Content uploaded by Zhipeng Deng
Author content
All content in this area was uploaded by Zhipeng Deng on Jan 25, 2023
Content may be subject to copyright.
Development and validation of a smart HVAC control system for multi-occupant
1
offices by using occupants’ physiological signals from wristband
2
3
Zhipeng Deng1, Qingyan Chen1,*
4
1School of Mechanical Engineering, Purdue University, 585 Purdue Mall, West Lafayette,
5
IN 47907, USA
6
7
*Corresponding author: Qingyan Chen, yanchen@purdue.edu
8
9
Abstract
10
Since people spend most of their time indoors, it is important to create comfortable indoor
11
environments for building occupants. However, unsuitable thermostat settings lead to
12
energy waste and an undesirable indoor environment, especially in multi-occupant rooms.
13
This study aimed to develop and validate a control strategy for the HVAC systems in multi-
14
occupant offices using physiological parameters measured by wristbands. We used an
15
ANN model to predict thermal sensation from indoor environmental and physiological
16
parameters such as air temperature, relative humidity, clothing level, wrist skin temperature,
17
skin relative humidity and heart rate. The model was trained by data collected in seven
18
multi-occupant offices in the course of a year, and it was able to predict the thermal
19
sensation with high accuracy. Next, we developed a control strategy for the HVAC system
20
to improve the thermal comfort of all the occupants in the room. The control system was
21
smart and could adjust the thermostat set point automatically in real time. We validated the
22
system by means of both experiments and numerical simulations. In most cases, we
23
improved the occupants’ thermal comfort level. After using the wristband control, over
24
half of the occupants experienced a neutral sensation, and fewer than 5% still felt
25
uncomfortable. The energy consumption by the HVAC system with the wristband control
26
was almost the same as when the constant set point was used. After coupling with
27
occupancy-based control by means of lighting sensors or wristband Bluetooth, the heating
28
and cooling loads were reduced by 90% and 30%, respectively, in interior offices.
29
Therefore, the smart HVAC control system can effectively control the indoor environment
30
for thermal comfort and energy saving.
31
32
Keywords
33
Thermal comfort, artificial neural network, air temperature, skin temperature, skin relative
34
humidity, heart rate, thermostat set point
35
36
37
1. Introduction
38
Currently, people in North America spend roughly 90% of their time indoors [1]. Therefore,
39
it is important to create comfortable, healthy, and productive indoor environments for
40
occupants. Such environments are typically achieved by the use of heating, ventilating, and
41
air-conditioning (HVAC) systems. Unfortunately, our resulting indoor environments are
42
still very poor, as demonstrated by a survey in which the predominant complaint by office
43
occupants was that “it is too hot and too cold simultaneously” [2].
44
In many buildings, although occupants can actively adjust indoor environment settings,
45
studies [3-4] have shown that the occupants know little about the thermostat or the HVAC
46
control system, and thermostats often have unsuitable settings. The resulting overheating
47
and overcooling issues in buildings reportedly waste ten billion dollars per year in the US
48
[5]. In addition, a previous study [6] found that in multi-occupant offices, unawareness of
49
others’ feelings and the need to compromise with other people worsened the indoor
50
environment and sometimes made it more extreme. To solve these issues, we need to
51
automate HVAC control systems and create “smart” systems that can ascertain occupants’
52
thermal sensation [7].
53
Thermal sensation encompasses the physiological and subjective response of occupants to
54
the thermal environment in buildings, in vehicles and outdoors [8-9]. To evaluate thermal
55
sensation, Fanger developed predictive mean vote (PMV) and predicted percentage
56
dissatisfied (PPD) models in the 1970s [10]. However, the PMV model was developed by
57
conducting a questionnaire with a large group of occupants. The model does not consider
58
individual differences and parameters and thus cannot be used for personalized control.
59
Subsequently, many researchers have developed personalized thermal comfort models [11]
60
that achieve better accuracy with more individualized parameters, and may also be used to
61
control the indoor environment [12-14]. These personalized models [15-18] have
62
correlated individual thermal comfort with various parameters of human physiology. The
63
most frequently used physiological parameters for evaluating thermal sensation were local
64
skin temperature [19-21], facial temperature [22-23], heart rate (HR) [24-25], blood
65
pressure [26-27], pulse waves [28], brain waves [16, 29], and sweat rate [30-31]. These
66
personalized thermal comfort models were able to predict occupants’ thermal sensation
67
with high accuracy. Several studies [30, 32, 33] also found that, when the occupants felt
68
uncomfortable in a transient thermal environment, their skin temperature, HR and sweat
69
rate exhibited a noticeably different pattern from comfortable condition. Hence, a
70
correlation exists between the physiological parameters and occupants’ thermal sensation
71
and behavior. It is possible to use this correlation to control HVAC systems for thermal
72
comfort.
