
accurate and rapid predictions can be made from newly input data; and this can be achieved without
115
requiring the original building information or conducting any computational simulations. This domain is
116
one of the most interesting paradigms research community is shifting towards over the past decade. It
117
has long been employed in numerous research activities to predict energy consumption in buildings,
118
paralleled with several review studies. Still, machine learning is underexploited to predict daylighting
119
inside buildings, where architects and researchers are just beginning to leverage from its capabilities.
120
121
2. Related Review Studies
122
Lately, MLAs are being utilized to predict internal daylighting (Kazanasmaz et al., 2009; Lorenz and
123
Jabi, 2017; Yoon et al., 2016; Zhou and Liu, 2015), visual comfort (Chatzikonstantinou and Sariyildiz,
124
2016; Nasrollahi and Shokri, 2016; Navada et al., 2016), sky and efficacy models (Janjai and Plaon,
125
2011; Li et al., 2010; López and Gueymard, 2007; Pattanasethanon et al., 2008), weather elements
126
(Chen et al., 2011; Zhang et al., 2017; Zou et al., 2016), shading systems performances (Jain and Garg,
127
2018), High Dynamic Range Imagery (HDRI) illumination from images (Gardner et al., 2017; Zhang and
128
Lalonde, 2017) or 3D objects (Weber et al., 2018), artificial lighting (Bellocchio et al., 2011; Şahin et al.,
129
2015), among other applications. Moreover, there have been many studies that focused on reviewing
130
the use of MLAs to predict solar radiation (Voyant et al., 2017; Yadav and Chandel, 2014; Zendehboudi
131
et al., 2018), sky model classification (Li and Lou, 2018), weather data (Lou et al., 2017) and building
132
performance (Chakraborty and Elzarka, 2019).
133
134
Among recent reviews that considered the prediction of energy consumption is the work of Swan and
135
Ugursal (2009), who identified two approaches to predict energy consumption in buildings, regarding
136
the order of data input: bottom-up that involves statistical and engineering methods, also called white-
137
box; and top-down that is attributed to technological and economic methods. The review compared
138
those approaches in terms of data type and size, input of information, complexity and applications. Zhao
139
and Magoulès (2012) drew four approaches to predict energy consumption in buildings: engineering
140
methods that utilize physically-based simulations; statistical methods that relate energy use with the
141
associated design parameters via relevant data; artificial intelligence algorithms that perform predictions
142
using the related learning data; and hybrid or grey-box methods that combine white- and black-box
143
methods to eliminate their shortfalls. The review evaluated those methods against model complexity,
144
running speed, ease of use, inputs complexity and accuracy. Similarly, Foucquier et al. (2013)
145
highlighted three approaches to predict energy consumption: physically-based white-box that entails
146
computational fluid dynamics, besides zonal and nodal methods; statistical black-box that includes
147
5
Multiple Linear Regression (MLR), Artificial Neural Network (ANN), Support Vector Machine (SVM) and
148
Genetic Algorithms (GA); and grey-box methods. Each method was discussed in terms of applications,
149
complexity, ease of use, computation time, data size and algorithm hyperparameters. Ahmad et al.
150
(2014) outlined two approaches to predict energy use: MLAs, such as ANN and SVM; and hybrid
151
methods. They were compared against complexity, ease of use, speed, inputs type and accuracy.
152
153
An interesting review by Fumo (2014) summarized different attempts to classify energy prediction,
154
focusing on whole building energy, model verification and weather data. The review grouped four
155
approaches to predict energy: hybrid; statistical; engineering; and steady-state or dynamic methods,
156
acknowledging that the best approach would depend on type, scale and size of data, besides design
157
stage. Wang and Srinivasan (2017) identified three categories of energy prediction: white-, black- and
158
grey-box approaches, focusing on black-box single algorithms, such as MLR, ANN and Support Vector
159
Regression (SVR); besides other ensemble algorithms (Hansen and Salamon, 1990). The review
160
compared those approaches against ease of use, applications, accuracy, computation speed and
161
difficulty, providing suggestions for future trends. Another review on the use of black-box approaches
162
to predict energy consumption in buildings was conducted by Amasyali and El-Gohary (2018), who cited
163
63 academic research papers, and presented a comparison of the used MLAs against building type,
164
temporal graduality, type of energy consumption, purpose of prediction, data type, data size, data pre-
165
processing, feature type and performance evaluation. The review also discussed the related future
166
research directions. Likewise, Wei et al. (2018) compared three approaches to predict energy: white-,
167
grey- and black-box, along with their applications. The review focused on black-box method, and
168
reviewed 61 research papers that utilized ANN, SVM, MLR, Decision Tree (DT), GA, K-means, Self-
169
Organizing Map and Hierarchal Clustering, offering a comparison among them against scale, measure
170
length, input variables and data source.
171
172
The accumulations of those review studies mainly identified three approaches to predict building
173
performance: white-, black- and grey-box (Fig. 1), confirming the focus that was given to MLAs to solve
174
regression, classification and clustering problems. Regarding the black-box approach, which is of main
175
concern to this review, the used criteria and considerations to compare MLAs themselves can be
176
realized by algorithm-related features that involve, according to previous reviews: scope of prediction,
177
applications, available data, complexity, ease of use, speed, accuracy and hyperparameters. Herein,
178
choosing the best algorithm mainly depends on case-specific, data-related attributes, which include the
179
available data type, scale and size.
180
181
6
182
183
184
185
186
187
188
189
190
Fig. 1. The reviewed approaches to predict building performance with different criteria to compare MLAs.
191
192
3. Research Aim and Methodology
193
To the author’s best knowledge, hardly any attempt can be found in literature that addresses reviewing
194
previous research attempts to use MLAs to predict luminous conditions and daylighting performance in
195
buildings. Such scarcity can be attributed to the relatively novel, yet promising, field of application. Thus,
196
this research aims to present a review on the growing directions found in the scholarly literature that
197
exploit different MLAs to rapidly predict daylighting performance inside buildings. This review will only
198
consider MLAs that offer accelerated approximations, rather than other optimization algorithms that
199
take a longtime to find global optima. Also, it will not reflect the research landscape of neither electrical
200
lighting nor external daylighting, as they are out of the scope, and cannot fit the extent of its specified
201
target; yet, a study that considers those issues can be undertaken separately in the future.
202
203
The work of this review builds on previous reviews on different machine learning approaches to predict
204
energy consumption in buildings (Ahmad et al., 2014; Amasyali and El-Gohary, 2018; Foucquier et al.,
205
2013; Fumo, 2014; Swan and Ugursal, 2009; Wang and Srinivasan, 2017; Wei et al., 2018; Zhao and
206
Magoulès, 2012). It is based on a thorough analysis and citation of up-to-date academic research
207
papers, considering a set of criteria to compare MLAs, derived from previous review studies (Fig. 1).
208
They include four main groups, each of which comprises a number of subcategories: (i) Scope: building
209
types, climate zones and locations; (ii) MLAs: problem type, selected MLAs with their hyperparameters,
210
input parameters and output parameters; (iii) Data: data sources, data sizes and temporal granularities;
211
as well as (iv) Accuracy: evaluation metrics. Finally, the review offers a detailed discussion about those
212
criteria. This review seeks to be more inclusive, yet easier to be realized by architects of different
213
backgrounds and expertise. That should increase the applicability to predict internal daylighting using
214
machine learning, and consequently, support architects’ design decision-making. Another implication
215
of this work is to point towards knowledge gaps and missing opportunities in the related research
216
domain, revealing future trends and perspectives that allow for such innovative approaches to be
217
exploited more commonly in Architectural practice. It is expected that, sooner than later, machine
218
learning will contribute to more challenging problems in Architecture, and specifically, in the field of
219
daylighting prediction.
220
221
7
To achieve the research aim, the methodological procedure is realized through several steps. First, a
222
Boolean keyword search is conducted on the scholarly literature by Google Scholar and Egyptian
223
Knowledge Bank, allowing to combine the related keywords with search operators such as ‘AND’ and
224
‘OR’ to find detailed results. Such engines facilitate a broad search across many disciplines and sources
225
within accredited academic databases and publishers. To focus on the related research domain and
226
the targeted scope, the resulting scholarly literature is screened and analytically reviewed to identify
227
their purpose and scope. Each of which is then filtered to distinguish relevant candidate articles within
228
the scope of this review. In the refined scholarly literature, the articles that were cited by other articles
229
are filtered too using the same relevance criteria. Those that passed the filtration are also identified as
230
additional candidate articles. All candidate articles are then categorized as per the topical structure of
231
the entire review, which is divided into the following sections:
232
• Section 1 provides a brief introduction on white- and black-box approaches;
233
• Section 2 highlights the related review studies that focused on the use of MLAs in building
234
performance, defining their scope;
235
• Section 3 defines the research aim and describes the methodological procedure of this review;
236
• Section 4 describes machine learning, its process and categories, along with a brief background
237
on some of the famous MLAs;
238
• Section 5 analyzes previous studies that utilized MLAs to predict internal daylighting conditions,
239
in terms of the identified four criteria groups: Scope of Prediction; MLAs; Data Sources and
240
Sizes; and Evaluation Metrics;
241
• Section 6 discusses those studies in detail considering the criteria and their subcategories to
242
reveal the missing opportunities in the related research domain;
243
• Section 7 directs future research perspectives and trends; and
244
• Section 8 concludes the review findings.
245
246
4. Background
247
4.1. Machine Learning Process and Categories
248
Derived from computer science, machine learning is the study of statistical models and algorithms that
249
use a dataset’s variables to recognize relevant spatiotemporal patterns and information. Machine
250
learning offers computer systems with whole new abilities, as they can perform rapid predictions from
251
newly input data, without being instructed to perform such task (Mitchell, 1997). Basically, machine
252
learning searches sample training data to build and formulate a mathematically-fit model (Kalogirou,
253
2001; Liu et al., 2019), which maps the complex relationships between independent inputs and target
254
outputs. This is where evaluation metrics are used to assess the model’s performance (Li and Lou,
255
2018). Machine learning process encompasses a set of steps: (i) Data Collection entails gathering
256
relevant historical, experimental, observational, or simulation-derived data, before constructing the
257
model. The quality of data directly affects how accurate the predictive model will be. (ii) Data Preparation
258
involves data normalization, scaling and randomization; and this is essential, as the ranges of data
259
variables can vary widely, influencing how MLAs function. Data randomization is conducted as well, to
260
8
prevent the order of data from affecting the learning. (iii) Data Exploration and Visualization by means
261
of statistical and visualization techniques is vital to test if there are relevant relationships between input
262
variables, and to detect any data imbalances so that the developed model will not be biased towards
263
predicting a particular range of variables. (iv) Data Pre-Processing includes splitting the data into 3
264
parts. The first, and the largest, is used for training the model; the second is for testing or evaluating
265
the performance of the model during the training; and the third part is for validating or fine-tuning MLAs
266
hyperparameters’ to offer further improvements to the model’s performance. Finally, (v) Prediction are
267
then conducted, where the worth of machine learning is realized accordingly.
268
269
MLAs differ in the used technique, the type of variables to handle, and the nature of tasks to solve. Still,
270
they can be grouped according to the learning technique into 3 categories: supervised, unsupervised,
271
and reinforcement (Bishop, 2006) (Fig. 2); each of which has its strengths and weaknesses. Supervised
272
learning analyzes possibly generalized dataset to build the model that links input variables with output
273
ones, based on training dataset of many input-output pairs. The model can then predict target variables
274
from newly input data. For classification and regression, the supervised learning algorithms include
275
MLR, ANN, SVM, Naïve Bayes (NB), DT and Random Forest (RF). Unsupervised learning detects
276
unknown patterns in a dataset to recognize target variables without the output data. The model realizes
277
unsorted information based on similarities, differences and patterns. For clustering, feature selection
278
and dimensionality reduction, the unsupervised learning algorithms include Hierarchical Clustering, K-
279
means and Principal Component Analysis (PCA). Reinforcement learning pursues a specific target in a
280
dynamic environment without knowing if the model converges to the target or not. It takes an action as
281
input, and generates a maximum expected reward as output, where the process is driven by the
282
feedback from the environment. Reinforcement learning is studied in many disciplines, such as
283
operations research, swarm intelligence and GA. Well-established models would require adequate and
284
unbiased databases that include wide ranges of inputs and outputs for training, validating and testing
285
(Basheer and Hajmeer, 2000; Kaytez et al., 2015).
286
287
288
289
290
291
292
293
294
295
296
Fig. 2. Different categories of MLAs according to learning technique, with examples.
297
298
9
4.2. Artificial Neural Network
299
ANN is a non-linear function approximation algorithm that basically performs numerical predictions via
300
supervised, unsupervised and reinforcement learning techniques (Yao, 1999). Other ANN variations
301
are used for classification, clustering and pattern recognition (Jain et al., 1996). The biological structure
302
of neural networks that constitutes the brain and nervous system is the main concept behind ANN.
303
Essentially, ANN is composed of a number of layers; each of which contains several processing
304
elements, called neurons. The input layer receives the input data; the hidden layer(s) performs most of
305
the computations; while the output layer predicts the results (Hecht-Nielsen, 1992) (Fig. 3). Neurons of
306
one layer are interconnected to other neurons in the consequent layer by connections that are assigned
307
with weights (Jain et al., 1996). Inputs (xj) are multiplied by corresponding weights (wij), and the
308
weighted sum is adjusted by bias (θi) to minimize the error that stems from the difference between the
309
actual and predicted outputs. Then, this value passes through an activation function (fi), such as Unit
310
Step, Sigmoid, Piecewise Linear and Gaussian (Bishop, 2006), which acts as a gate that determines if
311
the neurons should transmit new data to the next layer or not. The training of ANN entails tuning the
312
weights in the neurons depending on the outputs. According to the flow of information and the
313
connection weights, data is propagated throughout the whole network; and when it reaches the output
314
layer, the neuron with highest value defines the output. This is where the difference between the
315
predicted and actual outputs is attained, changing back the connection weights to minimize that
316
difference. Such Back-Propagation training process (Rumelhart and Hintonf, 1986) is iteratively
317
updated until a termination condition is met. Readers can refer to (Schmidhuber, 2015) for a historical
318
overview and additional features on ANN.
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
Fig. 3. An example of ANN architecture, showing different layers, their corresponding weights, and various types of activation
337
functions that can be applied to the weighted sum.
338
339
10
Key hyperparameters to tune ANN before training include: the number of hidden layers; the number of
340
hidden neurons; the learning rate that defines the update frequency to adjust the errors; the number of
341
epochs that identifies the number of iterations such ANN undergoes for the whole training dataset; the
342
batch size to control the number of propagated training samples through the network before updating
343
the model; and the activation function that introduces non-linear properties to ANN and converts inputs
344
of neurons to outputs. Variations of ANN include Feed-Forward Neural Network (FFNN) (Ivakhnenko,
345
1971), Back-Propagation Neural Network (BPNN) (Rumelhart and Hintonf, 1986), Radial Basis
346
Function Neural Networks (RBFNN) (Broomhead and Lowe, 1988), Recurrent Neural Networks (RNN)
347
(Elman, 1990), Adaptive Neuro-Fuzzy Inference System (ANFIS) (Jang, 1993) and Convolutional
348
Neural Networks (CNN) (LeCun et al., 1998).
349
350
4.3. Support Vector Machine
351
SVM is a discriminative classifier that is used fundamentally for classification, pattern recognition and
352
regression by supervised learning techniques (Boser et al., 1992). The concept behind SVM is to
353
differentiate between groups of linearly separable data features, called vectors, which belong to different
354
categories (Fig. 4). This is where an optimal separating hyperplane is defined geometrically by the
355
nearest vectors, called support vectors. The hyperplane should be as far as possible from those support
356
vectors; thus, SVM seeks to maximize the decision boundary, or the margins (ω), of such hyperplane.
357
This will ensure the accuracy of classification, allowing for more space for newly input vectors that can
358
be classified according to the hyperplane side they belong (Wu et al., 2008). To this point, the data is
359
assumed to be linearly separable, where it is possible to find a separating hyperplane that will perfectly
360
classify the data; but this rarely occurs in real-life. Therefore, instead of letting SVM classify data
361
correctly and risk overfitting, some mistakes are intentionally allowed by introducing a penalty for each
362
data vector, the slack variable (ξi). In this sense, SMV seeks to lessen the slack variables by minimizing
363
the sum of their distances from the margins, while maximizing the hyperplane’s margins themselves.
364
365
Yet again, this may be too simple for practical applications; thus, SVM can handle complex non-linear
366
problems by transforming the data features into a higher-dimensional space using a non-linear kernel
367
function, such as Polynomial, Gaussian Radial Basis Function (RBF) and Hyperbolic Tangent
368
(Aiserman et al., 1964). In a (d) dimensional features space, the hyperplane is a (d-1) dimensional
369
separator; for instance, in a 3D space, the hyperplane is a 2D plane. As the separating hyperplane of
370
maximized margin is deployed, different vectors can be classified as linearly separable data in such
371
higher-dimensional space (Fig. 4) (Vapnik, 1999). Then, the non-linear solution is implicitly attained by
372
projecting the results to the original lower-dimensional space (Kotsiantis, 2007). For further details and
373
overview on SVM, readers may refer to (Burges, 1998). Essential hyperparameters to tune SVM
374
include: parameter (C) that defines the margin width for penalizing the misclassified vectors to avoid
375
overfitting; the used kernel function; and the kernel’s specific parameters, such as the degree of
376
Polynomial Kernel and the width of RBF Kernel. Other diversifications of SVM include Least-Squares
377
Support Vector Machine (LS-SVM) (Suykens and Vandewalle, 1999), Support Vector Clustering (SVC)
378
(Ben-Hur et al., 2001), Bayesian SVM (Polson and Scott, 2011) and SVR (Drucker et al., 1997).