73
To measure and monitor human physiological parameters, some studies [18, 30, 34, 35]
74
used specialized sensors and medical equipment, which were not convenient for occupants’
75
everyday work or for longtime monitoring. In recent years, the development of personal
76
health monitoring devices, such as wristbands and smart watches, have provided the means
77
for nonintrusive monitoring of physiological parameters in real time [34, 36, 37]. However,
78
only a few studies have used human physiological data for HVAC system control. For
79
example, Li et al. [37] used the collected data from wristbands and a smart thermostat to
80
build a random forest model for predicting thermal preference, and then used the model to
81
test a smart phone application framework for determining the optimal room conditioning
82
mode and HVAC setting. Yi [22] and Cosma [23, 38] used facial skin temperature from a
83
thermographic camera for a building control system that provided individualized thermal
84
comfort. However, the occupants’ clothing level and metabolic rate could not be recorded
85
with the use of the thermographic camera. Li et al. [36, 39] used skin temperature and HR
86
to develop an environment optimization algorithm for thermal comfort and energy saving,
87
but the control model was linear, and it could only be used for a single occupant. Currently,
88
there is no smart control algorithm using human physiological data that can be applied to
89
multi-occupant offices.
90
Hence, the purpose of this study was to develop and validate a control strategy for HVAC
91
systems in multi-occupant offices using wristbands to provide thermal comfort. For this
92
purpose, we collected the indoor environmental parameters and occupants’ thermal
93
sensation and human physiological data in several multi-occupant offices. Next, we trained
94
an artificial neural network (ANN) model to predict the thermal sensation. Based on this
95
model, we developed an HVAC control algorithm that could better ascertain the thermal
96
sensation of multiple occupants. Occupants can then effectively operate and control the
97
indoor environment for thermal comfort.
98
The layout of this paper is organized as follows: Section 2 describes the methods for
99
collecting data, predicting thermal sensation, and developing and validating HVAC control
100
strategies. Section 3 provides the results of data collection and comfort control system
101
analysis. Sections 4 and 5 discuss the results and summarize conclusions of this study,
102
respectively.
103
104
2. Methods
105
To develop an HVAC control system for overall thermal comfort that uses occupants’
106
physiological parameters, we first collected data on the indoor environment, thermal
107
sensation, and physiological parameters in seven multi-occupant offices. Subsequently, we
108
built and trained an ANN model using the collected data. Finally, we developed and
109
validated the control strategies for the HVAC systems according to the correlation between
110
the physiological parameters and occupants’ thermal sensation.
111
112
2.1 Data collection
113
This study collected data on air temperature, relative humidity (RH), clothing level, thermal
114
sensation, wrist skin temperature, wrist skin RH, and HR in seven multi-occupant offices
115
at Purdue University, US. The offices were located on the first and second floors of a three-
116
story building as shown in Fig. 1. We chose offices in which the occupants spent a
117
considerable amount of time. A total of 24 students (16 males and eight females) of
118
different ages participated in the data collection.
119
120
(a)
(b)
Figure 1. Layout of (a) the first floor and (b) the second floor of the building used for the
121
data collection. The red dots indicate the multi-occupant offices used.
122
123
Each of the offices had a thermostat (Siemens 544–760A) as shown in Fig. 2(a) which
124
enabled the building automation system (BAS) to control the room air temperature. The
125
occupants could adjust the set point of the thermostat within the range of 18.3°C (65°F) to
126
26.7°C (80°F). We used data loggers (Sper Scientific 800,049) as shown in Fig. 2(b) in
127
each office to record the room air temperature and RH every ten minutes. In the early
128
mornings before the occupants’ arrival, we adjusted the thermostat set point in each office
129
to a different temperature to expand the data range. We used wristbands (Hesvit S3) as
130
shown in Fig. 2(c) to record the occupants’ physiological data in the Ray W. Herrick
131
Laboratories (HLAB) offices, including wrist skin temperature, wrist skin RH and HR,
132
every ten minutes. Each wristband had a unique serial number and could communicate with
133
a cellphone via Bluetooth. The working distance of the Bluetooth connection was 5 m, and
134
thus we could use it to detect the presence of each occupant in the offices.