379
11
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
Fig. 4. An example of linear and non-linear SVM, showing the definition of optimization problems and the optimal separating
400
hyperplane in both cases.
401
402
4.4. Other Popular Algorithms
403
Other algorithms, including Regression, DT and RF, have received much recognition, and besides
404
daylighting, they were implemented to predict different aspects of building performance. Regression
405
entails a set of methods that mainly predict the relationship between two sets of variables: a response
406
(dependent) variable and explanatory (independent) variables. As an inherently supervised algorithm,
407
MLR is one of the most exploited form of regression, by which a linear equation fits the training dataset,
408
making assumptions about the linearity, variance and independence of errors. There are many
409
examples of Regression algorithms designed for different purposes according to the data type and
410
distribution, such as Polynomial Regression (Smith, 1918), Logistic Regression (Cramer, 2002;
411
Kleinbaum et al., 2002), Bayesian Regression (Fienberg, 2006) and Auto Regression (AR) (Akaike,
412
1969).
413
414
DT is a supervised algorithm that is widely used to solve both classification and regression problems.
415
For a given set of inputs with specific features, DT maps an inverted treelike structure that consists of
416
nodes and branches to predict the output that results from hierarchical steps of decisions. DT starts
417
with a root node that checks if the prediction of a given input values or classes is satisfied. This is where
418
the data is split into branches that connect the nodes, representing yes or no decision responses. Then,
419
each internal (non-leaf) node evaluates the next input feature; and the decisions continue to traverse
420
12
downwards, leading eventually to a leaf node that corresponds to the final prediction output. Such DT
421
structure can grow flexibly according to the input data. There are many diversifications of DT algorithm,
422
which include Iterative Dichotomiser 3 (ID3) (Quinlan, 1986), C4.5 (Quinlan, 1993), Classification and
423
Regression Tree (CART) (Breiman et al., 1984), RF (Ho, 1995) and Chi-Square Automatic Interaction
424
Detection (CHAID) (Kass, 1980).
425
426
All those MLAs rely on supervised learning approach, and can be used for classification and regression
427
problems considering discrete or continuous data. They received much attention by many researchers;
428
and the next section reviews previous studies that utilized MLAs to predict luminous conditions and
429
daylighting performance inside buildings.
430
431
5. Previous Studies on Machine Learning Algorithms to
432
Predict Internal Daylighting
433
The landscape of scholarly literature that put the focus on employing different MLAs to predict internal
434
daylighting conditions is limited, yet continues to grow. To help architects comprehend the most recent
435
directions, while offering guidance for future research trends and areas of development as well, a total
436
of 27 related studies, which were published in the last 13 years (2006-2019), were found, then reviewed
437
thoroughly. Every study involves a number of investigational attempts that vary according to the
438
objective, building type, location, selected algorithm and output parameters. Consequently, the total
439
number of investigational attempts exceeds the number of studies themselves (109 attempts for 27
440
research paper). (Table 1) presents a comprehensive outline on such recent work, which is arranged
441
chronologically in terms of: Scope of Prediction, MLAs, Data Sources and Sizes, and Evaluation
442
Metrics; each of which is discussed in the following section with the same order. The summary of the
443
coordinates plot’s results and the distributions of their frequency rates are shown in (Fig. 5) according
444
to: (a) Building Type, (b) Climate Zone, (c) Location, (d) Problem Type, (e) Selected MLAs, (f)
445
Hyperparameters, (g) Input Parameters, (h) Output Parameters, (i) Data Source, (j) Simulation Tool, (k)
446
Temporal Granularity and (l) Error Metrics.
447
448
13
Table 1. Summary of the reviewed researches that considered predicting internal daylighting using machine learning, in terms of scope, the used algorithm, data and accuracy.
449
Author(s)
Scope
Machine Learning Algorithm
Data
Accuracy
Building
Type
Location and
Climate Zone
Problem
Type
Selected
Algorithm
Hyperparameters
Input Parameters
Output
Parameters
Data Source
Data Size
Temporal
Granularity
Evaluation Metrics
(Kurian et al., 2006)
N/S
Manipal,
Karnataka, India
(Tropical
Monsoon, Am)
Regression
ANN
No. Epochs
Window Width
Window Height
Windows Location
Room Width
Room Length
Room Height
Room Elevation
Obstruction Height
Obstruction Length
Obstruction Width
Global Horizontal Radiation
Direct Normal Radiation
Diffuse Horizontal Radiation
Illuminance
Values
Simulation-
Derived
(Radiance)
500 for Training
500 for Testing
Annual
Timeseries
(Hourly Basis)
0.01-3.50 lx (PE)
0.13-1.12% (RMSE)
AR
N/S
0.19-0.91 lx (PE)
(Binol, 2008)
Office
Izmir, Turkey (Hot-
Summer
Mediterranean,
Csa)
Regression
ANN
No. Hidden Neurons
No. Epochs
Orientation
Sensor Point Identification
Distance from Windows
Floor/Room Identification
Number of Windows
Room Width
Room Length
Room Height
Global Horizontal Radiation
Relative Humidity
UV Index
UV Dose
Time of the Day
Day of the Year
Illuminance
Values
Field
Measurements
3,168 for Training
792 for Testing
Selected
Instances (Hourly
Basis)
2.20-2.54% (PE)
(Conraud-Bianchi,
2008)
Educational
Grong, Norway
(Subarctic, Dfc)
Regression
ANN
No. Hidden Neurons
Orientation
Window Width
Window Height
Shading Device
Shading Parameters
Dry Bulb Temperature
DF
Simulation-
Derived (ESP-r)
1,200 for Training
300 for Testing
Annual (Hourly
Basis)
0.72-2.50% (ARE)
(Kurian et al., 2008)
N/S
Mangalore,
Karnataka, India
(Tropical
Monsoon, Am)
Regression
ANN
No. Epochs
Orientation
Direct Normal Radiation
Diffuse Horizontal Radiation
Dry Bulb Temperature
Time of the Day
Illuminance
Values
Simulation-
Derived (Ecotect)
500 for Training
500 for Testing
Annual
Timeseries
(Hourly Basis)
0.01-3.50 lx (PE)
0.13-1.12% (RMSE)
AR
N/S
N/S
(Kazanasmaz et al.,
2009)
Office
Izmir, Turkey (Hot-
Summer
Mediterranean,
Csa)
Regression
ANN
No. Hidden Neurons
Orientation
Sensor Point Identification
Distance from Windows
Floor/Room Identification
Number of Windows
Room Width
Room Length
Room Height
Global Horizontal Radiation
Dry Bulb Temperature
Relative Humidity
UV Index
UV Dose
Time of the Day
Day of the Year
Illuminance
Values
Field
Measurements
8,000 for Training
2,000 for Testing
Selected
Instances (Hourly
Basis)
1.08-3.32% (PE)
(Ahmed et al.,
2011a)
Office
Cork, Ireland
(Temperate
Oceanic, Cfb)
Classification
NB
N/S
Orientation
Window Width
Window Height
Room Width
Room Length
Room Height
Room Elevation
Wall Material
Floor Material
Ceiling Material
Obstruction Height
Illuminance
Values
Field
Measurements
Simulation-
Derived (Ecotect)
906,540 for
Training
140,218 for
Testing
Selected
Instances (Sub-
Hourly Basis)
99.47% (PC)
DT
N/S
99.39% (PC)
Laboratory
14
SVM
N/S
Obstruction Length
Obstruction Width
Global Horizontal Radiation
Global Horizontal Illuminance
99.72% (PC)
(Ahmed et al.,
2011b)
Office
Laboratory
Cork, Ireland
(Temperate
Oceanic, Cfb)
Regression
DT
N/S
Orientation
Window Width
Window Height
Room Width
Room Length
Room Height
Room Elevation
Work Plane Height
Wall Material
Floor Material
Ceiling Material
Obstruction Height
Obstruction Length
Obstruction Width
Global Horizontal Radiation
Diffuse Horizontal Radiation
Time of the Day
Illuminance
Values
Field
Measurements
Simulation-
Derived (Ecotect)
906,540 for
Training
140,218 for
Testing
Selected
Instances (Sub-
Hourly Basis)
99.39% (PC)
SVM
N/S
94.99% (PC)
(Hu and Olbina,
2011)
Office
Gainesville,
Florida, US (Humid
Subtropical, Cfa)
Regression
ANN
No. Hidden Layers
No. Hidden Neurons
Global Horizontal Radiation
Direct Normal Radiation
Diffuse Horizontal Radiation
Horiz. Infrared Radiation Intensity
Global Horizontal Illuminance
Diffuse Horizontal Illuminance
Exterior Horiz. Beam Illuminance
Dry Bulb Temperature
Relative Humidity
Time of the Day
Solar Altitude
Solar Azimuth
Solar Declination
Solar Hour
Illuminance
Values
Simulation-
Derived (DElight)
1,404 for Training
936 for Testing
Selected
Instances (Hourly
Basis)
1.50-5.30% (PE)
(da Fonseca et al.,
2013)
Office
Florianópolis,
Brazil (Humid
Subtropical, Cfa)
Regression
ANN
No. Hidden Neurons
No. Epochs
Learning Rate
Activation Function
Orientation
Window Width
Window Height
Glazing Properties
Shading Device
Shading Parameters
Illuminance
Values
Simulation-
Derived
(DAYSIM)
195 for Training
21 for Testing
N/S
0.03% (MSE)
MLR
N/S
0.05% (MSE)
0.80 (R2)
(Inanici, 2013)
N/S
Seattle,
Washington, US
(Warm-Summer
Mediterranean,
Csb)
Regression
LSMR
N/S
Orientation
Global Horizontal Illuminance
Internal Illuminance
Pixel Luminance
Time of the Day
Day of the Year
Solar Altitude
Solar Azimuth
Illuminance
Values
Field
Measurements
(High Dynamic
Range Imagery)
50 for Training
50 for Testing
Selected
Instances (Sub-
Hourly Basis)
0.69-0.99 (R2)
Simulation-
Derived
(Radiance)
0.87-0.90 (R2)
(Colaco et al., 2014)
N/S
Manipal,
Karnataka, India
(Tropical
Monsoon, Am)
Regression
ANN
No. Hidden Layers
No. Hidden Neurons
Sensor Point Identification
Glazing Properties
Wall Material
Floor Material
Ceiling Material
Illuminance
Values
Field
Measurements
49,180 for
Training
21,080 for
Testing
Selected
Timeseries (Sub-
Hourly Basis)
2.20-4.60% (RMSE)
0.98 (R2)
AR
N/S
7.50-9.20% (RMSE)
0.94-0.96 (R2)
(Logar et al., 2014)
Office
Ljubljana,
Slovenija
(Temperate
Oceanic, Cfb)
Regression
ANN
No. Membership
Functions
Shading Device
Shading Parameters
Global Horizontal Radiation
Diffuse Horizontal Radiation
Global Horizontal Illuminance
Illuminance
Values
Field
Measurements
42,000 for
Training
28,000 for
Testing
Selected
Timeseries (Sub-
Hourly Basis)
12.60-18.10% (RMS)
7.76-12.2% (MBE)
(Liu et al., 2015)
Industrial
City of Singapore,
Singapore
(Tropical
Rainforest, Af)
Clustering
DBSCAN
N/S
Occupancy Schedule
Glazing Properties
Room Width
Room Length
Room Height
Wall Material
Floor Material
Ceiling Material
Shading Device
Shading Parameters
DA
Simulation-
Derived
(Radiance)
N/S
Annual (Hourly
Basis)
N/S
15
(Zhou and Liu,
2015)
Office
Virginia, US
(Humid
Subtropical, Cfa)
Classification
ANN
No. Hidden Layers
Occupancy Schedule
Sensor Point Identification
Window Width
Window Height
Windows Location
Glazing Properties
Wall Material
Floor Material
Ceiling Material
Diffuse Horizontal Illuminance
Exterior Horiz. Beam Illuminance
Solar Altitude
Solar Azimuth
UDI
Simulation-
Derived
(OpenStudio)
729 for Training
729 for Testing
Annual (Hourly
Basis)
with PCA: 3.65-7.49% (PE)
without PCA: 9.71-16.35% (PE)
SVM
Kernel Functions
with PCA: 29.86-37.85% (PE)
without PCA: 26.12-29.92% (PE)
(Chatzikonstantinou
and Sariyildiz, 2016)
Office
Izmir, Turkey (Hot-
Summer
Mediterranean,
Csa)
Regression
ANN
Activation Function
No. Hidden Layers
No. Hidden Neurons
Orientation
Occupancy Schedule
Sensor Point Identification
View Direction
Window Width
Window Height
Windows Location
Room Width
Room Length
Room Height
Time of the Day
Day of the Year
DA
DGP
Simulation-
Derived
(Honeybee)
2,000 for Training
500 for Testing
Annual (Hourly
Basis)
0.07-0.32% (RMSE)
0.68-0.81 (R2)
SVM
Kernel Functions
(C) Parameter
Width of RBF Kernel
0.08-0.13% (RMSE)
0.59-0.94 (R2)
RF
No. Trees
No. Features
0.07-0.16% (RMSE)
0.66-0.91 (R2)
k-NN
N/S
0.09-0.14% (RMSE)
0.38-0.92 (R2)
(Navada et al.,
2016)
Office
Manipal,
Karnataka, India
(Tropical
Monsoon, Am)
Regression
ANN
No. Hidden Layers
No. Hidden Neurons
No. Epochs
Room Width
Room Length
Room Height
Work Plane Height
Shading Device
Shading Parameters
Global Horizontal Radiation
Direct Normal Radiation
Diffuse Horizontal Radiation
Illuminance
Values
Field
Measurements
5,696 for Training
2,848 for Testing
Selected
Timeseries (Sub-
Hourly Basis)
0.97-5.24% (PE)
(Ahmad et al., 2017)
Educational
Cardiff, Wales, UK
(Temperate
Oceanic, Cfb)
Regression
ANN
No. Hidden Layers
No. Hidden Neurons
Occupancy Schedule
Shading Device
Shading Parameters
Global Horizontal Radiation
Direct Normal Radiation
Diffuse Horizontal Radiation
Dry Bulb Temperature
Relative Humidity
Wind Speed
Time of the Day
Day of the Year
Solar Altitude
Solar Azimuth
Illuminance
Values
Simulation-
Derived
(EnergyPlus)
N/S
Annual (Hourly
Basis)
278.49-282.35 (RMSE)
0.98 (R2)
RF
No. Trees
No. Features
227.87-235.738 (RMSE)
0.97-0.98 (R2)
44.91-86.80 (MAD)
(Lorenz and Jabi,
2017)
Office
N/S
Regression
ANN
No. Hidden Layers
No. Hidden Neurons
Orientation
Sensor Point Identification
Distance from Windows
Floor/Room Identification
Room Width
Room Length
Room Height
Work Plane Height
Wall Material
Floor Material
Ceiling Material
DA
Simulation-
Derived (DIVA-
for-Rhino)
270 for Training
30 for Testing
Annual (Hourly
Basis)
0.11-0.28% (MSE)
(Nault et al., 2017)
Office
Geneva,
Switzerland
(Temperate
Oceanic, Cfb)
Regression
MLR
N/S
Orientation
Window Width
Window Height
Windows Location
Glazing Properties
Room Width
Room Length
Room Height
Room Elevation
sDA
Simulation-
Derived (DIVA-
for-Rhino)
280 for Training
280 for Testing
Annual (Hourly
Basis)
2.82-7.14% (RMSE)
0.59-0.98 (R)
4.30-20.90% (PE)
16
GPs
Obstruction Height
Obstruction Length
Obstruction Width
Obstruction Angle
Distance from Obstruction
Global Horizontal Radiation
Direct Normal Radiation
Diffuse Horizontal Radiation
3.16-21.09% (RMSE)
0.48-0.97 (R)
33.40-48.90% (PE)
(Uribe et al., 2017)
Office
Antofagasta, Chile
(Hot-Desert, BWh)
Regression
ANN
N/S
Window Width
Window Height
Glazing Properties
Wall Material
Floor Material
Ceiling Material
Shading Device
Shading Parameters
sDA
ASE
Simulation-
Derived
(Groundhog)
5400 for Training
Annual (Hourly
Basis)
N/S
Santiago, Chile
(Warm-Summer
Mediterranean,
Csb)
Punta Arenas,
Chile (Subpolar
Oceanic, Cfc)
(Verso et al., 2017)
Office
Berlin, Germany
(Temperate
Oceanic, Cfb)
Regression
MLR
N/S
Orientation
Window Width
Window Height
Windows Location
Glazing Properties
Room Length
Work Plane Height
Obstruction Height
Obstruction Angle
Shading Device
Shading Parameters
DA
CDA
sDA
Simulation-
Derived
(DAYSIM)
510 for Training
48 for Testing
Annual (Sub-
Hourly Basis)
2.76-13.92% (RMSE)
0.03-0.11 (CV)
Turin, Piemont,
Italy (Humid
Subtropical, Cfa)
Catania, Sicily,
Italy (Hot-Summer
Mediterranean,
Csa)
(Yacine et al., 2017)
Educational
Biskra, Algeria
(Hot-Desert, BWh)
Regression
ANN
No. Hidden Neurons
Internal Illuminance
DGP
CGI
Field
Measurements
(Experimental
Survey)
59 for Training
31 for Testing
Selected
Instances (Hourly
Basis)
39.20-69.60% (RER)
(Beccali et al., 2018)
Laboratory
Palermo, Italy
(Hot-Summer
Mediterranean,
Csa)
Regression
ANN
N/S
Sensor Point Identification
Global Horizontal Radiation
Global Horizontal Illuminance
Time of the Day
Day of the Year
Solar Altitude
Solar Azimuth
Illuminance
Values
Field
Measurements
64,786 for
Training
11,432 for
Testing
Selected
Instances (Sub-
Hourly Basis)
0.22-0.47% (MSE)
0.91-0.95 (R2)
(Liu et al., 2018)
N/S
Seattle,
Washington, US
(Warm-Summer
Mediterranean,
Csb)
Regression
ANN
N/S
Pixel Location
Room Width
Room Length
Room Height
Wall Material
Floor Material
Ceiling Material
Direct Normal Radiation
Diffuse Horizontal Radiation
Global Horizontal Illuminance
Pixel Luminance
Solar Altitude
Solar Azimuth
Illuminance
Values
Simulation-
Derived
(Radiance)
3,879 for Training
500 for Testing
Annual (Hourly
Basis)
0.01072-0.06389% (MSE)
0.01 (RER)
(Lorenz et al., 2018)
N/S
N/S
Regression
ANN
No. Hidden Neurons
Sensor Point Identification
Distance from Windows
Window Width
Window Height
Windows Location
Obstruction Height
Distance from Obstruction
DA
Simulation-
Derived (DIVA-
for-Rhino)
2057 for Training
685 for Testing
Annual (Hourly
Basis)
0.51-0.55% (MAE)
0.63-0.65% (RMSE)
0.06-0.09% (MBE)
(Radziszewski and
Waczyńska, 2018)
Office
Denver, Colorado,
US
(Cold Semi-Arid,
BSk)
Regression
ANN
No. Hidden Layers
No. Hidden Neurons
No. Epochs
Learning Rate
Activation Function
Orientation
Window Width
Window Height
Windows Location
Room Width
Room Length
Room Height
Time of the Day
Day of the Year
DF
DA
DGP
Simulation-
Derived (DIVA-
for-Rhino)
2963 for Training
200 for Testing
Annual (Hourly
Basis)
0.20-0.80% (MSE)
17
(Ayoub, 2019a)
Residential
Alexandria, Egypt
(Hot-Desert, BWh)
Regression
MLR
N/S
Window Width
Window Height
Windows Location
Glazing Properties
Obstruction Height
Obstruction Angle
Obstruction Material
Distance from Obstruction
sDA
ASE
Simulation-
Derived (DIVA-
for-Rhino)
1296 for Training
Annual (Hourly
Basis)
5.28-8.13% (RMSE)
0.87-0.98 (R2)
N/A: Not Available
N/S: Not Specified
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
18
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
Fig. 5. Summary of the review results and the distributions of their frequency rates according to: (a) building type, (b) climate zone, (c) location, (d) problem type, (e) selected MLAs, (f)
493
hyperparameters, (g) input parameters, (h) output parameters, (i) data source, (j) simulation tool, (k) temporal granularity and (l) error metrics.