135
We also used a questionnaire to collect the thermal sensation vote (TSV) [40] according to
136
a seven-point scale (−3 for cold, −2 for cool, −1 for slightly cool, 0 for neutral, +1 for
137
slightly warm, +2 for warm, and +3 for hot) and clothing level from the occupants every
138
ten minutes when they were inside the offices.
139
140
(a)
(b)
(c)
Figure 2. Data collection devices used in this study. (a) thermostat on wall, (b) data
141
logger, (c) wristband.
142
143
With the above effort, we were able to collect the necessary data. Note that all data
144
collection in this study was approved by the Purdue University Institutional Review Board
145
Protocol # 1902021796.
146
147
2.2 Artificial neural network model for thermal comfort
148
149
With the collected data, we built a model to correlate the indoor environmental and
150
physiological data with occupants’ TSV. This study began with the following hypothesis:
151
• The impact of the outdoor weather and solar radiation on the indoor environment
152
and TSV was neglected, since all the data were collected in interior offices as shown
153
in Fig. 1.
154
• Based on the collected data and published literature [22, 41], the wrist skin
155
temperature difference in each time step was linearly related to the air temperature
156
difference. Therefore, we used linear regression to determine the correlation
157
coefficient with the collected data, as the following equation shows:
158
, , , ,
()
skin f skin i air f air i
T T C T T− = −
(1)
159
where
skin
T
is the skin temperature,
air
T
the air temperature,
C
the coefficient, and
160
the subscripts
i
and
f
represent the initial values before the control system
161
adjusted the thermostat set point and the final values after control, respectively.
162
• Room air RH changed when the thermostat set point was adjusted. We assumed
163
that the air pressure and humidity ratio remained constant, and the skin RH
164
variation was the same as the air RH variation. Thus, we have
165
, , , ,
( , ) ( , )
air i air i air f air f
AH T RH AH T RH=
(2)
166
, , , ,skin f skin i air f air i
RH RH RH RH− = −
(3)
167
where
( , )
air air
AH T RH
is the absolute humidity at a specific air temperature and
168
air RH, and
skin
RH
is the wrist skin relative humidity.
169
• Metabolic rate was related to HR, according to the literature [42] and ISO8996 [43].
170
Occupants’ clothing level and HR remained the same in the offices before and after
171
the control system adjusted the thermostat set point. Thus,
172
if
HR HR=
(4)
173
if
Clo Clo=
(5)
174
175
With the above hypothesis, this study could predict the occupants’ TSV. In many previous
176
studies, [6, 44-46] ANN models have been very effective in dealing with the highly
177
complex correlations between input parameters and TSV. Therefore, the present study also
178
employed this type of model. An ANN model uses machine learning methods to learn a
179
particular relationship between input and output, and it can identify the relationship after
180
being trained with sufficient data. This study sought to correlate occupants’ thermal
181
sensation with indoor environmental parameters and physiological parameters.
182
183
As shown in Fig. 3, an ANN model has a layered structure, typically comprised of an input
184
layer, a hidden layer and an output layer. The number of neurons in the hidden layer
185
indicates the model’s complexity, and adjusting this number allows one to control the
186
complexity. However, increasing the number of neurons could result in overfitting and a
187
longer training time. In this study, we found that six neurons in the hidden layer could
188
predict the TSV accurately without overfitting. The transfer function in the hidden layer is
189
a given function that can provide the corresponding output value for each possible input.
190
In this study, we used the logistic function as the transfer function because it can provide
191
the TSV for any possible input.
192
193
194
Figure 3. Structure of the ANN model in this study. There are six input parameters, six
195
neutrons in the hidden layer and one output parameter.
196
197
Hence, the mathematical form of the ANN model in this study can be expressed as
198
199
1
{1 exp[ ( )]}
output hidden hidden output
TSV b
−
= + − + +w w X b
(6)
200
201
where
X
is an
1n
input vector for the n input parameters,
hidden
w
is a
6n
weight matrix
202
in the hidden layer,
hidden
b
is a
61
vector representing bias in the hidden layer,
output
w
is
203
a
16
weight matrix in the output layer,
output
b
is a number representing bias in the output
204
layer, and TSV represents the output thermal sensation vote.