494
19
5.1. Scope of Prediction
495
5.1.1. Building Types
496
The Scope of Prediction can be categorized into building types, climate zones and locations. Regarding
497
the building types, 5 categories are identified in the reviewed studies (Office, Educational, Laboratorial,
498
Industrial and Residential). (Fig. 5-a) summarizes their distribution, indicating that 60% of the studies
499
focused on office buildings, 11% and 6% considered educational and laboratorial spaces respectively,
500
2% included industrial and residential spaces equally, but the remaining 19% did not specify the studied
501
building type. Of the reviewed studies that considered office spaces, 32% were performed in Temperate
502
Oceanic climate, whilst 25% and 21% were conducted in both Hot-Summer Mediterranean and Humid
503
Subtropical climates, respectively. On the problem type, 82% were used to solve regression problems,
504
only 18% considered classifications, while clustering was not studied. About the used MLAs, 46% of
505
the studies that considered studying office spaces relied on ANN, 18% used MLR, while 11% utilized
506
both DT and SVM. Most of those studies that considered office spaces, 78%, depended on simulation
507
tools to acquire training data, whereas the remaining 22% obtained data from field measurements. 57%
508
of them covered long-term daylighting predictions, 36% selected certain temporal instances, but the
509
remaining did not specify the used temporal granularity.
510
511
5.1.2. Climate Zones and Locations
512
In compliance with Köppen-Geiger classification (Critchfield, 1983; Geiger, 1954; Kottek et al., 2006),
513
10 climate zones are found in the reviewed studies (Temperate Oceanic, Hot-Summer Mediterranean,
514
Tropical Monsoon, Humid Subtropical, Hot-Desert, Warm-Summer Mediterranean, Subarctic, Tropical
515
Rainforest, Subpolar Oceanic and Cold Semi-Arid climates), as indicated in (Table 1). Those climate
516
zones cover only 10 out of 30 sub-types of the updated Köppen-Geiger classification (Beck et al., 2018).
517
This can be attributed to the small number of researches, where a researcher, or a group of researchers,
518
conducted more than one investigation considering the same location, including (Ahmed et al., 2011a)
519
and (Ahmed et al., 2011b) in Cork, Ireland; (Binol, 2008), (Kazanasmaz et al., 2009) and
520
(Chatzikonstantinou and Sariyildiz, 2016) in Izmir, Turkey; (Kurian et al., 2006), (Colaco et al., 2014)
521
and (Navada et al., 2016) in Manipal, Karnataka, India; besides (Inanici, 2013) and (Liu et al., 2018) in
522
Seattle, Washington, US. Still, 69% of the reviewed studies focused on climate class C (temperate),
523
while the remaining 18%, 11% and 2% included classes A (tropical), B (dry) and D (continental),
524
respectively, but class E (polar) was not considered. (Fig. 5-b) reveals that Temperate Oceanic, Hot-
525
Summer Mediterranean, Tropical Monsoon and Humid Subtropical climates are the most frequently
526
used to predict internal daylighting, accounting for 28%, 17%, 15% and 13% of the attempts in the
527
reviewed studies, respectively.
528
529
Those climate zones are representative of 23 locations around the world, as demonstrated in (Fig. 5-
530
c) and (Fig. 6) (Cork, Ireland; Izmir, Turkey; Manipal, Karnataka, India; Seattle, Washington, US; Grong,
531
Norway; Mangalore, Karnataka, India; Florianopolis, Brazil; Virginia, US; Cardiff, Wales, UK; Geneva,
532
Switzerland; Biskra, Algeria; Gainesville, Florida, US; Ljubljana, Slovenia; City of Singapore, Singapore;
533
20
Antofagasta, Chile; Santiago, Chile; Punta Arenas, Chile; Berlin, Germany; Turin, Piemonte, Italy;
534
Catania, Sicily, Italy; Palermo, Italy; Denver, Colorado, US; and Alexandria, Egypt).
535
536
537
538
539
540
541
542
543
544
545
546
Fig. 6. World map showing the distribution and the magnitudes of geographical locations that were considered in the reviewed
547
studies.
548
549
5.2. Machine Learning Algorithms
550
5.2.1. Problem Types
551
The criterion of Machine Learning Algorithms can be realized through problem types, selected MLAs
552
along with their hyperparameters, input parameters and output parameters. Regarding the problem
553
types, 3 categories are recognized in previous studies (Regression, Classification and Clustering). (Fig.
554
5-d) summarizes their distribution, showing that 87% of the reviewed studies sought to solve regression
555
problems, but only 11% and 2% of them examined classification and clustering, respectively.
556
557
5.2.2. Selected Machine Learning Algorithms and Hyperparameters
558
According to publishing year, (Fig. 7) reveals the distribution of attempts to use MLAs to predict internal
559
daylighting in previous studies. Until 2012, only 4 MLAs have been used (ANN, SVM, AR, DT) to solve
560
regression problems, in addition to one attempt to classify illumination categories using NB (Ahmed et
561
al., 2011a). Later, 6 more MLAs were additionally employed to solve regression, classification and
562
clustering problems (MLR, RF, LSMR, DBSCAN, k-NN and Gaussian Processes (GPs)). (Fig. 5-e)
563
illustrates their distribution of application, where almost half of previous studies used ANN to predict
564
internal daylighting, whereas 13%, 11%, 7% and 6% utilized MLR, SVM, AR and DT, respectively. As
565
stated earlier, the research scope focuses on MLAs that offer accelerated approximations, rather than
566
other optimization algorithms that take a longtime to find global optima; thus, only one MLA was cited
567
from (Conraud-Bianchi, 2008), although ANN and GA have been actually used. Overall, 37% of the
568
studies sought to compare two or more MLAs against prediction accuracy, speed, among other criteria.
569
They include (Kurian et al., 2008, 2006) who compared ANN and AR; (Ahmed et al., 2011a, 2011b)
570
who compared NB, DT and SVM; (da Fonseca et al., 2013) who compared ANN and MLR; (Colaco et
571
al., 2014) who compared ANN and AR; (Zhou and Liu, 2015) who compared ANN and SVM;
572
(Chatzikonstantinou and Sariyildiz, 2016) who compared ANN, SVM, RF and k-NN; (Ahmad et al.,
573
2017) who compared ANN and RF; and (Nault et al., 2017) who compared MLR and GPs.
574
21
575
Considering the used hyperparameters, 35% of the reviewed studies either did not mention the method
576
to tune the used MLAs before training begins, or did not use any hyperparameters at all. The remaining
577
65% considered 3 groups of hyperparameters according to the used MLA (ANN: number of Hidden
578
Neurons, Hidden Layers, Epochs, Activation Function and Learning Rate; SVM: Kernel Functions, C
579
Parameter and Width of RBF Kernel; RF: number of Trees and Features). (Fig. 5-f) demonstrates their
580
overall distribution, and since ANN is the most used MLA, its hyperparameters are also the most
581
frequently utilized, where number of Hidden Neurons, Hidden Layers, Epochs, Activation Function and
582
Learning Rate were employed with frequency rates of 21%, 11%, 10%, 5% and 3%, respectively.
583
584
585
586
587
588
589
590
591
592
593
Fig. 7. Distribution of the previous attempts to use MLAs to predict internal daylighting according to publishing year.
594
595
5.2.3. Input Parameters
596
As mentioned before, MLAs can predict internal luminous conditions based on a set of input parameters.
597
The reviewed studies show that the identified parameters are categorized into 2 major groups according
598
to the input data: External and Internal; each of which can be subcategorized into 3 sets of parameters:
599
(Climate Conditions; Temporal Settings; External Obstructions) and (Building Physical Features;
600
Openings and Shading Devices; Occupancy and Sensor Data), respectively (Fig. 8). While the
601
investigational purposes differ from study to another, a total of 48 input parameters can be identified
602
that were selected to construct prediction models, based on the available data. (Fig. 5-g) shows their
603
overall distribution.
604
605
606
607
608
609
610
611
612
613
Fig. 8. The identified different sets of input parameters to predict internal daylighting
614
615
22
Regarding the (i) External Input parameters, (i-i) Climate Conditions include dry bulb temperature,
616
relative humidity, wind speed, UV index, UV dose, global horizontal radiation, direct normal radiation,
617
diffuse horizontal radiation, horizontal infrared radiation intensity from sky, global horizontal illuminance,
618
diffuse horizontal illuminance and exterior horizontal beam illuminance. (i-ii) Temporal Settings include
619
time of the day, day of the year, solar altitude, solar azimuth, solar declination and solar hour. (i-iii)
620
External Obstructions include obstruction height, obstruction length, obstruction width, obstruction
621
angle, obstruction material and distance from obstruction. Considering the (ii) Internal Input parameters,
622
(ii-i) Building Physical Features include orientation, room width, room length, room height, room
623
elevation, work plane height, wall material, floor material and ceiling material. (ii-ii) Openings and
624
Shading Devices include window width, window height, windows location, number of windows, glazing
625
properties, shading device and shading parameters. (ii-iii) Occupancy and Sensor Data includes
626
occupancy schedule, sensor point identification, distance from windows, floor/room identification, view
627
direction, pixel location, pixel luminance and internal illuminance.
628
629
In general, it can be concluded that the usage frequency of the external input parameters to perform
630
predictions by the investigated attempts is 39%, while the frequency of the internal parameters is 61%.
631
Specifically, Climate Conditions have a noteworthy impact on the internal daylighting. As they can easily
632
be acquired from representative weather data files, they accounted for 19% of the used input
633
parameters. Temporal Settings and External Obstructions, though important, represented only 11% and
634
8%, respectively. On the other hand, Building Physical Features, as they immediately affect the internal
635
daylighting, comprised more than 30% of the input parameters, but Openings and Shading Devices, as
636
well as Occupancy and Sensor Data, constituted 22% and 9%, respectively. Of all the input parameters,
637
the most frequently used in the previous studies are: orientation, window width, window height, room
638
width, room length, room height, global horizontal radiation, time of the day, diffuse horizontal radiation,
639
windows location, sensor point identification, shading device, shading parameters, wall material, floor
640
material, ceiling material and glazing properties.
641
642
5.2.4. Output Parameters
643
Considering the output parameters, they can be categorized into 2 main types: (Absolute Illuminance
644
Values and Performance Metrics); both were predicted almost equally in previous studies, with
645
frequencies of 53% and 47%, respectively. For clarity, the term ‘Performance Metrics’ is used in this
646
review to signify daylight metrics, while ‘Evaluation Metrics’, which will be discussed later, denotes the
647
accuracy measures used to assess how well MLAs models would perform. (Fig. 5-h) summarizes the
648
overall distribution of output parameters, confirming that nearly 13%, 11%, 7%, 4%, 4%, 4%, 2% and
649
2% of the reviewed attempts focused on predicting DA, sDA300/50%, Daylight Glare Probability (DGP),
650
Daylight Factor (DF), ASE1000/250h, UDI, Continuous Daylight Autonomy (CDA) and CIE Glare Index
651
(CGI), respectively. According to publishing year, (Fig. 9) reveals the distribution of previous attempts
652
to predict daylighting in terms of absolute illuminance values and performance metrics.
653
654
655
23
656
657
658
659
660
661
662
663
664
665
Fig. 9. Distribution of the previous attempts to predict daylighting in terms of absolute illuminance values and performance
666
metrics according to publishing year.
667
668
With regard to daylight metrics, Trotter (1911) introduced DF to assess daylighting, which has been
669
widely used since then. DF is the ratio between the received indoor illuminance (Ei) on a given plane to
670
the corresponding level of unobstructed outdoor illuminance (Eo), excluding the direct sunlight from both
671
illuminances (IES, 1972; Love, 1992), as follows:
672
(1)
673
Later, DA was proposed by Association Suisse des Electriciens in 1989, which was later improved by
674
Reinhart and Walkenhorst (2001). DA is the percentage of the annual occupied timesteps when the
675
illuminance exceeds a predefined threshold:
676
,
(2)
677
where (ti) is the occupied hour of the year, (w.fi) is a weighting factor that depends on (EDaylight) and
678
(ELimit) that are the horizontal illuminance at a given point due to daylight and the illuminance limit value,
679
respectively. It is the first metric to account for dynamic temporal occurrences of daylight due to
680
changing climate conditions. Still, DA does not consider target illuminances neither below nor above
681
that threshold, lacking the indications of higher lighting energy demand, or visual discomforts due to
682
sunlight penetration. Rogers and Goldman (2006) proposed CDA by assigning a partial credit to
683
illuminances below a suggested limit. CDA is the percentage of the annual occupied timesteps when
684
the illuminance is over or under a predefined threshold:
685
,
(3)
686
Similarly, UDI was developed by Nabil and Mardaljevic (2006), which is the first metric to address
687
different ranges of daylighting performance by introducing the idea of ranged thresholds. UDI is the
688
percentage of the annual occupied timesteps when the illuminance is useful (within the lower and upper
689
limits), underlit (< lower limit), or overlit (> upper limit):
690
24
,
(4)
691
The approved method of daylight metrics IES LM-83-12 (2012), issued by IES in 2012, then expanded
692
those temporal metrics with sDA300/50% and ASE1000/250h to account for spatial consideration. sDA300/50%
693
and ASE1000/250h describe how much percentage of a space receives adequate amount of daylight, and
694
excessive amount of sunlight, respectively. sDA300/50% is the percentage of a space that receives a
695
minimum target illuminance of 300 lx for temporal fraction threshold of 50% of the annual occupied
696
hours:
697
,
(5)
698
S(i) represents a function, where si is the occurrences exceeding the illuminance target at points (pi),
699
is the temporal fraction threshold and ty is the annual occupied timesteps. ASE1000/250h is the percentage
700
of a space that receives a minimum direct sunlight of 1000 lx for absolute hour threshold of 250 hours
701
of the annual occupied hours:
702
,
(6)
703
where A(i) represents a function, where ai is the occurrences exceeding the illuminance target at points
704
(pi) and Ty is the annual absolute hour threshold.