205
206
We used the ANN model to predict the TSV of the occupants. According to the PMV
207
thermal comfort model [10], six parameters have an impact on thermal comfort: air
208
temperature, air RH, clothing insulation, air velocity, metabolic rate, and mean radiant
209
temperature. Our measurements showed that the surface temperature of the surrounding
210
walls was almost the same as the room air temperature. Therefore, we assumed that the
211
radiant temperature was the same as the room air temperature. Our measurements also
212
indicated that the air velocity in the offices was lower than 0.2 m/s. According to ASHRAE
213
Standard 55 [47], acceptable comfort zones have air velocity below 0.2 m/s, and thus the
214
impact of air velocity on thermal comfort could be neglected in this study. As for the
215
metabolic rate, a review paper [42] and ISO 8996 [43] have identified a correlation between
216
HR and metabolic rate. High HR typically indicates high metabolic rate. Choi [24] and
217
Kizito [25] also showed that HR was an important factor for predicting individual TSV.
218
Therefore, the HR could be used to predict TSV, replacing metabolic rate in this study.
219
Meanwhile, skin temperature is related to radiative, convective and evaporative heat loss
220
from human skin [48] and is therefore a crucial factor in individual thermal comfort [49].
221
In addition, previous studies have found a correlation between thermal sensation and sweat
222
rate or skin wetness [30-31].
223
224
To predict individual TSV, then, the ANN model in this study required six input parameters:
225
two indoor environmental parameters (air temperature and air RH) and four individual
226
parameters (wrist skin temperature, wrist skin RH, HR and clothing insulation). Therefore,
227
n = 6 in Eq. (6), and the input vector X of the six input parameters is
228
229
[ , , , , , ]T
air air skin skin
T RH T RH HR Clo=X
(7)
230
231
The model output TSV can be expressed as a number from −3 to 3. The collected data were
232
used to train the ANN model so that the predicted TSV would be nearly the same as the
233
collected data in the offices.
234
235
This study used Matlab Neural Network Toolbox [50] in Matlab R2018a to build and train
236
the ANN model. The training targets were the actual TSV that had been collected. For the
237
training process, the Levenberg-Marquardt (LM) algorithm [51] used the following
238
approximation to approach the unknown weight coefficients:
239
240
1
1[]
TT
kk
−
+= − +x x JJ I J e
(8)
241
242
where
J
is the Jacobian matrix that contains first derivatives of the errors with respect to
243
the weights and biases,
I
is the identity matrix, and
e
is the error vector. The damping
244
factor µ was adjusted at each iteration.
245
246
2.3 HVAC control algorithm for thermal comfort
247
248
After training the ANN model, we developed a control strategy for the HVAC system by
249
using the correlation between the physiological data and occupants’ TSV. Fig. 4 shows the
250
working principle of the control strategy for the HVAC system. The lighting occupancy
251
sensor on the ceiling and the Bluetooth receiver in the wristband can sense the occupant’s
252
arrival and departure in the offices. Thus, the BAS can control the on/off status of the
253
HVAC system automatically. The wristband measures the physiological data, including
254
skin temperature, skin RH and HR, every ten minutes. The ANN model then uses the
255
correlation to predict the TSV, and the control system determines whether or not the
256
occupants feel comfortable and the indoor environmental parameters need to be adjusted.
257
If the occupants feel cold, the thermostat set point needs to be raised, and vice versa. If the
258
occupants feel comfortable, the thermostat set point remains unchanged. The process
259
updates every ten minutes, or whenever a new occupant enters or an occupant leaves the
260
room. When the room is unoccupied, the HVAC system is shut down.
261
262
263
Figure 4. Working principle of the control algorithm for using physiological parameters
264
from wristbands
265
266
With the above working principle, the control system was able to calculate comfortable
267
indoor environmental parameters. In a single-occupant office, “comfort” means a neutral
268
feeling and
0TSV =
on the part of the occupant. To enable this neutral feeling, we can
269
solve the air temperature in the input vector with the following equation based on Eq. (6):
270
ln 1 output
hidden hidden
output
b
= − − − −
w
w X b
(9)
271
272
However, in multi-occupant offices, it is typically impossible for every occupant to feel
273
neutral simultaneously, which means that all
0TSV =
is impossible. Rather, thermal
274
comfort in a multi-occupant office implies that all TSV values are close to 0 [51], which
275
means
276
2
min ( 0)
air
TTSV −
277
278
where the summation symbol represents the adding of the
2
TSV
for all occupants of the
279
room. According to the study’s hypothesis and Eqs. (1) through (5), the room air
280
temperature is the variable. Thus, at the minimum we have
281
2
()
20
air air
TSV TSV
TSV
TT
==
(10)
282
With Eq. (10), it is clear that for single-occupant offices,
283
2
()
00
air
TSV TSV
T
= =
284
For multi-occupant offices, by applying the chain rule and using Eq. (6) in Eq. (10), we
285
obtain
286
2
2
()
2 exp( ) [1, , , , 0, 0]
[1 exp( )]
output T
air skin skin
hidden hidden hidden
air hidden hidden air air air
w
TSV RH T RH
TSV
T T T T
= − − − −
+ − −
w X b w
w X b
287
(11)
288
where
air
air
RH
T
can be calculated directly by psychrometric relationship, while
skin
air
T
T
and
289
skin
air
RH
T
can be calculated by Eqs. (2) and (3).
290
Hence, the control system can solve the above equations to find the comfortable air
291
temperature for the offices. Because the current thermostats and control system only accept
292
integers for the set point, the system identifies the closest integer as the thermostat set point
293
with the optimal air temperature.
294
295
As a previous study [6] observed the occupants of multi-occupant offices may compromise
296
according to others’ thermal preferences. When occupants do not know the thermal needs
297
of others, they often choose not to adjust the HVAC system. As a result, thermal comfort
298
for all occupants is hard to achieve, and the room air temperature may become extremely
299
hot or cold. However, because it receives physiological signals from all the occupants in
300
the room, the smart control system knows if some occupants feel uncomfortable and the
301
indoor environment needs to be adjusted. Even if the occupants do not communicate with
302
one another, the smart control system is able to determine the indoor environmental
303
parameters. Therefore, the problem of thermal comfort in multi-occupant offices can be
304
solved.
305
306
2.4 Validation of HVAC control algorithm
307
308
After designing this smart HVAC control system, we needed to experimentally validate its
309
ability to improve thermal comfort in the offices. To do so, we applied the control strategy
310
for the HVAC system in several multi-occupant offices. We used the actual indoor air
311
temperature and RH measured by the data loggers, and physiological parameters measured
312
by the wristbands, as input to the control system. The system adjusted the thermostat set
313
point in response to the measured data. We used a questionnaire to record the occupants’
314
TSV before and after the adjustment. We validated the system in summer, fall and winter,
315
since the clothing levels of the occupants and the indoor RH pattern varied from season to
316
season. However, the experimental validation was time-consuming. The limited number of
317
validation cases covered only a small group of occupants and limited ranges of the control
318
parameters. It was hard that the actual control parameters such as air temperature, skin
319
temperature, skin RH and HR went to extremes, so validating these extreme cases was hard.
320
Therefore, we also used numerical simulations to validate the control system. We simulated
321
the number of occupants in the office, from one to five. Next, we randomly generated the
322
air temperature, air RH, clothing insulation, skin temperature, skin RH and HR as the inputs
323
to the control system. With these inputs, we evaluated and compared the TSV before and
324
after the control of the indoor environment by the developed control system.
325
326
2.5 Energy analysis
327
328
We also analyzed the energy use of the HVAC system in the offices with the developed
329
control strategy by using an energy simulation program. We simulated the heating and
330
cooling loads in the offices with EnergyPlus. We constructed a building geometry model
331
based on the actual HLAB building as shown in Fig. 5, and used the actual properties of
332
the HLAB building and the HVAC system in the energy simulation. This model had been
333
validated previously, and detailed information can be found in [53]. The weather data used
334
in the simulation was that for a typical meteorological year (TMY3). The developed control
335
system with wristbands was able to adjust the thermostat settings based on the indoor
336
environmental and human physiological parameters. We simulated these parameters
337
numerically for the control system and generated the schedule and settings of the HVAC
338
system for the energy simulation.
339
340
(a)
(b)
Figure 5. (a) Photograph of the HLAB building and (b) geometric model of the HLAB
341
building for EnergyPlus simulation
342
343
3. Results
344
The above methods collected the data to train the ANN models for predicting TSV. The
345
correlations between the physiological parameters and occupants’ TSV were then used to
346
improve the overall thermal comfort in multi-occupant offices. Finally, we analyzed the
347
HVAC control system experimentally and though simulations.
348
349
3.1 Data collection
350
Data were collected during three seasons of 2019. In each season, we collected the data for
351
more than three weeks in every multi-occupant office. We obtained over 500 data points
352
from the 24 occupants to train the ANN model. The average data collection duration for
353
each occupant exceeded 5 h.