705
706
About glare indices, Einhorn (1979) introduced Glare Index, which that was later recommended by CIE,
707
referred to as CGI. It considers direct and diffuse light evaluated on a horizontal plane passing through
708
a viewpoint. Then, Wienold and Christoffersen (2006) modified the original glare formula by presenting
709
DGP, which combines vertical eye illuminance with glare source luminance assessed at the same point.
710
Like most glare indices, DGP requires the source luminance, size and relative position, as follows:
711
(7)
712
where (Ls) is luminance of the glare source luminance, (Lb) is the luminance of the background or field
713
of view from the observer’s position, (ωs) is the angular size of the glare source from the observer’s
714
position, (P) is the relative position of the glare source regarding the observer’s focal point, and
715
exponents (e), (f) and (g) are weight factors that vary according to glare formulas.
716
717
25
5.3. Data Sources, Sizes and Temporal Granularities
718
5.3.1. Data Sources and Sizes
719
The used Data for training and testing machine learning models can be categorized into Data Sources
720
and Sizes, along with Temporal Granularities. Considering the data sources, 2 types can be found in
721
previous studies (Simulation-Derived and Field Measurements). (Fig. 5-i) summarizes their distribution,
722
indicating that 73% of the studies focused on simulation-derived data, while the remaining collected
723
data from field measurements, including HDRI (Inanici, 2013) and experimental survey (Yacine et al.,
724
2017). The Simulation-derived data is used to construct machine learning models, considering both
725
existing spaces and conceptual designs, using simulation tools, such as DIVA-for-Rhino, Radiance,
726
DAYSIM, Ecotect, Honeybee, Groundhog, OpenStudio, EnergyPlus, DElight and ESP-r. (Fig. 5-j)
727
reveals their distribution, where they were used in previous investigational attempts with frequencies of
728
19%, 15%, 15%, 12%, 12%, 9%, 6%, 6%, 3% and 3%, respectively. The reader can refer to (Ayoub,
729
2020, 2019b) for additional information on daylight simulation tools capabilities and the used light
730
transport algorithms. Regarding the data sizes, they can be classified into 4 types, according to data
731
sources and their usage (Simulation-Derived: Training and Testing Data, in addition to Field
732
Measurements: Training and Testing Data). Overall, the ranges of data size for Simulation-Derived
733
Training Data are from 5,400 to 195, Simulation-Derived Testing Data are from 729 to 21, Field
734
Measurements Training Data are from 906,540 to 50, and Field Measurements Testing Data are from
735
140,218 to 31, as shown in (Table 1). (Fig. 10) plots the reviewed studies against data sizes, distributed
736
over a logarithmic scale. It can be concluded that the average ratio between training and testing data is
737
approximately 3:1.
738
739
740
741
742
743
744
745
746
747
748
749
750
Fig. 10. Distribution of data sizes used to train and test MLAs in previous studies according to author and publishing year.
751
752
5.3.2. Temporal Granularities
753
Temporal granularity is the time of the year and its resolution to consider for a building performance
754
indicator, either a metric or a set of absolute values. In the reviewed studies, they are categorized into
755
2 main types (Annual and Selected Instances); including 3 subcategorical resolutions (Hourly, Sub-
756
Hourly and Timeseries). (Fig. 5-k) summarizes their overall distribution, showing that almost 43%, 19%,
757
11%, 9%, 8% and 6% of the studies accounted for Annual (Hourly Basis), Selected Instances (Sub-
758
26
Hourly Basis), Selected Instances (Hourly Basis), Annual Timeseries (Hourly Basis), Annual (Sub-
759
Hourly Basis) and Selected Timeseries (Sub-Hourly Basis), respectively; but the remaining studies did
760
not identify the selected temporal granularity. 57% of the reviewed attempts considered Annual
761
predictions, while 38% of them focused on Selected Instances. Considering the temporal resolutions,
762
Hourly resolutions were considered 62% of the attempts, whereas Sub-Hourly were accounted for 34%.
763
764
5.4. Evaluation Metrics
765
Evaluation metrics are accuracy measures used to assess how well machine learning models’ training
766
and testing would perform; and this can be achieved by comparing the differences between the
767
predicted and the observed/simulated values. 12 evaluation metrics are identified in the previously
768
reviewed studies, including Root Mean Square Error (RMSE), Percentage of Error (PE), Coefficient of
769
Determination (R2), Mean Square Error (MSE), Predictive Confidence (PC), Mean Biased Error (MBE),
770
Relative Error Rate (RER), Average Relative Error (ARE), Coefficient of Correlation (R), Coefficient of
771
Variation (CV), Mean Absolute Deviation (MAD) and Mean Absolute Error (MAE). Some of the famous
772
metrics can be acquired from equations (8) to (14).
773
(8)
774
(9)
775
(10)
776
(11)
777
(12)
778
(13)
779
(14)
780
where is the predicted illuminance value for times ; is the observed/simulated illuminance value
781
for times ; is the average illuminance value; and is the number of data points used for evaluation.
782
(Fig. 5-l) summarizes their distribution, indicating that 23%, 19%, 16% and 12% of the reviewed studies
783
used RMSE, PE, R2 and MSE. About 14% of them utilized PC, MBE and RER with the same frequency
784
rate of nearly 5%; while almost 12% used ARE, CV, MAD, MAE and R with the same frequency of about
785
2%. Other researchers, including (Liu et al., 2015; Uribe et al., 2017), did not specify the used evaluation
786
metrics.
787
788
27
6. Discussion
789
Next subsections include a discussion of the review results, organized in accordance with the same
790
structural order of the previous section: Scope of Prediction; MLAs; Data Source, Size and Temporal
791
Granularities; and Evaluation Metrics.
792
793
6.1. Scope of Prediction
794
6.1.1. Building Types
795
In daylighting studies, selecting a specific building type is crucial to determine the applied occupancy
796
schedules and the targeted illuminance levels. Occupancy schedules normally define the hours of the
797
year when a space is occupied by users; hours outside such ranges are not considered in the evaluation
798
of daylighting performance. LEED v4 (USGBC, 2013) and LM-83 (IES, 2012) specify typical occupancy
799
schedules from 8:00 AM to 6:00 PM, including all weekdays. This represents common working hours
800
from 9:00 AM to 5:00 PM, in addition to one hour for arrival and another for departure. Targeted daylight
801
levels vary according to building types and functions in a way that significantly affects the simulation
802
outputs. For suggestions on adequate illuminance levels, the readers are referred the handbooks of the
803
Illuminating Engineering Society of North America (IESNA) (IESNA, 2018) and the American Society of
804
Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) (ASHRAE, 2007).
805
806
While residential buildings represent a large part of building energy consumption in many countries
807
around the world (Pérez-Lombard et al., 2008), recent reviews, such as (Wang and Srinivasan, 2017)
808
and (Amasyali and El-Gohary, 2018), confirmed that more than 80% of the reviewed researches
809
focused the investigations extensively on non-residential buildings to predict energy consumption by
810
MLAs. This coincides with a much higher percentage, 98%, of the investigated studies in this review
811
that focused on predicting the luminous conditions of non-residential buildings. Such tendency may be
812
attributed to several reasons. (i) Due to the issue of privacy (Jetcheva et al., 2014), it is not easy to
813
regularly collect building performance data from residential buildings. Whereas, the relatively accessible
814
public spaces, such as offices and educational buildings, encourage researchers to investigate those
815
types of buildings instead. (ii) Many test models actually resemble reference office spaces (Reinhart et
816
al., 2013), to achieve some sort of standardized evaluations, allowing to facilitate comparing results
817
from different studies. Still, it is worth mentioning that during the early stages of design, commonly used
818
3D ‘shoe-box’ test models (Reinhart et al., 2013) that represent standalone buildings do not always
819
reflect reality. In fact, it is more accurate to explicitly account for the actual space geometry (Ayoub,
820
2019b), surrounding obstructions (Ayoub, 2019a) and external ground, since they have a great impact
821
on daylighting performance (Ayoub, 2019b; Ng, 2001; Reinhart and Walkenhorst, 2001). Nonetheless,
822
such simple models may still be satisfactory for initial estimations using MLAs to achieve generalization
823
in such conceptual stage of design. (iii) Specifically, energy-based predictions using machine learning,
824
as shown in previous reviews, depend largely on a derived data from field measurements and historical
825
datasets. However, this is rarely found in daylighting studies, where many researchers depend mostly
826
on simulation-derived data rather than scarce measurements. (iv) Perhaps the focus on non-residential
827
buildings might stem from researchers who seek to simplify data acquisition by minimizing the time
828
28
required for annual simulations; since offices and educational spaces have limited and pre-determined
829
occupancy schedules, while residential buildings are regularly occupied, requiring to perform annual
830
simulations on an hourly basis for all daylit hours; adding more stress to the computations. The more
831
hours of the day are considered, the more stress is added to the computational power, especially when
832
the target is to obtain accelerated preliminary approximations for different design solutions.
833
834
6.1.2. Climate Zones and Locations
835
The inequitable distribution of cities and climate zones in which the studies reflected (Fig. 6) reveals
836
some opposing concerns. For instance, in cooling-dominant climates, such as Hot-Desert, solar heat
837
gain due to excessive sunlight is inevitable, and has long been regarded as a major driver for energy-
838
efficient building envelops (Ayoub, 2019a, 2018; Elnokaly et al., 2019). The need to decrease heat
839
gains, by minimizing window-to-wall ratio, contradicts with the necessity to maintain adequate level of
840
daylight, by maximizing such ratio (Ayoub, 2018; Xue et al., 2016). Moreover, the precise identification
841
of the study location is of great importance to select adequate weather dataset for building performance
842
simulations. Typically, weather files are metrological data, collected from weather stations, and include
843
single-years that represent historical records of specific locations, for different temporal resolutions,
844
over extended observation periods (Herrera et al., 2017; Moazami et al., 2019). Nonetheless, most of
845
the reviewed studies did not explicitly identify the used weather file, but rather stated the location; and
846
this may entail some confusions. (i) Form which the weather files are created, it is typical to select the
847
nearest weather station to the project location, although there might not necessarily be a nearby station.
848
In some locations, such as California that has 16 different climate zones (EnergyPlus, 2019), few
849
kilometres can make a difference (Reinhart, 2018). Thus, understanding the local weather conditions is
850
needed to select suitable files. (ii) Also, selecting files of unobstructed weather stations, typically in
851
airports, may be unsuitable for daylighting studies in heavily obstructed context, where many weather
852
elements, such as solar irradiances and wind velocities, can significantly be affected by such settings
853
(Ayoub, 2019b). However, studies that rely on simulation-derived data can adjust the used weather files
854
by additional field measurements (Charles and Crawley, 2011), expensive satellite data, or by machine
855
learning to generate synthetic data (Chauhan and Thakur, 2014).
856
857
6.2. Machine Learning Algorithms
858
6.2.1. Problem Types
859
The majority of reviewed studies tend to investigate regression problems, and this is quite justifiable,
860
as the non-linear nature of internal daylight patterns (Lorenz and Jabi, 2017; Suykens et al., 2012) can
861
tolerably be handled by regression. This also corresponds with the intent to obtain numerical
862
estimations of illuminance values or performance metrics, rather than 2-dimensionally mapping the
863
spatial distribution of internal daylight at sensor points (Carlucci et al., 2015; Galatioto and Beccali,
864
2016). Classification and clustering problems were only addressed in (Ahmed et al., 2011a; Liu et al.,
865
2015; Zhou and Liu, 2015).
866
867
29
6.2.2. Selected Machine Learning Algorithms and Hyperparameters
868
As mentioned earlier, after 2012 many MLAs were used to solve regression, classification and clustering
869
problems (Fig. 7). From a daylight simulation perspective, 2012 checkpoint is particularly important to
870
the building design community, as it corresponds with the emergence of multi-core CPUs that were
871
derived by economic and thermodynamic limits (Boix, 2015; Koomey et al., 2011; Parkhurst et al.,
872
2006), at the expense of obsolete single-core designs. Such boost in the computational power was
873
accompanied by the development of: (i) test cases that are used to validate daylight simulation tools
874
(Aizlewood, 1993; CIE, 2006; Mardaljevic, 2001; Schregle and Wienold, 2004) ; (ii) weather datasets
875
that are representative of extreme climate conditions (CIBSE, 2019; Ferrari and Lee, 2008); (iii) 3- and
876
5-phase CBDM that brought the accurate and fast modelling of the optical properties of Complex
877
Fenestration Systems (CFS) via Radiance’s bi-directional raytracing (Lee et al., 2018; Saxena et al.,
878
2010); (iv) various dynamic metrics that report temporal daylight occurrences (Nabil and Mardaljevic,
879
2006; Rogers and Goldman, 2006); in addition to (v) a new interesting approach to estimate illuminance
880
from an HDRI (Gardner et al., 2017; Zhang and Lalonde, 2017) and 3D objects (Weber et al., 2018).
881
Not only such factors are bringing innovative daylighting studies, but also can offer new opportunities
882
for potential applications of machine learning in the field of building performance simulation.
883
884
Compared to engineering methods, machine learning does not require any prior knowledge about the
885
building physical attributes, including geometry, surrounding context, opening locations and
886
configurations, furniture and material properties. Such information is already embedded in the
887
parameters of the developed model, which in turn, are reflected in the prediction outputs. In this sense,
888
the training and testing datasets are derived from the same feature space and represent the same
889
distribution. To make accurate predictions using MLA, training relevant data is required to construct
890
mathematically-fit models; but they are only applicable in the data range from which they were
891
developed. Once a model is successfully constructed, it can be easily used for accelerated predictions.
892
However, during the early stages of design process, preliminary building designs are iteratively changed
893
to search for optimum alternatives; consequently, the data distribution is modified as well. This can
894
significantly affect the outputs, where it becomes difficult to generalize the daylighting performance for
895
new design proposals. In such case, machine learning models need to be entirely reconstructed and
896
retrained by newly collected set of data. To reduce the effort of such time-consuming work, Transfer
897
Learning between the old and new data domains offer a solution for this complication (Pan and Yang,
898
2009). Another discrepancy that is associated with using MLAs themselves as holistic approach of
899
univariate model output is that they can be tedious, where every building performance output would
900
entail creating a dedicated algorithm. This is apparent in the reviewed studies and the larger number of
901
investigational attempts they involve.
902
903
The choice of suitable MLAs depends on the problem that is being addressed (Wang and Srinivasan,
904
2017), the collected data type, and model-specific criteria that include accuracy, complexity and speed.
905
However, some decisions that were made in previous studies to use a single algorithm, or compare a
906
number of algorithms, were not supported neither by technical aspects, nor by heuristic methods. The
907
30
latter method, while important, is time-consuming, since it requires an extensive empirical trial and error
908
(Lokuciejewski et al., 2010), coupled with a higher degree of experience to be handled properly. Still,
909
machine learning is not commonly practiced by architects (Khean et al., 2018); but with more focused
910
implementations of machine learning in the field, such lack of integration will eventually diminish.
911
912
6.2.3. Input Parameters
913
Apart from (Chatzikonstantinou and Sariyildiz, 2016; Radziszewski and Waczyńska, 2018; Zhou and
914
Liu, 2015), previous studies that predicted annual metrics did not use any Temporal Settings to
915
construct their machine learning models, as such annual data is already embedded in the parameters
916
of the developed models. This coincides with their selected temporal granularity of Annual (Hourly and
917
Sub-Hourly Basis), except for (Yacine et al., 2017), where the authors sought to create models to study
918
the association between lighting quality and glare indices, only for selected timesteps. This is because
919
annual glare studies require creating a hemispherical fisheye image at each timestep, yielding
920
thousands of images per year, putting additional stress on the computations, especially for multiple
921
viewpoints (Ayoub, 2019b). As stated earlier, although the External Obstructions are essential inputs to
922
consider, as they have a profound impact on the internal daylighting conditions (Ayoub, 2019b; Ng,
923
2001; Reinhart and Walkenhorst, 2001), they were not accounted for that much in previous studies.
924
This can be attributed to the aim to achieve generalization, where simple ‘shoe-box’ models are
925
sufficient for such initial estimations in the early stages of design.
926
927
On the other hand, some of the reviewed studies addressed a small number of input parameters, such
928
as (Colaco et al., 2014; Conraud-Bianchi, 2008; da Fonseca et al., 2013; Kurian et al., 2008; Logar et
929
al., 2014; Yacine et al., 2017). This suggests that they represent isolated cases of which the developed
930
machine learning models are only applicable on, and do not offer generalization. Yet again, even with
931
larger amount of input parameters, many studies did not test the predictive power of the selected inputs.
932
In fact, machine learning models can be constructed from any number of input parameters. However,
933
this does not mean to add as much inputs as possible to the model, since more inputs, while can offer
934
more accuracy, can come on the expense of interpretability. In contrast, less inputs, while can ensure
935
interpretability, but might yield less accuracy and increase the risk of overfitting. Thus, Feature Selection
936
(Kira and Rendell, 1992) can be used to identify those that would contribute to the model construction,
937
while increasing its prediction accuracy and interpretability by discarding insignificant inputs. Evaluation
938
metrics are also used to gauge different combinations of inputs. For clarity, the term (input parameters)
939
refers to the original input data, but the term (feature) denotes those that are identified from the feature
940
selection. When the number of parameters is small, searching all possible combinations of features via
941
brute-force can be used to find the best set. But with higher dimensional datasets, it might be inefficient
942
and time-consuming to go through an even larger number of inputs combinations; especially this would
943
require repeating the model training with every new combination. Thus, other feature selection methods
944
can be used, including wrappers, filters and embedded approaches (Chandrashekar and Sahin, 2014;
945
Vergara and Estévez, 2014). In this sense, feature selection explores the relationship among inputs
946
rather than their value ranges. On the contrary, Sensitivity Analysis (Davis, 1989; Saltelli et al., 2004)
947
31
investigates the relationship between inputs, together with a model's architecture, and outputs. In other
948
words, it measures the variation in outputs due to uncertainty in various ranges of inputs and model
949
architectures (Kazanasmaz et al., 2009). Sensitivity analysis can explore the variation in the outputs
950
when inputs vary around their original values, or can account for a wider range of inputs values. Using
951
both of those tests can help designers to (i) realize of the relationship between input and output
952
parameters; (ii) identify the most influential parameters on the predicted outputs; (iii) detect sources of
953
prediction errors that may stem from unexpected relationships between inputs and outputs; and (iv)
954
offer further simplification and interpretability to the developed models.
955
956
6.2.4. Output Parameters
957
Apart from DF, the oldest daylight metrics, researchers did not actually start predicting performance
958
metrics until 2015, as shown in (Fig. 9). This is rather surprising, since the performance metrics that
959
considered temporal occurrences of horizontal illuminances were basically originated back to 1989.
960
This can be attributed to the desire to acquire absolute illuminance values before 2015, rather than
961
performance metrics; especially in studies that considered temporal granularity of Annual Timeseries,
962
such as (Kurian et al., 2008, 2006), as well as the ones that accounted for Selected Instances that
963
include (Ahmed et al., 2011a, 2011b; Binol, 2008; Colaco et al., 2014; Hu and Olbina, 2011; Inanici,
964
2013; Kazanasmaz et al., 2009; Logar et al., 2014). The majority of the latter cases depended on Field
965
Measurements to collect the training data, as will be explained in the following subsections. Between
966
2008 and 2015, there was no research on MLAs that focused on predicting daylighting performance
967
metrics, and this can be attributed to two distinct stages of significant developments: before and during
968
that period. (i) Before 2008, during the early years of the new millennium, the first stage has witnessed
969
the progress of multi-core CPUs, coupled with the emergence of validation datasets and test cases
970
(Aizlewood, 1993; CIE, 2006; Mardaljevic, 2001; Schregle and Wienold, 2004) , Standard General
971
Sky (CIE, 2003; ISO, 2004), weather datasets of extreme climate conditions (CIBSE, 2019; Ferrari and
972
Lee, 2008), CBDM (Mardaljevic, 2000b; Reinhart and Herkel, 2000) and dynamic metrics that report
973
temporal daylight occurrences (Nabil and Mardaljevic, 2006; Rogers and Goldman, 2006). (ii) During
974
the second stage that started by 2010, the growth of Graphics Processing Units (GPUs) and parallel
975
computation offered unprecedented real-time prediction of building performance (Jones and Reinhart,
976
2015; McNeil and Lee, 2012; Schardl, 2016). This was coupled by the development of more accurate
977
weather datasets, including International Weather for Energy Calculations (IWEC2) (Joe et al., 2014)
978
and Typical Meteorological Year (TMY3) (Charles and Crawley, 2011), 3- and 5-PM of CBDM (Lee et
979
al., 2018; Saxena et al., 2010) and daylight metrics that consider the spatiotemporal evaluations of
980
daylighting performance (IES, 2012). Thus, it took the research community a few years employ those
981
accelerating developments and MLAs to predict daylighting performance metrics.
982
983
Several daylight simulation tools are now supporting a wide range of daylight metrics to report different
984
daylighting performance of various purposes and target values. They capture the spatial and temporal
985
illuminance conditions and encapsulate such information into single values. This seems convenient,
986
particularly in cases that involve annual predictions with thousands of hourly, or sub-hourly, output data.
987
32
Still, relying on daylight metrics as outputs for machine learning models yields some inconsistencies.
988
Some of the used daylight metrics in previous studies confusingly do not specify exact threshold values.
989
For example, the definition of DA does not state a specific threshold, but it was suggested to set it at
990
500 lx (Olbina and Beliveau, 2009). Also, there is no consensus on the illuminance limits for UDI, though
991
many researchers suggested different upper and lower thresholds (David et al., 2011; Mardaljevic et
992
al., 2009; Nabil and Mardaljevic, 2005; Olbina and Beliveau, 2009). The choice of certain thresholds
993
has a significant impact on how the resulting metric value would be. For instance, DA reports the
994
percentage of the annual occupied timesteps when the internal illuminance exceeds a predefined
995
threshold. But if this threshold is changed, the resulting DA value changes as well, given that the internal
996
daylighting condition remains the same in both cases; and this can yield misleading results. Thus, such
997
metrics are not consistent with each other, and do not necessarily yield common resolutions, requiring
998
data post-processing to retrieve the original illuminance values. Alternatively, absolute illuminance
999
values can be more specific in machine learning field of application, especially in cases that require
1000
regression; reflecting accurate representation of the study space, rather than indicating the daylighting
1001
performance via metrics that rise above or fall below given thresholds. In other words, the process of
1002
assessing daylighting by performance metrics should be after predicting absolute illuminance values.
1003
In this, (Ayoub, 2018) suggested an interesting approach to convert absolute illuminances into their
1004
equivalent metrics, and vice versa.
1005
1006
6.3. Data Source, Size and Temporal Granularities
1007
6.3.1. Data Sources and Sizes
1008
Field measurements are real data collected from sensors, HDRI or surveys. In building performance
1009
simulation, they are more accurate than scale models, which may introduce uncertainties related to
1010
geometries and photometric properties of materials (Maamari and Fontoynont, 2003; Thanachareonkit
1011
and Scartezzini, 2010). Therefore, the latter method of data collection was not used in previous studies.
1012
Field measurements typically take a long time to collect, and they become inefficient to investigate long-
1013
term analysis under a variety of sky conditions. As a result, this data collection approach was only used
1014
in studies that accounted for selected timesteps under limited sky conditions. However, such
1015
measurements can be useful to examine post-occupancy performance of existing buildings after
1016
construction, not during the conceptual design. On the other hand, long-term assessments of internal
1017
daylighting using simulation approach would require inputting a complex set of parameters, as indicated
1018
before. But they can easily produce estimations of extended periods of time that cover an entire year,
1019
under different sky conditions (Mardaljevic, 2000b). Several studies have confirmed that dynamic
1020
daylight simulations using DC method, coupled with backward raytracing (Ward, 1994; Ward and
1021
Shakespeare, 1998) and Perez All-Weather (Perez et al., 1993) can predict internal illuminance with
1022
errors below 20-25%, compared to the limited field observations (Mardaljevic, 2000a; Reinhart and
1023
Andersen, 2006; Reinhart and Breton, 2009; Reinhart and Walkenhorst, 2001). This error may seem
1024
high, but the sensitivity of human eye can hardly perceive changes in the internal illuminance that vary
1025
in a wide range of magnitudes over time (Reinhart, 2011).
1026
1027
33
2 of the used 10 simulation tools are in fact legacy simulation engines (Radiance and EnergyPlus) that
1028
are not easy to use due to their complicated settings and features (Ayoub, 2019b). This, while can be
1029
valued by researchers and skilled practitioners, hinders such tools from being casually utilized by
1030
unexperienced users (Reinhart and Fitz, 2006). Therefore, many of the newly developed tools are
1031
actually visual interfaces to those engines, without integrating some of their complex features. 7 of the
1032
used tools are Radiance-based (DIVA-for-Rhino, DAYSIM, Ecotect, Honeybee, Groundhog, ESP-r and
1033
OpenStudio). 4 of them include EnergyPlus engine (Honeybee, ESP-r, OpenStudio and DElight); and
1034
this allows to conduct daylighting and energy simulations, simultaneously. Radiance is among the oldest
1035
simulation tools, which is highly regarded as a standard tool by the building design community (Ayoub,
1036
2019b). For practical applications, both Radiance and its modified version, DAYSIM, have undergone
1037
heavy validations against measurements of real room and physical model (Jarvis and Donn, 1997),
1038
clear glazing and shading systems (Reinhart and Walkenhorst, 2001), internal illuminance (Reinhart et
1039
al., 2006), translucent panels (Reinhart and Andersen, 2006), measurements of goniophotometer
1040
(Schregle and Wienold, 2004), and 3-phase method to model CFS (McNeil and Lee, 2012). Radiance
1041
was also validated against test cases from the datasets of BRE-IDMB (Mardaljevic, 2000b, 1997, 1995)
1042
and CIE (Donn et al., 2007; Geisler-Moroder and Dür, 2008; Osborne, 2012).
1043
1044
While limited data sizes can seemingly be acquired rapidly, they only capture samples of minor effects
1045
on the investigated building performance. Therefore, machine learning models should reliably be built
1046
using sufficient amount of data; although this would require additional computational effort. Such
1047
datasets should also be representative, offering data distribution that is embedded with as much
1048
information as possible in the feature space for accurate predictions. It might be surprising that the
1049
average data sizes of field measurements that focused on illuminance values are much larger than
1050
those of simulation-derived that considered daylight metrics. The latter approach, as explained earlier,
1051
captures long-term illuminance conditions and encapsulates thousands of hourly, or sub-hourly, data
1052
entries into single values that reflect the building performance for an extended period. However, the
1053
uncertainty and non-linearity in simulation-derived calculations are unavoidable (Tregenza, 2017), due
1054
to the changing sky luminance distribution patterns that cannot be predicted with complete confidence.
1055
Daylighting studies should, still, rely on long-term analysis, which is imperative for inclusive decision-
1056
making. Thus, after 2015 (Fig. 10), almost all previous studies depended on daylight simulations to
1057
collect the required data, except for (Beccali et al., 2018; Navada et al., 2016; Yacine et al., 2017).
1058
1059
6.3.2. Temporal Granularities
1060
It can be concluded from (Table 1) that Annual predictions were basically related to estimating different
1061
daylight metrics using simulation-derived data of reduced sizes; apart from (Ahmad et al., 2017; Kurian
1062
et al., 2008, 2006; Liu et al., 2018) that predicted illuminance values. On the other hand, predictions
1063
made from Selected Instances were mostly associated with illuminance values, collected by field
1064
measurements of larger data sizes; apart from (Yacine et al., 2017) that predicted glare indices of DGP
1065
and CGI. Nonetheless, commonly used temporal granularities in previous studies primarily focus on the
1066
long-term analysis of occupied hours and daylit hours. The former is related to studies that consider
1067
occupancy schedules as a time basis, such as LEED v4 (USGBC, 2013) and LM-83 (IES, 2012); while
1068
34
the latter is associated with the study of always-occupied spaces, or when the space function changes
1069
from time to time, causing alterations to the occupancy schedule.
1070
1071
For whole building performance investigations, several daylight and energy simulation tools, such as
1072
DIVA-for-Rhino (Jakubiec and Reinhart, 2011) and Honeybee (Roudsari and Pak, 2014), generate
1073
occupancy schedules, where illuminance values can be linked directly to lighting energy consumption
1074
for the same temporal granularities, in addition to heating and cooling energies, to estimate electric
1075
energy reduction due to natural daylighting. However, this may cause an overestimated energy
1076
consumption, if the daylit hours exceed the actual occupancy schedule, in office spaces for instance.
1077
Therefore, more realistic occupancy schedules should be used instead for such studies. Another
1078
important issue related to selecting temporal granularities is the consideration of weather data. With
1079
regard to the same location, the impact of using weather files of different temporal resolutions and
1080
periods on building performance was previously investigated, confirming that reducing time intervals
1081
from hours into minutes should be considered (Walkenhorst et al., 2002). Also, weather fluctuations
1082
cause a significant impact as well, as short-term investigations showed unreliable results and minor
1083
effects on daylighting performance (Bellia et al., 2015a, 2015b; Crawley, 1998; Crawley and Lawrie,
1084
2015; Iversen et al., 2013; Yang et al., 2008).
1085
1086
6.4. Evaluation Metrics
1087
After constructing a machine learning model, evaluation metrics are used to assess its performance
1088
during training and testing. Evaluating models is an essential step to test their accuracy before
1089
predictions can be undertaken in actual applications. Most of the reviewed studies that addressed
1090
regression problems depended on RMSE and MSE. Both metrics emphasis on outliers and extreme
1091
error values by squaring the differences between the predicted and observed values before averaging
1092
them. Since the square term in RMSE and MSE, along with the absolute term in MAE, cancel negative
1093
errors, they do not indicate whether the model over- or underestimates the predictions. Alternatively,
1094
other metrics, such as PE and MBE, indicate both positive and negative errors.
1095
1096
Selecting an adequate error metric depends mainly on the way the resulting outliers and extreme errors
1097
will be treated. Every metric reveals different qualities about the predicted errors. For example, metrics,
1098
such as MBE and MAE, are relatively easy to be interpreted, but they do not make error values
1099
contribute as much to the overall model accuracy. Whereas the contribution of the square term in MSE
1100
makes low errors quadratically bigger than that of RMSE, penalizing the model for making erroneous
1101
predictions that slightly differ from their corresponding observations. Thus, the difference between MSE
1102
and MAE suggests a small error of large penalty or a larger error of moderate penalty (Twomey and
1103
Smith, 1995). However, as MSE values get larger, it becomes harder to understand how well the model
1104
performs; thus, RMSE is rather used in such situation, as it converts the squared error values into their
1105
initial unit, making them easier to be interpreted and compared with the original observations. In this
1106
sense, RMSE, as it aggregates the magnitudes of the errors in a single value, it is highly influenced by
1107
outliers; thus, they need to be removed from the dataset before using such metric.
1108
35
1109
7. Recommendations and Future Research Trends
1110
Based on the reviewed studies, and in attempt to point towards knowledge gaps and missing research
1111
opportunities, the summary of recommendations and future trends in the related domain includes:
1112
• Scope of Prediction:
1113
o It is recommended to pay more attention towards regularly occupied spaces, including residential buildings, where
1114
using simple ‘shoe-box’ models as a template is satisfactory for initial MLAs estimations to achieve generalization
1115
during the conceptual design stage.
1116
o More focus should be put to exploit classification and clustering, as they can be used in many situations that entail
1117
classifying daylighting conditions by UDI and DGP.
1118
o It is suggested to expand the use of MLAs to investigate challenging climate zones and harsh weather conditions,
1119
while explicitly identifying the studied location along with the used weather dataset file.
1120
o No study has considered investigating the impact of using MLAs to predict complex daylighting situations of CFS.
1121
Such systems, including optical light redirecting systems, mirrored louver systems, prismatic window films and
1122
angular screens, can be simulated via 3- and 5-phase analysis (Lee et al., 2018; Saxena et al., 2010). They can
1123
play an important role as energy-efficient techniques to optimize façades designs, especially for building in dense
1124
contexts and highly obstructed built environment; and predicting those system’s performance using MLAs would
1125
be useful.
1126
o Again, there is a clear gap between the number of research efforts that has been developed for CBDM and the
1127
ones that addresses the applications of advanced types of CBDM (3- and 5-phase); and this needs to be
1128
addressed in future investigations.
1129
o With the discovery of specialized photoreceptors in human eye (Pauley, 2004), to the author’s best knowledge,
1130
no study has examined yet the impact of using MLAs to predict non-visual effects of daylighting (Andersen et al.,
1131
2012; Mardaljevic et al., 2014; Webb, 2006), which is a vital element of healthy living, responsible for regulating
1132
the internal circadian system.
1133
o As discussed earlier, the use of HDRI is a promising approach to estimate illuminance (Gardner et al., 2017;
1134
Zhang and Lalonde, 2017). Although such approach offers accuracy and timesaving, it needs more studies, as it
1135
only captures instantaneous conditions, not giving a future perspective of the total daylighting performance.
1136
• Machine Learning Algorithms:
1137
o To reduce the effort and consumed time to retrain machine learning models due to changes in design components
1138
during the conceptual design phase, Transfer Learning can be exploited, where the originally developed model
1139
can be reused as the starting point for additional tasks.
1140
o As explained earlier, machine learning is not commonly practiced by architects, and it is urged to incorporate its
1141
concepts and ideas into Architectural education. machine learning can have a significant role to play not only to
1142
surrogate building performance simulations, but also to quantify and classify large amounts of data, support form-
1143
finding and optimize different building components for better energy-efficiency.
1144
o Feature Selection and Sensitivity Analysis should always be incorporated to determine which set of inputs affects
1145
daylighting the most, in addition to examining the effect of adjusting weather files, and which weather elements
1146
influence the outcomes. Those might be investigated previously (Wang et al., 2019), yet their impact on MLAs
1147
has not yet been investigated.
1148
o It is also recommended to integrate absolute illuminance values when using MLAs to predict internal daylighting
1149
performance. Arguably, for comparison and optimization, it is more convenient to use single value to evaluate
1150
different design scenarios in the same format.
1151
• Data Source, Size and Temporal Granularities:
1152
o It is suggested to use the same temporal granularities, preferably daylit hours in daylighting studies, so that
1153
different spaces can be compared, incorporating daily and seasonal changes, different climate zones, besides
1154
geographical locations and latitudes.