354
355
Fig. 6 shows the distributions of clothing level, wrist skin temperature and HR in the
356
collected data. In winter, shoulder seasons and summer, the typical clothing level was a
357
sweater with thick pants, a long/short sleeve shirt with pants, and a short sleeve shirt with
358
pants/shorts, respectively. However, some occupants kept almost the same clothing level
359
indoors all year around. As for the skin temperature and HR, the obtained distributions
360
were very similar to those in previous studies [54-55], collected from over 2000 people.
361
Therefore, the bias of the data collection was small.
362
363
(a)
(b)(c)
Figure 6. Distribution of the collected data: (a) clothing level; (b) wrist skin temperature;
364
(c) heart rate. The probability density curves of the collected data were lognormal
365
distributions.
366
367
3.2 ANN model training
368
We used the above collected data from the seasons to train the ANN model by means of
369
the LM algorithm. Fig. 7 displays the training results for TSV. The ANN model was able
370
to predict occupants’ TSV with six input parameters. After training, the prediction fitted
371
the collected data with R2 = 0.89. Compared with the R2 = 0.75 [6] when physiological
372
parameters were not used, the ANN model in this study was more accurate.
373
374
375
376
Figure 7. The training results for the ANN model.
377
378
3.3 Control system for multi-occupant office
379
3.3.1 GUI of the control system
380
381
After training the ANN model, we developed an HVAC control system for the offices. Fig.
382
8 shows the graphical user interface (GUI) of the control system that incorporates
383
wristband data. It is a dynamic GUI and can update the input and control information
384
automatically. The left and right halves of the panel are the input and output fields,
385
respectively. As shown at the top of the input field, we need the indoor air temperature and
386
RH data from the data logger in the office. The small thermometer next to is updated
387
automatically with the input data. Next, in the middle of the GUI, we must select
388
checkboxes to indicate the presence of occupants in the office based on the Bluetooth
389
transmissions from their wristbands. The current system supports a maximum of five
390
occupants. The greater the number of occupants, the more input fields are available for the
391
physiological data. We then need the physiological data from the occupants’ wristbands
392
and their clothing insulation values for the input fields. With these data, the program uses
393
the algorithm presented in Section 2.3 to calculate the optimal air temperature set point and
394
the control behavior. It also uses the ANN model to predict the occupants’ TSV before and
395
after the control behavior. Finally, the GUI displays the TSV, control behavior and
396
thermostat diagram in the output field, on the right side of Fig. 8. The light red arrow and
397
the red arrow in the thermostat diagram point to the current and optimal set points,
398
respectively. For example, Fig. 8 portrays a case with three occupants. The current air
399
temperature is 22.3°C, and the control behavior is raising of the set point by 2°C. Occupants
400
No. 1 and 2 feel slightly cool before the control behavior. After control, they feel almost
401
neutral. However, occupant No. 3 feels neutral before control, but slightly warm after
402
control. The overall thermal comfort in this three-occupant office is improved, but not for
403
all the occupants. If the TSV of some occupants contradicts that of others, the current
404
system can satisfy most but not all of the occupants. This occurs because the goal of the
405
control algorithm is to minimize the summation of TSV2. The system can control only the
406
room air temperature. Hence, further study is needed to provide personalized
407
environmental control and satisfy all occupants.
408
409
To avoid the impact of incorrect measurements by the wristband and data logger on the
410
control system, and to enhance the system’s robustness, the input fields accept only
411
reasonable inputs. The acceptable range of the air temperature is from 15°C to 35°C, wrist
412
skin temperature from 28°C to 36°C, and HR from 50 to 160 bpm. If any input data are
413
outside the acceptable range, the system will display an error message and will not adjust
414
the set point.
415
416
417
Figure 8. The GUI of the developed control system using wristbands with three
418
occupants.
419
420
3.3.2 Experimental validation of the control system
421
After designing this control system, we validated it by experiments in summer, shoulder
422
and winter seasons as described in Section 2.4. Most of the validation cases were conducted
423
in offices with two or three occupants. Only 10% of the cases had one occupant, and 15%
424
had four occupants. As for the clothing level, most occupants wore short sleeve shirts and
425
pants in summer, and long sleeve shirts and pants in shoulder seasons, and adding sweaters
426
in winter. In summer some wore short sleeve shirts and shorts, and only a few wore long
427
sleeve shirts with pants. Some occupants wore short sleeve shirts and pants all year around.