1155
36
o Long-term resolutions should always be considered, typically on an annual basis. Typical weather files constructed
1156
from most recent 20-30 years of observed data, at least 8 years of observed data (Hubbard et al., 2004), can be
1157
representative of future situations of the following decade (Eames et al., 2012). Yet again, harsh climate zones of
1158
extreme weather excitations (Kalamees et al., 2012; Pernigotto et al., 2014) buildings may encounter should be
1159
integrated into typical weather files (CIBSE, 2009; Jentsch et al., 2013; Mylona, 2012).
1160
o To reduce the uncertainty of daylight calculations that stem from continuously changing sky luminance distribution
1161
patterns, the statistical significance of simulation results should explicitly be estimated, and possible changes to
1162
the building design or its surrounding context should also probabilistically be predicted (Tregenza, 2017).
1163
o Other growing research trends that are related to the study of energy-efficient daylighting strategies are occupancy
1164
behavior and control systems (Beccali et al., 2018; Hu and Olbina, 2011), which have a major impact on improving
1165
the prediction performance.
1166
1167
8. Conclusion
1168
Machine learning methods should be exploited more frequently in viable resolutions to surrogate
1169
complex daylight simulations. From the author’s viewpoint, this requires a mutual consideration of both
1170
education and practice standpoints. (i) Computational innovations revolutionize how Architecture is
1171
being practiced, as tools of building performance simulation, along with 3D modeling and representation
1172
courses, are commonly being introduced to students in many Schools of Architecture. Still, machine
1173
learning, statistical methods and optimization techniques unfortunately are not. (ii) Architects are now
1174
confronted with the urge to adopt new subdomains of Architecture that machine learning are already
1175
contributing to. Still, due to the lack of specialized experience, this shift may raise justified concerns
1176
among traditional practitioners to integrate what they consider complex methods into their practices.
1177
Besides, some practitioners, who got familiar with machine learning methods, might be unaccustomed
1178
to handle their difficulties with ease. Indeed, the amount of required training to master such techniques
1179
is considerable. This would require more integration of machine learning as early as possible in
1180
Architectural education before expecting an effective increase in their use in practice.
1181
1182
Daylight simulation is an active area of research in architecture that provides designers with accurate
1183
quantifications of internal luminous conditions. Nonetheless, such approach is time-consuming and
1184
computationally expensive, as it demands inputting a complex set of parameters, making it inherently
1185
inflexible, especially during the early stages of design, when preliminary assessment of daylighting is
1186
required. The review work of this research highlights an alternative approach to exploit the outputs of
1187
such complex computational simulations to construct predictive models that build on MLAs. The use of
1188
this approach is granting much attention from researchers over past years due to its ability to act as
1189
proxy to simulations. This is coupled with the increasing number of studies that have been published
1190
previously. Such growing directions consider exploiting different algorithms for both daylighting
1191
predictions as well as on glare evaluations. This research particularly focuses on outlining the scopes
1192
of prediction, the used algorithms, data sources and sizes, besides evaluation metrics, providing a
1193
detailed discussion on the findings of previous studies. Supported by lower error rates, the
1194
accumulations of the previously reviewed studies confirmed the ability of MLAs to offer fast and accurate
1195
predictions, when compared to complex simulations. Finally, knowledge gaps are also discussed,
1196
revealing future trends to use such approaches in the Architectural practice. However, this research
1197
37
domain is not mature enough, and more investigations in the related architectural applications are
1198
strongly recommended.
1199
1200
References
1201
Ahmad, A.S., Hassan, M.Y., Abdullah, M.P., Rahman, H.A., Hussin, F., Abdullah, H., Saidur, R., 2014.
1202
A review on applications of ANN and SVM for building electrical energy consumption forecasting.
1203
Renew. Sustain. Energy Rev. 33, 102–109.
1204
Ahmad, M.W., Hippolyte, J.L., Mourshed, M., Rezgui, Y., 2017. Random Forests and Artificial Neural
1205
Network for Predicting Daylight Illuminance and Energy Consumption, in: International Building
1206
Performance Simulation Association (IBPSA) 2017 Conference, California, United States.
1207
Ahmed, A., Korres, N.E., Ploennigs, J., Elhadi, H., Menzel, K., 2011a. Mining building performance data
1208
for energy-efficient operation. Adv. Eng. Informatics 25, 341–354.
1209
Ahmed, A., Otreba, M., Korres, N.E., Elhadi, H., Menzel, K., 2011b. Assessing the performance of
1210
naturally day-lit buildings using data mining. Adv. Eng. Informatics 25, 364–379.
1211
Aiserman, M.A., Braverman, E.M., Rozonoer, L.I., 1964. Theoretical foundations of the potential
1212
function method in pattern recognition. Autom. Remote Control 25, 917–936.
1213
Aizlewood, M.E., 1993. Innovative daylighting systems: An experimental evaluation. Light. Res.
1214
Technol. 25, 141–152.
1215
Aizlewood, M.E., Littlefair, P.J., 1996. Daylight prediction methods: A survey of their Use. Conf. Pap.
1216
CIBSE Natl. Light. Conf. 126–140.
1217
Akaike, H., 1969. Fitting autoregressive models for prediction. Ann. Inst. Stat. Math. 21, 243–247.
1218
Allouhi, A., El Fouih, Y., Kousksou, T., Jamil, A., Zeraouli, Y., Mourad, Y., 2015. Energy consumption
1219
and efficiency in buildings: current status and future trends. J. Clean. Prod. 109, 118–130.
1220
Amasyali, K., El-Gohary, N.M., 2018. A review of data-driven building energy consumption prediction
1221
studies. Renew. Sustain. Energy Rev. 81, 1192–1205.
1222
Andersen, M., Mardaljevic, J., Lockley, S.W., 2012. A framework for predicting the non-visual effects of
1223
daylight–Part I: photobiology-based model. Light. Res. Technol. 44, 37–53.
1224
ASHRAE, 2019. International Weather Files For Energy Calculations 2.0 (IWEC2) [WWW Document].
1225
URL https://www.ashrae.org/technical-resources/bookstore/ashrae-international-weather-files-
1226
for-energy-calculations-2-0-iwec2 (accessed 10.2.19).
1227
ASHRAE, 2007. ASHRAE Handbook of Fundamentals. American Society of Heating, Refrigerating and
1228
Air-Conditioning Engineers, Atlanta, Georgia.
1229
Athienitis, A.K., Tzempelikos, A., 2002. A methodology for simulation of daylight room illuminance
1230
distribution and light dimming for a room with a controlled shading device. Sol. Energy 72, 271–
1231
281.
1232
Ayoub, M., 2020. A Review on Light Transport Algorithms and Simulation Tools to Model Daylighting
1233
inside Buildings. Sol. Energy 198, 623–642.
1234
Ayoub, M., 2019a. A multivariate regression to predict daylighting and energy consumption of residential
1235
buildings within hybrid settlements in hot-desert climates. Indoor Built Environ. 28, 848–866.
1236
https://doi.org/326X18798164
1237
38
Ayoub, M., 2019b. 100 years of daylighting: A chronological review of daylight prediction and calculation
1238
methods. Sol. Energy 194, 360–390.
1239
Ayoub, M., 2018. Integrating illuminance and energy evaluations of cellular automata controlled
1240
dynamic shading system using new hourly-based metrics. Sol. Energy 170, 336–351.
1241
https://doi.org/10.1016/j.solener.2018.05.041
1242
Basheer, I.A., Hajmeer, M., 2000. Artificial neural networks: fundamentals, computing, design, and
1243
application. J. Microbiol. Methods 43, 3–31.
1244
Beccali, M., Bonomolo, M., Ciulla, G., Brano, V. Lo, 2018. Assessment of indoor illuminance and study
1245
on best photosensors’ position for design and commissioning of Daylight Linked Control systems.
1246
A new method based on artificial neural networks. Energy 154, 466–476.
1247
Beck, H.E., Zimmermann, N.E., McVicar, T.R., Vergopolan, N., Berg, A., Wood, E.F., 2018. Present
1248
and future Köppen-Geiger climate classification maps at 1-km resolution. Sci. Data 5.
1249
https://doi.org/https://doi.org/10.1038/sdata.2018.214
1250
Bellia, L., Pedace, A., Fragliasso, F., 2015a. Dynamic daylight simulations: Impact of weather file’s
1251
choice. Sol. Energy 117, 224–235.
1252
Bellia, L., Pedace, A., Fragliasso, F., 2015b. The role of weather data files in Climate-based Daylight
1253
Modeling. Sol. Energy 112, 169–182.
1254
Bellocchio, F., Ferrari, S., Lazzaroni, M., Cristaldi, L., Rossi, M., Poli, T., Paolini, R., 2011. Illuminance
1255
prediction through SVM regression, in: Environmental Energy and Structural Monitoring Systems
1256
(EESMS), IEEE Workshop. pp. 1–5.
1257
Ben-Hur, A., Horn, D., Siegelmann, H.T., Vapnik, V.N., 2001. Support vector clustering. J. Mach. Learn.
1258
Res. 2, 125–137.
1259
Binol, S., 2008. A prediction model for daylighting illuminance for office buildings. Izmir Institute of
1260
Technology, Izmir, Turkey.
1261
Bishop, C.M., 2006. Pattern recognition and machine learning. Springer, Singapore.
1262
Boix, C., 2015. Technological Progress, in: Political Order and Inequality. pp. 92–127.
1263
https://doi.org/10.1017/cbo9781316105504.004
1264
Boser, B.E., Guyon, I.M., Vapnik, V.N., 1992. A training algorithm for optimal margin classifiers, in:
1265
Proceedings of the Fifth Annual Workshop on Computational Learning Theory. ACM, pp. 144–
1266
152.
1267
Breiman, L., Friedman, J., Olshen, R., Stone, C., 1984. Classification and regression trees. Chapman
1268
and Hall, Wadsworth, New York.
1269
Broomhead, D.S., Lowe, D., 1988. Radial basis functions, multi-variable functional interpolation and
1270
adaptive networks.
1271
Burges, C.J., 1998. A tutorial on support vector machines for pattern recognition. Data Min. Knowl.
1272
Discov. 2, 121–167.
1273
Carlucci, S., Causone, F., De Rosa, F., Pagliano, L., 2015. A review of indices for assessing visual
1274
comfort with a view to their use in optimization processes to support building integrated design.
1275
Renew. Sustain. Energy Rev. 47, 1016–1033.
1276
Chakraborty, D., Elzarka, H., 2019. Advanced machine learning techniques for building performance
1277
39
simulation: a comparative analysis. J. Build. Perform. Simul. 12, 193–207.
1278
Chandrashekar, G., Sahin, F., 2014. A survey on feature selection methods. Comput. Electr. Eng. 40,
1279
16–28.
1280
Charles, B.S., Crawley, D.B., 2011. Weather Data for Building Performance Simulation, in: Hensen,
1281
J.L., Lamberts, R. (Eds.), Building Performance Simulation for Design and Operation. Spon Press,
1282
New York, United States, pp. 37–55.
1283
Chatzikonstantinou, I., Sariyildiz, S., 2016. Approximation of simulation-derived visual comfort
1284
indicators in office spaces: A comparative study in machine learning. Archit. Sci. Rev. 59, 307–
1285
322.
1286
Chauhan, D., Thakur, J., 2014. Data mining techniques for weather prediction: A review. Int. J. Recent
1287
Innov. Trends Comput. Commun. 2, 2184–2189.
1288
Chen, J.L., Liu, H.B., Wu, W., Xie, D.T., 2011. Estimation of monthly solar radiation from measured
1289
temperatures using support vector machines–a case study. Renew. Energy 36, 413–420.
1290
Cianfrani, C.A., Tsiakals, J.J., West, J., 2009. ISO 9001: 2008 explained. ASQ Quality Press.
1291
CIBSE, 2019. CIBSE - Weather Data [WWW Document].
1292
CIBSE, 2009. The use of Climate Change Scenarios for Building Simulation: the CIBSE Future Weather
1293
Years. London.
1294
CIE, 2006. Test cases to assess the accuracy of lighting computer programs. Vienna, Austria.
1295
CIE, 2003. Spatial distribution of daylight - CIE standard general sky.
1296
Colaco, S.G., Colaco, A.M., Kurian, C.P., George, V.I., 2014. An Adaptive predictive framework to
1297
online prediction of interior daylight illuminance, in: 2014 International Conference on Advances
1298
in Energy Conversion Technologies (ICAECT). pp. 174–180.
1299
Conraud-Bianchi, J., 2008. A Methodology for the Optimization of Building Energy, Thermal, and Visual
1300
Performance. Montreal, Quebec, Canada.
1301
Copping, B.W., 1987. CIBSE, applications manual: window design. London.
1302
Cramer, J.S., 2002. The origins of logistic regression. Amsterdam.
1303
Crawley, D.B., 1998. Which Weather Data Should You Use for Energy Simulations of Commercial
1304
Buildings? Trans. Soc. Heat. Refrig. air Cond. Eng. 104, 498–515.
1305
Crawley, D.B., Lawrie, L.K., 2015. Rethinking the TMY: is the “typical” meteorological year best for
1306
building performance simulation?, in: Conference: Building Simulation.
1307
Critchfield, H.J., 1983. Criteria for classification of major climatic types in modified Köppen system.
1308
da Fonseca, R.W., Didoné, E.L., Pereira, F.O.R., 2013. Using artificial neural networks to predict the
1309
impact of daylighting on building final electric energy requirements. Energy Build. 61, 31–38.
1310
David, M., Donn, M., Garde, F., Lenoir, A., 2011. Assessment of the Thermal and Visual Efficiency of
1311
Solar Shades. Build. Environ. 46, 1489–1496.
1312
Davis, G.W., 1989. Sensitivity analysis in neural net solutions. IEEE Trans. Syst. Man. Cybern. 19,
1313
1078–1082.
1314
De Cian, E., Lanzi, E., Roson, R., 2007. The impact of temperature change on energy demand: a
1315
dynamic panel analysis, No. I. Milano.
1316
DiLaura, D.L., Houser, K.W., Mistrick, R.G., Steffy, G.R., 2011. The lighting handbook: reference and
1317
40
application. Illuminating Engineering Society of North America, New York.
1318
Donn, M., Xu, D., Harrison, D., Maamari, F., 2007. Using simulation software calibration tests as a
1319
consumer guide-A geasibility study using lighting simulation software, in: Proceedings of the 10th
1320
Conference of International Building Performance Simulation Association (IBPSA). China, pp.
1321
1999–2006.
1322
Drucker, H., Burges, C.J., Kaufman, L., Smola, A.J., Vapnik, V.N., 1997. Support vector regression
1323
machines, in: In Proceedings of the Neural Information Processing Systems. pp. 155–161.
1324
Dufton, A.F., 1946. Protractors for the computation of daylight factors. D.S.I.R. building research
1325
technical paper no. 28.
1326
Eames, M.E., Kershaw, T., Coley, D.A., 2012. A comparison of future weather created from morphed
1327
observed weather and created by a weather generator. Build. Environ. 56, 252–264.
1328
Einhorn, H.D., 1979. Discomfort glare: A formula to bridge differences. Light. Res. Technol. 11, 90–94.
1329
Elman, J.L., 1990. Finding structure in time. Cogn. Sci. 14, 179–211.
1330
Elnokaly, A., Ayoub, M., Elseragy, A., 2019. Parametric investigation of traditional vaulted roofs in hot-
1331
arid climates. Renew. Energy 138, 250–262.
1332
EnergyPlus, 2019. Weather Data Sources, EnergyPlus. U.S. Department of Energy’s (DOE) Building
1333
Technologies Office (BTO), and managed by the National Renewable Energy Laboratory (NREL)
1334
[WWW Document].
1335
Ferrari, D., Lee, T., 2008. Beyond TMY: climate data for specific applications, in: Proceedings 3rd
1336
International Solar Energy Society Conference-Asia Pacific Region (ISES-AP-08).
1337
Fienberg, S.E., 2006. When did Bayesian inference become “Bayesian”? Bayesian Anal. 1, 1–40.
1338
Foucquier, A., Robert, S., Suard, F., Stéphan, L., Jay, A., 2013. State of the art in building modelling
1339
and energy performances prediction: A review. Renew. Sustain. Energy Rev. 23, 272–288.
1340
Fumo, N., 2014. A review on the basics of building energy estimation. Renew. Sustain. Energy Rev.
1341
31, 53–60.
1342
Galatioto, A., Beccali, M., 2016. Aspects and issues of daylighting assessment: A review study. Renew.
1343
Sustain. Energy Rev. 66, 852–860.
1344
Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F., 2017.
1345
Learning to predict indoor illumination from a single image. ACM Trans. Graph. 36, 176–189.
1346
Geiger, R., 1954. Klassifikation der Klimate nach W. Köppen, in: Landolt-Börnstein – Zahlenwerte Und
1347
Funktionen Aus Physik, Chemie, Astronomie, Geophysik Und Technik, Alte Serie. Springer,
1348
Berlin, pp. 603–607.
1349
Geisler-Moroder, D., Dür, A., 2008. Validation of Radiance against CIE171: 2006 and improved
1350
adaptive subdivision of circular light sources. Fribourg, Switzerland.
1351
Hansen, L.K., Salamon, P., 1990. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell.
1352
10, 993–1001.
1353
Hecht-Nielsen, R., 1992. Theory of the backpropagation neural network, in: Wechsler, H. (Ed.), Neural
1354
Networks for Perception. Academic Press, pp. 65–93.
1355
Herrera, M., Natarajan, S., Coley, D.A., Kershaw, T., Ramallo-Gonzalez, A.P., Eames, M.E., Fosas, D.,
1356
Wood, M., 2017. A Review of Current and Future Weather Data for Building Simulation. Build.
1357
41
Serv. Eng. Res. Technol. 38, 602–627.
1358
Ho, T.K., 1995. Random decision forests, in: Proceedings of the 3rd International Conference on
1359
Document Analysis and Recognition. Montreal, pp. 278–282.
1360
Hopkinson, R.G., Petherbridge, P., Longmore, J., 1966. Daylighting. Butterworth-Heinemann Ltd,
1361
London, United Kingdom.
1362
Hu, J., Olbina, S., 2011. Illuminance-based slat angle selection model for automated control of split
1363
blinds. Build. Environ. 46, 786–796.
1364
Hubbard, K., Kunkel, K., DeGaetano, A., Redmond, K., 2004. Sources of uncertainty in the calculation
1365
of the design weather conditions in the ASHRAE Handbook of Fundamentals.
1366
Ibarra, D.I., Reinhart, C.F., 2009. Daylight factor simulations-How close do simulation beginners “really”
1367
get? 11th Int. IBPSA Conf. Build. Simul. 2009 196–203.
1368
IEA, 2019a. Global energy & CO2 status report [WWW Document]. URL https://www.iea.org/geco/
1369
(accessed 9.30.19).
1370
IEA, 2019b. Energy efficiency: buildings [WWW Document]. URL
1371
https://www.iea.org/topics/energyefficiency/buildings/ (accessed 10.1.19).
1372
IEA, 2019c. Energy efficiency [WWW Document]. URL https://www.iea.org/reports/energy-efficiency-
1373
2019 (accessed 10.1.19).
1374
IEA, 2018. CO2 emissions from fuel combustion-highlights.
1375
IEA, 2006. Light’s labour’s lost: policies for energy-efficient lighting. Paris.
1376
IEA, 2000. Daylight in buildings. California.
1377
IES, 2012. Approved Method: IES Spatial Daylight Autonomy (sDA) and Annual Sunlight Exposure
1378
(ASE). Illuminating Engineering Society, New York.
1379
IES, 1972. Daytime Lighting in Buildings:
1380
IESNA, 2018. IESNA Lightning Handbook, IESNA Lightning Handbook. https://doi.org/99-76610
1381
Inanici, M., 2013. Dynamic daylighting simulations from static High Dynamic Range imagery using
1382
etrapolation and Daylight Coefficient methodologies, in: Proceedings of the International Building
1383
Performance Simulation Association (IBPSA) 2013 Conference. Chambery, France, pp. 3392–
1384
3399.
1385
Iqbal, I., Al-Homoud, M.S., 2007. Parametric analysis of alternative energy conservation measures in
1386
an office building in hot and humid climate. Build. Environ. 42, 2166–2177.
1387
ISO, 2004. Spatial Distribution of Daylight CIE Standard General Sky.
1388
Ivakhnenko, A.G., 1971. Polynomial theory of complex systems. IEEE Trans. Syst. Man. Cybern. 4,
1389
364–378.
1390
Iversen, A., Svendsen, S., Nielsen, T.R., 2013. The effect of different weather data sets and their
1391
resolution on climate-based daylight modelling. Light. Res. Technol. 45, 305–316.
1392
Jain, A.K., Mao, J., Mohiuddin, K.M., 1996. Artificial neural networks: A tutorial. Computer (Long. Beach.
1393
Calif). 29, 31–44. https://doi.org/10.1109/2.485891
1394
Jain, S., Garg, V., 2018. A review of open loop control strategies for shades, blinds and integrated
1395
lighting by use of real-time daylight prediction methods. Build. Environ. 135, 352–364.
1396
Jakubiec, J.A., Reinhart, C.F., 2011. DIVA 2.0: Integrating daylight and thermal simulations using
1397
42
Rhinoceros 3D, Daysim and EnergyPlus, in: Proceedings of the International Building Simulation
1398
Association Conference. Sydney, pp. 2202–2209.
1399
Jang, J.S., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man.
1400
Cybern. 23, 665–685.
1401
Janjai, S., Plaon, P., 2011. Estimation of Sky Luminance in the Tropics Using Artificial Neural Networks:
1402
Modeling and Performance Comparison with the CIE Model. Appl. Energy 88, 840–847.
1403
Jarvis, D., Donn, M., 1997. Comparison of Computer and Model Simulations of a Daylit Interior with
1404
Reality, in: Proceedings of the 5th Conference of International Building Performance Simulation
1405
Association (IBPSA). Prague, Czech Republic.
1406
Jentsch, M.F., James, P.A., Bourikas, L., Bahaj, A.S., 2013. Transforming existing weather data for
1407
worldwide locations to enable energy and building performance simulation under future climates.
1408
Renew. Energy 55, 514–524.
1409
Jetcheva, J.G., Majidpour, M., Chen, W.P., 2014. Neural network model ensembles for building-level
1410
electricity load forecasts. Energy Build. 84, 214–223.
1411
Joe, Y., Fenxian, H., Seo, D., Krarti, M., 2014. Development of 3012 IWEC2 Weather Files for
1412
International Locations (RP-1477). ASHRAE Trans. 120, 340–355.
1413
Jones, N.L., Reinhart, C.F., 2015. Fast daylight coefficient calculation using graphics hardware, in: 14th
1414
Conference of International Building Performance Simulation Association, Hyderabad, India, Dec.
1415
7-9, 2015. pp. 1237–1244.
1416
Kalamees, T., Jylhä, K., Tietäväinen, H., Jokisalo, J., Ilomets, S., Hyvönen, R., Saku, S., 2012.
1417
Development of weighting factors for climate variables for selecting the energy reference year
1418
according to the EN ISO 15927-4 standard. Energy Build. 47, 53–60.
1419
Kalogirou, S.A., 2001. Artificial neural networks in renewable energy systems applications: a review.
1420
Renew. Sustain. Energy Rev. 5, 373–401.
1421
Kass, G. V, 1980. An exploratory technique for investigating large quantities of categorical data. J. R.
1422
Stat. Soc. Ser. C (Applied Stat. 29, 119–127.
1423
Kaytez, F., Taplamacioglu, M.C., Cam, E., Hardalac, F., 2015. Forecasting electricity consumption: A
1424
comparison of regression analysis, neural networks and least squares support vector machines.
1425
nternational J. Electr. Power Energy Syst. 67, 431–438.
1426
Kazanasmaz, T., Günaydin, M., Binol, S., 2009. Artificial neural networks to predict daylight illuminance
1427
in office buildings. Build. Environ. 44, 1751–1757.
1428
Khean, N., Fabbri, A., Haeusler, M.H., 2018. Learning machine learning as an architect, how to?
1429
Presenting and evaluating a Grasshopper based platform to teach architecture students machine
1430
learning, in: Proceedings of the 36th International Conference on Education and Research in
1431
Computer Aided Architectural Design in Europe. Lodz, Poland, pp. 95–102.
1432
Kira, K., Rendell, L.A., 1992. A practical approach to feature selection, in: Machine Learning
1433
Proceedings 1992. Morgan Kaufmann Publishers, pp. 249–256.
1434
Kleinbaum, D.G., Dietz, K., Gail, M., Klein, M., Klein, M., 2002. Logistic regression. Springer-Verlag,
1435
New York.
1436
Koch-Nielsen, H., 2013. Stay cool: a design guide for the built environment in hot climates. Routledge,
1437
43
London. https://doi.org/https://doi.org/10.4324/9781315074429
1438
Koomey, J.G., Berard, S., Sanchez, M., Wong, H., 2011. Implications of historical trends in the electrical
1439
efficiency of computing. IEEE Ann. Hist. Comput. 33, 46–54.
1440
https://doi.org/10.1109/MAHC.2010.28
1441
Kotsiantis, S.B., 2007. Supervised machine learning: A review of classification techniques. Informatica
1442
31, 249–268.
1443
Kottek, M., Grieser, J., Beck, C., Rudolf, B., Rubel, F., 2006. World map of the Köppen-Geiger climate
1444
classification updated. Meteorol. Zeitschrift 15, 259–263.
1445
Kurian, C.P., Aithal, R.S., Bhat, J., George, V.I., 2008. Robust control and optimisation of energy
1446
consumption in daylight—artificial light integrated schemes. Light. Res. Technol. 40, 7–24.
1447
Kurian, C.P., George, V.I., Bhat, J., Aithal, R.S., 2006. ANFIS model for the time series prediction of
1448
interior daylight illuminance. ICGST Int. J. Artif. Intell. Mach. Learn. 6, 35–40.
1449
Labayrade, R., Fontoynont, M., 2009. Use of CIE 171:2006 test cases to assess the scope of lighting
1450
simulation programs, in: Proceedings of the CIE Light and Lighting Conference. Budapest,
1451
Hungary.
1452
LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998. Gradient-based learning applied to document
1453
recognition. Proc. IEEE 86, 2278–2324.
1454
Lee, E.S., Geisler-Moroder, D., Ward, G.J., 2018. Modeling the direct sun component in buildings using
1455
matrix algebraic approaches: Methods and validation. Sol. Energy 160, 380–395.
1456
Li, D.H., Lou, S., 2018. Review of solar irradiance and daylight illuminance modeling and sky
1457
classification. Renew. Energy 126, 445–453.
1458
Li, D.H., Tang, H.L., Lee, E.S., Muneer, T., 2010. Classification of CIE standard skies using probabilistic
1459
neural networks. Int. J. Climatol. A J. R. Meteorol. Soc. 30, 305–315.
1460
Lim, Y.W., Ahmad, M.H., Ossen, D.R., 2010. Empirical validation of daylight simulation tool with
1461
physical model measurement. Am. J. Appl. Sci. 7, 1426–1431.
1462
Liu, Y., Colburn, A., Inanici, M., 2018. Computing Long-Term Daylighting Simulations from High
1463
Dynamic Range Imagery Using Deep Neural Networks, in: The Building Performance Analysis
1464
Conference and SimBuild Co-Organized by ASHRAE and IBPSA-USA, Chicago, United States.
1465
Liu, Y., Huang, Y., Stouffs, R., 2015. Using a data-driven approach to support the design of energy-
1466
efficient buildings. J. Inf. Technol. Constr. 20, 80–96.
1467
Liu, Z., Wu, D., Liu, Y., Han, Z., Lun, L., Gao, J., Jin, G., Cao, G., 2019. Accuracy analyses and model
1468
comparison of machine learning adopted in building energy consumption prediction. Energy
1469
Explor. Exploit. 37, 1426–1451.
1470
Logar, V., Kristl, Ž., Škrjanc, I., 2014. Using a fuzzy black-box model to estimate the indoor illuminance
1471
in buildings. Energy Build. 70, 343–351.
1472
Lokuciejewski, P., Stolpe, M., Morik, K., Marwedel, P., 2010. Automatic Selection of Machine Learning
1473
Models for Compiler Heuristic Generation, in: Proceedings of the 4th Workshop on Statistical and
1474
Machine Learning Approaches to Architecture and Compilation SMART. Pisa, Italy, pp. 3–17.
1475
López, G., Gueymard, C.A., 2007. Clear-sky solar luminous efficacy determination using artificial neural
1476
networks. Sol. Energy 81, 929–939.
1477
44
Lorenz, C.L., Jabi, W., 2017. Predicting daylight autonomy metrics using machine learning, in:
1478
Proceedings of the International Conference for Sustainable Design of the Built Environment
1479
(SDBE 2017). London, United Kingdom, pp. 991–1002.
1480
Lorenz, C.L., Packianather, M., Spaeth, A.B., De Souza, C.B., 2018. Artificial Neural Network-Based
1481
Modelling for Daylight Evaluations, in: The Symposium on Simulation for Architecture and Urban
1482
Design (SimAUD 2018), Delft, Netherlands.
1483
Lou, S., Li, D.H., Lam, J.C., 2017. CIE standard sky classification by accessible climatic indices. Renew.
1484
Energy 113, 347–356.
1485
Love, J.A., 1992. The evolution of performance indicators for the evaluation of daylighting systems,
1486
Conference Record - IAS Annual Meeting (IEEE Industry Applications Society). IEEE, Texas,
1487
United States. https://doi.org/10.1109/IAS.1992.244539
1488
Maamari, F., Fontoynont, M., 2003. Analytical tests for investigating the accuracy of lighting programs.
1489
Light. Res. Technol. 35, 225–239.
1490
Mardaljevic, J., 2015. Climate-based daylight modelling and its discontents, in: Presented at the Simple
1491
Buildings Better Buildings? Delivering Performance through Engineered Solutions, CIBSE
1492
Technical Symposium. London, United Kingdom.
1493
Mardaljevic, J., 2001. The BRE-IDMP Dataset: A New Benchmark for the Validation of Illuminance
1494
Prediction Techniques. Trans. Illum. Eng. Soc. 33, 117–134.
1495
Mardaljevic, J., 2000a. Daylight Simulation: Validation, Sky Models and Daylight Coefficients. Ph.D. De
1496
Montfort University, Leicester, United Kingdom.
1497
Mardaljevic, J., 2000b. Simulation of annual daylighting profiles for internal illuminance. Light. Res.
1498
Technol. 32, 111–118. https://doi.org/10.1177/096032710003200302
1499
Mardaljevic, J., 1997. Validation of a lighting simulation program: A study using measured sky
1500
brightness distributions, in: Proceedings of the 8th European Lighting Conference. Amsterdam,
1501
pp. 555–569.
1502
Mardaljevic, J., 1995. Validation of a lighting simulation program under real sky conditions. Light. Res.
1503
Technol. 27, 181–188. https://doi.org/10.1177/14771535950270040701
1504
Mardaljevic, J., Andersen, M., Roy, N., Christoffersen, J., 2014. A framework for predicting the non-
1505
visual effects of daylight–Part II: The simulation model. Light. Res. Technol. 46, 388–406.
1506
Mardaljevic, J., Heschong, L., Lee, E.S., 2009. Daylight Metrics and Energy Savings. Light. Res.
1507
Technol. 41, 261–283.
1508
McNeil, A., 2014. The Three-Phase Method for Simulating Complex Fenestration with Radiance [WWW
1509
Document]. Lawrence Berkley Natl. Lab. (LNBL), Berkeley, Calif.
1510
McNeil, A., 2013. The Five-Phase Method for Simulating Complex Fenestration with Radiance [WWW
1511
Document]. Lawrence Berkley Natl. Lab. (LNBL), Berkeley, Calif.
1512
McNeil, A., Lee, E.S., 2012. A validation of the Radiance three-phase simulation method for modelling
1513
annual daylight performance of optically complex fenestration systems. J. Build. Perform. Simul.
1514
6, 24–37. https://doi.org/10.1080/19401493.2012.671852
1515
Millet, M.S., Adams, C., Bedrick, J.R., 1980. Graphic daylighting design method: Including clear sky
1516
conditions, in: Massachusetts, U.S. (Ed.), Proceedings of the 5th National Passive Solar
1517
45
Conference. pp. 1184–1188.
1518
Mitchell, T.M., 1997. Machine Learning. McGraw-Hill Higher Education, New York.
1519
Moazami, A., Nik, V.M., Carlucci, S., Geving, S., 2019. Impacts of future weather data typology on
1520
building energy performance–Investigating long-term patterns of climate change and extreme
1521
weather conditions. Appl. Energy 238, 696–720.
1522
Mylona, A., 2012. The use of UKCP09 to produce weather files for building simulation. Build. Serv. Eng.
1523
Res. Technol. 33, 51–62.
1524
Nabil, A., Mardaljevic, J., 2006. Useful daylight illuminances: A replacement for daylight factors. Energy
1525
Build. 38, 905–913. https://doi.org/10.1016/j.enbuild.2006.03.013
1526
Nabil, A., Mardaljevic, J., 2005. Useful Daylight Illuminance: A New Paradigm for Assessing Daylight in
1527
Buildings. Light. Res. Technol. 37, 41–57.
1528
NASA, 2019. 2018 Fourth Warmest Year in Continued Warming Trend, According to NASA, NOAA
1529
[WWW Document]. URL https://www.nasa.gov/press-release/2018-fourth-warmest-year-in-
1530
continued-warming-trend-according-to-nasa-noaa (accessed 9.30.19).
1531
Nasrollahi, N., Shokri, E., 2016. Daylight illuminance in urban environments for visual comfort and
1532
energy performance. Renew. Sustain. Energy Rev. 66, 861–874.
1533
Nault, E., Moonen, P., Rey, E., Andersen, M., 2017. Predictive models for assessing the passive solar
1534
and daylight potential of neighborhood designs: A comparative proof-of-concept study. Build.
1535
Environ. 116, 1–16.
1536
Navada, S.G., Adiga, C.S., Kini, S.G., 2016. Prediction of daylight availability for visual comfort. Int. J.
1537
Appl. Eng. Res. 11, 4711–4717.
1538
Ng, E., 2001. A Study of the Accuracy of Daylighting Simulation of Heavily obstructed Buildings in Hong
1539
Kong, in: Proceedings of the 7th Conference of the International Building Performance Simulation
1540
Association (IBPSA). Rio de Janeiro, Brazil, pp. 1215–1222.