428
429
Fig. 9(a) displays the actual TSV of the occupants, as recorded on a questionnaire before
430
and after the control behavior. Fig. 9(b) shows the TSV predicted by the ANN model that
431
was used in the control system. The predicted TSV exhibited a similar pattern to that of the
432
actual TSV, and this finding further validated the accuracy of the ANN model. The figure
433
also demonstrates that the control system was able to improve the thermal comfort in the
434
office. Before using the control system, over half of the occupants felt uncomfortable,
435
ranging from cool (TSV=-2) to warm (TSV=2). After using the system, almost all the
436
occupants reported a neutral feeling. Fewer than 10% of the occupants still felt slightly
437
cool or slightly warm, while none of the occupants felt cool or warm. Because the control
438
system optimized the overall thermal comfort for all the occupants, some occupants still
439
compromised for the sake of others’ thermal preferences, as in the example shown in Fig.
440
8. Thus, we experimentally validated the developed control system that uses wristbands
441
and the ANN model.
442
(a)
(b)
Figure 9. Results for the experimental validation cases before and after using the
443
developed control system: (a) actual TSV; (b) TSV predicted by ANN model.
444
445
3.3.3 Numerical validation of the control system
446
Since the experimental validation tests were very time-consuming, we also performed
447
numerical simulations. The purpose of these simulations was to increase the test size and
448
explore more cases, especially extreme cases. For numbers of office occupants ranging
449
from 1 to 5, we ran 1000 numerical cases each. We used a uniform distribution to randomly
450
generate air temperatures from 18°C to 25°C, clothing insulation values from 0.36 to 1.3,
451
and RH from 0% to 100% as the input parameters of the control system. As for generating
452
the human skin temperature and HR data, previous studies [56-57] had found that a
453
lognormal distribution was suitable for describing biological and medical phenomena such
454
as growth and metabolic rate. The probability density function of the lognormal distribution
455
was
456
22
(ln ) /2
1
2
x
Pe
x
−−
=
(12)
457
where
and
are the mean and standard deviation of the collected data, respectively.
458
Thus, we randomly generated the skin temperature and HR by using the lognormal
459
distribution, and the probability density curves are shown in Fig. 6. These curves fitted the
460
collected data well.
461
Fig. 10 shows the TSV distribution in the numerical validations before and after the
462
developed wristband control system was used. Since we generated the parameters
463
randomly within a large range, the simulated TSV before control was distributed from cold
464
to hot almost evenly. In a single-occupant room, the wristband control system would
465
always find the neutral temperature at which TSV = 0, as shown in Fig. 10(a). However, if
466
the number of occupants in the room was greater than one, the TSVs of most occupants
467
would still not be optimized after control. Some occupants might still feel slightly cool or
468
warm because of the necessary compromise among different occupants, as in the example
469
shown in Fig 8(b). A similar phenomenon occurred more often in the offices with greater
470
numbers of occupants, as shown in Fig. 10(b) through (e).
471
472
(a)
(b)
(c)
(d)
(e)
Figure 10. Distribution of TSV in the numerical validations before and after the
473
developed wristband control system was used. The number of occupants ranges from one
474
to five in (a) through (e).
475
476
Table 1 lists the improved TSV results in the simulated validation cases. “Improved” TSV
477
means that the absolute value of TSV was reduced after control. In single-occupant offices,
478
TSV would certainly be improved, while in multi-occupant offices most TSVs would be
479
improved. The greater the number of occupants in an office, the harder it would be to
480
improve the overall thermal comfort. This would occur because of the various thermal
481
preferences among the occupants, especially when some occupants have opposing
482
preferences. Therefore, the TSVs of most occupants were between -1 and 1 after control.
483
Feeling cool and feeling warm still existed, but only for a very small number of occupants
484
in extreme cases.
485
486
Table 1. Improved TSV results in the simulated cases
487
Percentage of
improved TSV
TSVf
around 0
TSVf
around 1 or -1
TSVf
around 2 or -2
1 occupant
100%
100%
0%
0%
2 occupants
97%
57%
41%
2%
3 occupants
93%
52%
45%
3%
4 occupants
89%
49%
47%
4%
5 occupants
85%
47%
48%
5%
488
3.3.4 Energy analysis of the control system
489
After analyzing the thermal comfort in the offices with the wristband control system, we
490
simulated the office heating/cooling load with the number of occupants ranging from 1 to
491
5. For each number of occupants, we simulated 1000 cases and obtained the average
492
heating/cooling load. We still generated the parameters randomly as in Section 3.3.3. The
493
wristband control system was able to calculate and adjust different set points for different
494
input values of air temperature, skin temperature, clothing level, etc. Note that every ten
495
minutes the set point was recalculated as shown in Fig. 4. We used the resulting set point
496
schedules in the energy simulation program. We compared the energy use per area in order
497
to eliminate the impact of room size, because the areas of these multi-occupant offices were
498
different. Since all the multi-occupant offices were in the interior zone as shown in Fig. 1,
499
the cooling load dominated. We compared the developed wristband control system with
500
the use of constant set points. On the basis of ASHRAE comfort zone specifications [40],
501
the set points for the winter, shoulder and summer seasons were 27°C, 25°C and 21°C,
502
respectively.