1541
NREL, 1995. User’s manual for TMY2s (typical meteorological years), and TMY2s, typical
1542
meteorological years derived from the 1961-1990 national solar radiation data base.
1543
NSRD, 2019. 1991- 2005 update: typical meteorological year 3 [WWW Document]. URL
1544
https://rredc.nrel.gov/solar/old_data/nsrdb/1991-2005/tmy3/ (accessed 5.20.19).
1545
Ochoa, C.E., Aries, M.B.C., Hensen, J.L.M., 2012. State of the art in lighting simulation for building
1546
science: A literature review. J. Build. Perform. Simul. 5, 209–233.
1547
https://doi.org/10.1080/19401493.2011.558211
1548
Olbina, S., Beliveau, Y., 2009. Developing a Transparent Shading Device as a Daylighting System.
1549
Build. Res. Inf. 37, 148–163.
1550
Osborne, J., 2012. Building a comprehensive dataset for the validation of daylight simulation software,
1551
using complex real architecture. Victoria University of Wellington, Wellington, New Zealand.
1552
Pan, S.J., Yang, Q., 2009. A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 22, 1345–1359.
1553
Parkhurst, J., Darringer, J., Grundmann, B., 2006. From single core to multi-core: Preparing for a new
1554
exponential, in: IEEE/ACM International Conference on Computer-Aided Design, Digest of
1555
Technical Papers, ICCAD. pp. 67–72. https://doi.org/10.1109/ICCAD.2006.320067
1556
Pattanasethanon, S., Lertsatitthanakorn, C., Atthajariyakul, S., Soponronnarit, S., 2008. An accuracy
1557
46
assessment of an empirical sine model, a novel sine model and an artificial neural network model
1558
for forecasting illuminance/irradiance on horizontal plane of all sky types at Mahasarakham,
1559
Thailand. Energy Convers. Manag. 49, 1999–2005.
1560
Pauley, S.M., 2004. Lighting for the human circadian clock: recent research indicates that lighting has
1561
become a public health issue. Med. Hypotheses 63, 588–596.
1562
Pérez-Lombard, L., Ortiz, J., Pout, C., 2008. A review on buildings energy consumption information.
1563
Energy Build. 40, 394–398.
1564
Perez, R., Seals, R., Michalsky, J., 1993. All-weather model for sky luminance distribution-Preliminary
1565
configuration and validation. Sol. Energy 50, 235–245. https://doi.org/10.1016/0038-
1566
092X(93)90017-I
1567
Pernigotto, G., Prada, A., Gasparella, A., Hensen, J.L.M., 2014. Analysis and improvement of the
1568
representativeness of EN ISO 15927-4 reference years for building energy simulation. J. Build.
1569
Perform. Simul. 7, 391–410.
1570
Polson, N.G., Scott, S.L., 2011. Data augmentation for support vector machines. Bayesian Anal. 6, 1–
1571
24.
1572
Quinlan, J.R., 1993. C4.5: Programs for machine learning. Morgan Kaufmann Publishers, San Mateo,
1573
California.
1574
Quinlan, J.R., 1986. Induction of decision trees. Mach. Learn. 1, 81–106.
1575
Radziszewski, K., Waczyńska, M., 2018. Machine Learning Algorithm-Based Tool and Digital
1576
Framework for Substituting Daylight Simulations in Early-Stage Architectural Design Evaluation,
1577
in: The Symposium on Simulation for Architecture and Urban Design (SimAUD 2018), Delft,
1578
Netherlands.
1579
Reinhart, C.F., 2018. Daylighting Handbook II. Building Technology Press.
1580
Reinhart, C.F., 2014. Daylighting handbook I: Fundamentals, designing with the sun, Muscle & Nerve.
1581
Building Technology Press. https://doi.org/10.1002/mus.24024
1582
Reinhart, C.F., 2011. Daylight Performance Predictions, in: Hensen, J.L., Lamberts, R. (Eds.), Building
1583
Performance Simulation for Design and Operation. Spon Press New York, United States, pp. 235–
1584
276.
1585
Reinhart, C.F., Andersen, M., 2006. Development and validation of a Radiance model for a translucent
1586
panel. Energy Build. 38, 890–904. https://doi.org/10.1016/j.enbuild.2006.03.006
1587
Reinhart, C.F., Breton, P.F., 2009. Experimental Validation of Autodesk 3ds Max Design 2009 and
1588
DAYSIM 3.0. Leukos 6, 7–35.
1589
Reinhart, C.F., Fitz, A., 2006. Findings from a Survey on the Current Use of Daylight Simulations in
1590
Building Design. Energy Build. 38, 824–835.
1591
Reinhart, C.F., Herkel, S., 2000. The Simulation of annual daylight illuminance distributions-a state-of-
1592
the-art comparison of six RADIANCE-based methods. Energy Build. 32, 167–187.
1593
https://doi.org/10.1016/S0378-7788(00)00042-6
1594
Reinhart, C.F., Jakubiec, J.A., Ibarra, D., 2013. Definition of a reference office for standardized
1595
evaluations of dynamic façade and lighting technologies. Proc. BS 2013 13th Conf. Int. Build.
1596
Perform. Simul. Assoc. 3645–3652. https://doi.org/10.1016/B978-008044722-3/50009-9
1597
47
Reinhart, C.F., LoVerso, V.R.M., 2010. A rules of thumb-based design sequence for diffuse daylight.
1598
Light. Res. Technol. 42, 7–31.
1599
Reinhart, C.F., Mardaljevic, J., Rogers, Z., 2006. Dynamic daylight performance metrics for sustainable
1600
building design. Leukos 3, 7–31.
1601
Reinhart, C.F., Walkenhorst, O., 2001. Validation of dynamic RADIANCE-based daylight simulations
1602
for a test office with external blinds. Energy Build. 33, 683–697. https://doi.org/10.1016/S0378-
1603
7788(01)00058-5
1604
Reinhart, C.F., Wienold, J., 2010. The daylighting dashboard - A simulation-based design analysis for
1605
daylit spaces, in: Fourth National Conference of IBPSA-USA. New York, pp. 410–418.
1606
Rogers, Z., Goldman, D., 2006. Daylighting metric development using daylight autonomy calculations
1607
in the sensor placement optimization tool. Boulder, Colorado, USA.
1608
Roudsari, M., Pak, M., 2014. Ladybug: A Parametric Environmental Plugin for Grasshopper to Help
1609
Designers Create an Environmentally-Conscious Design, in: Proceedings of the International
1610
Building Simulation Association Conference (IBPSA 2013). Chambéry, pp. 3128–3135.
1611
Rumelhart, D.E., Hintonf, G.E., 1986. Learning representations by back-propagating errors. Nature 323,
1612
533–536.
1613
Şahin, M., Oğuz, Y., Büyüktümtürk, F., 2015. Approximate and three-dimensional modeling of
1614
brightness levels in interior spaces by using artificial neural networks. J. Electr. Eng. Technol. 10,
1615
1822–1829.
1616
Saltelli, A., Tarantola, S., Campolongo, F., Ratto, M., 2004. Sensitivity analysis in practice: a guide to
1617
assessing scientific models. John Wiley & Sons Ltd, Chichester.
1618
Santamouris, M., Cartalis, C., Synnefa, A., Kolokotsa, D., 2015. On the impact of urban heat island and
1619
global warming on the power demand and electricity consumption of buildings — A review. Energy
1620
Build. 98, 119–124.
1621
Saxena, M., Ward, G.J., Perry, T., Heschong, L., Higa, R., 2010. Dynamic Radiance–Predicting annual
1622
daylighting with variable fenestration optics using BSDFs, in: Proceedings of the Fourth National
1623
Conference of IBPSA-USA. New York, pp. 402–409.
1624
Schardl, T.B., 2016. Performance engineering of multicore software : developing a science of fast code
1625
for the post-Moore era (PhD Thesis). Massachusetts Institute of Technology.
1626
Schmidhuber, J., 2015. Deep learning in neural networks: An overview. Neural Networks 61, 85–117.
1627
Schregle, R., Wienold, J., 2004. Physical validation of global illumination methods: measurement and
1628
error analysis. Comput. Graph. (ACM). 23, 761–781.
1629
Smith, K., 1918. On the standard deviations of adjusted and interpolated values of an observed
1630
polynomial function and its constants and the guidance they give towards a proper choice of the
1631
distribution of observations. Biometrika 12, 1–85.
1632
Suykens, J.A., Vandewalle, J., 1999. Least squares support vector machine classifiers. Neural Process.
1633
Lett. 9, 293–300.
1634
Suykens, J.A., Vandewalle, J.P., de Moor, B.L., 2012. Artificial neural networks for modelling and control
1635
of non-linear systems. Springer Science & Business Media.
1636
Swan, L.G., Ugursal, V.I., 2009. Modeling of end-use energy consumption in the residential sector: A
1637
48
review of modeling techniques. Renew. Sustain. Energy Rev. 13, 1819–1835.
1638
Thanachareonkit, A., Scartezzini, J.L., 2010. Modelling complex fenestration systems using physical
1639
and virtual models. Sol. Energy 84, 563–586.
1640
Tregenza, P.R., 2017. Uncertainty in daylight calculations. Light. Res. Technol. 49, 829–844.
1641
https://doi.org/10.1177/1477153516653786
1642
Trotter, A., 1911. Illumination; its Distribution and Measurement. Nature 88, 72–73.
1643
https://doi.org/10.1038/088072a0
1644
Twomey, J.M., Smith, A.E., 1995. Performance measures, consistency, and power for artificial neural
1645
network models. Math. Comput. Model. 21, 243–258.
1646
Tzempelikos, A., Athienitis, A.K., 2007. The impact of shading design and control on building cooling
1647
and lighting demand. Sol. Energy 81, 369–382.
1648
Ürge-Vorsatz, D., Eyre, N., Graham, P., Harvey, D., Hertwich, E., Jiang, Y., Kornevall, C., Majumdar,
1649
M., McMahon, J.E., Mirasgedis, S., Murakami, S., 2012. Energy end-use: buildings, in:
1650
Johansson, T.B., Patwardhan, A.P., Nakićenović, N., Gomez-Echeverri, L. (Eds.), Global Energy
1651
Assessment: Toward a Sustainable Future. Cambridge University Press, Cambridge, UK and New
1652
York, USA, pp. 649–760.
1653
Uribe, D., Veraand, S., Bustamante, W., 2017. Optimization of Complex Fenestration Systems using
1654
an Artificial Neural Network, in: The 51st International Conference of the Architectural Science
1655
Association (ANZAScA), Wellington, New Zealand.
1656
USDOE, 2019. 2017 renewable energy data book. (NREL) National Renewable Energy Laboratory.
1657
USGBC, 2013. LEED Reference Guide for Building Design and Construction, Version 4 [WWW
1658
Document].
1659
Vapnik, V.N., 1999. An overview of statistical learning theory. IEEE Trans. Neural Networks 10, 988–
1660
999.
1661
Vergara, J.R., Estévez, P.A., 2014. A review of feature selection methods based on mutual information.
1662
Neural Comput. Appl. 24, 175–186.
1663
Verso, V.R.M. Lo, Mihaylov, G., Pellegrino, A., Pellerey, F., 2017. Estimation of the daylight amount
1664
and the energy demand for lighting for the early design stages: Definition of a set of mathematical
1665
models. Energy Build. 155, 151–165.
1666
Voyant, C., Notton, G., Kalogirou, S., Nivet, M.L., Paoli, C., Motte, F., Fouilloy, A., 2017. Machine
1667
learning methods for solar radiation forecasting: A review. Renew. Energy 105, 569–582.
1668
Waldram, P.J., Waldram, J.M., 1923. Window design and the measurement and predetermination of
1669
daylight illumination. Illum. Eng. XVI, 90–122.
1670
Walkenhorst, O., Luther, J., Reinhart, C.F., Timmer, J., 2002. Dynamic annual daylight simulations
1671
based on one-hour and one-minute means of irradiance data. Sol. Energy 72, 385–395.
1672
https://doi.org/10.1016/S0038-092X(02)00019-1
1673
Wang, J., Wei, M., Chen, L., 2019. Does typical weather data allow accurate predictions of daylight
1674
quality and daylight-responsive control system performance. Energy Build. 184, 72–87.
1675
Wang, Z., Srinivasan, R.S., 2017. A review of artificial intelligence based building energy use prediction:
1676
Contrasting the capabilities of single and ensemble prediction models. Renew. Sustain. Energy
1677
49
Rev. 75, 796–808.
1678
Ward, G.J., 1994. The RADIANCE lighting simulation and rendering system, Proceedings of the 21st
1679
annual conference on Computer graphics and interactive techniques - SIGGRAPH ’94. ACM
1680
Press, Florida, United States. https://doi.org/10.1145/192161.192286
1681
Ward, G.J., 1992. Measuring and modeling anisotropic reflection, in: Proceedings of the 19th Annual
1682
Conference on Computer Graphics and Interactive Techniques. pp. 265–272.
1683
Ward, G.J., Shakespeare, R., 1998. Rendering with Radiance: The Art and Science of Lighting
1684
Visualization. Morgan Kaufman. https://doi.org/citeulike-article-id:9944262
1685
Webb, A.R., 2006. Considerations for lighting in the built environment: Non-visual effects of light. Energy
1686
Build. 38, 721–727.
1687
Weber, H., Prévost, D., Lalonde, J.F., 2018. Learning to estimate indoor lighting from 3d objects, in:
1688
Proceedings of the International Conference on 3D Vision (3DV). IEEE, pp. 199–207.
1689
Wei, Y., Zhang, X., Shi, Y., Xia, L., Pan, S., Wu, J., Han, M., Zhao, X., 2018. A review of data-driven
1690
approaches for prediction and classification of building energy consumption. Renew. Sustain.
1691
Energy Rev. 82, 1027–1047.
1692
Wienold, J., Christoffersen, J., 2006. Evaluation methods and development of a new glare prediction
1693
model for daylight environments with the use of CCD cameras. Energy Build. 38, 743–757.
1694
https://doi.org/10.1016/j.enbuild.2006.03.017
1695
Wong, S.L., Wan, K.K., Lam, T.N., 2010. Artificial neural networks for energy analysis of office buildings
1696
with daylighting. Appl. Energy 87, 551–557.
1697
Wu, X., Kumar, V., Quinlan, J.R., Ghosh, J., Yang, Q., Motoda, H., McLachlan, G.J., Ng, A., Liu, B.,
1698
Philip, S.Y., Zhou, Z.H., 2008. Top 10 algorithms in data mining. Knowl. Inf. Syst. 14, 1–37.
1699
Wyszecki, G., 1970. Development of new CIE standard sources for colorimetry. Die Farbe 19, 43–76.
1700
Xue, P., Mak, C.M., Huang, Y., 2016. Quantification of luminous comfort with dynamic daylight metrics
1701
in residential buildings. Energy Build. 117, 99–108.
1702
Yacine, S.M., Noureddine, Z., Piga, B.E.A., Morello, E., others, 2017. Developing neural networks to
1703
investigate relationships between lighting quality and lighting glare indices. Energy Procedia 122,
1704
799–804.
1705
Yadav, A.K., Chandel, S.S., 2014. Solar radiation prediction using artificial neural network techniques:
1706
A review. Renew. Sustain. Energy Rev. 33, 772–781.
1707
Yang, L., Lam, J.C., Liu, J., Tsang, C.L., 2008. Building energy simulation using multi-years and typical
1708
meteorological years in different climates. Energy Convers. Manag. 49, 113–124.
1709
Yao, X., 1999. Evolving artificial neural networks. Proc. IEEE 87, 1423–1447.
1710
https://doi.org/10.1109/5.784219
1711
Yoon, Y.B., Moon, J.W., Kim, S., 2016. Development of annual daylight simulation algorithms for
1712
prediction of indoor daylight illuminance. Energy Build. 118, 1–17.
1713
Zendehboudi, A., Baseer, M.A., Saidur, R., 2018. Application of support vector machine models for
1714
forecasting solar and wind energy resources: A review. J. Clean. Prod. 199, 272–285.
1715
Zhang, J., Lalonde, J.F., 2017. Learning high dynamic range from outdoor panoramas, in: Proceedings
1716
of the IEEE International Conference on Computer Vision and Pattern Recognition (CVPR 2017).
1717
50
pp. 4529–4538.
1718
Zhang, W., Zhang, H., Liu, J., Li, K., Yang, D., Tian, H., 2017. Weather prediction with multiclass support
1719
vector machines in the fault detection of photovoltaic system. IEEE/CAA J. Autom. Sin. 4, 520–
1720
525.
1721
Zhao, H.X., Magoulès, F., 2012. A review on the prediction of building energy consumption. Renew.
1722
Sustain. Energy Rev. 16, 3586–3592.
1723
Zhou, S., Liu, D., 2015. Prediction of daylighting and energy performance using artificial neural network
1724
and support vector machine. Am. J. Civ. Eng. Archit. 3, 1–8.
1725
Zou, L., Wang, L., Lin, A., Zhu, H., Peng, Y., Zhao, Z., 2016. Estimation of global solar radiation using
1726
an artificial neural network based on an interpolation technique in southeast China. J. Atmos.
1727
Solar-Terrestrial Phys. 146, 110–122.
1728
1729