503
504
Table 2 compares the average heating and cooling loads per area between the control
505
system using wristbands and the constant set point for a one-year period. The simulated
506
heating load was almost the same as that with the constant set point, but the cooling load
507
was slightly higher. The difference was less than 7%. We also compared the control
508
systems when coupled with occupancy-based control. There were lighting sensors in the
509
offices that could detect the room occupancy and shut down the HVAC system to save
510
energy. The developed control system could also use the Bluetooth connection with the
511
wristbands to detect the number of room occupants. We found that coupling with
512
occupancy-based control yielded an energy saving of about 90% for heating load and 30%
513
for cooling load, when either the constant set point or wristband control system was used.
514
The reason for the huge energy saving was that the largest heating load occurred when the
515
room was unoccupied. Shutting down the HVAC system could save energy during this
516
period. The energy saving of the wristband control system when the room was occupied
517
was close to that of a similar control system in a previous study [36].
518
519
Table 2 Comparison of average heating and cooling loads per area between the control
520
system using wristbands and the constant set point in a one-year period
521
Load
per area
(W/m2)
Constant
set point
Constant
set point
with
occupancy-
based
control
Wristband
control
Wristband
control
with
occupancy-
based
control
Heating
53.4
7
53.2
6.3
Cooling
98.8
72.7
106.4
72
522
4. Discussion
523
In this study, we used the ANN model to predict the occupants’ TSV by using human
524
physiological data such as wrist skin temperature and HR from wristbands. Because the
525
ANN model developed here was personalized, the accuracy was good. Heart rate was used
526
instead of metabolic rate because the actual metabolic rate was hard to measure under real
527
conditions [58]. However, the heart rate may be influenced by other individualized factors,
528
such as physical fitness, health, mood and age [59-60]. Furthermore, the sensors in the
529
wristbands may sometimes have measured the data inaccurately, for example, if the
530
occupants did not wear the wristbands properly (too tight or too loose). Although we
531
limited the input field, failed measurements would have interfered with and delayed the
532
control system. In addition, the developed control system required a large number of
533
parameters as input in order to control the HVAC system. All the indoor environmental
534
parameters and physiological parameters could be measured automatically, but the clothing
535
insulation level could not. Developing a smart system that detects occupants’ clothing level
536
automatically is a possible improvement for consideration in the future.
537
538
We collected the data in multi-occupant offices and simulated the energy use in these
539
offices. All the offices were in the interior zone, and we neglected the impact of solar
540
radiation and outdoor weather on the occupants’ thermal sensation and the energy use. The
541
developed control system was able to find one optimal set point for all the occupants in a
542
given office. However, it could not satisfy all the occupants if their thermal preferences
543
were in conflict. Therefore, in the future it is necessary to develop a smart HVAC system
544
with zonal control that can satisfy all the occupants.
545
546
5. Conclusions
547
548
In this study, we collected data on skin temperature, skin RH and HR from wristbands worn
549
by occupants in multi-occupant offices. We developed an HVAC control system and
550
validated it by means of experiments and numerical simulations. We also compared the
551
energy use of the wristband control system with constant set point control. This study led
552
to the following conclusions:
553
1) The ANN model predicted the occupants’ TSV accurately with physiological input
554
parameters such as skin temperature, HR and skin RH. This correlation between
555
the physiological parameters and occupants’ TSV could be used for the HVAC
556
control system.
557
2) The wristband control system was capable of improving the overall thermal comfort
558
in multi-occupant offices. The control system was smart and could adjust the
559
thermostat set point automatically in real time. We validated the system by means
560
of both experiments and numerical simulations. In most cases, we improved the
561
occupants’ thermal comfort level. Over half of the occupants reported a neutral
562
feeling, and fewer than 5% of the occupants still felt uncomfortable, after using the
563
control system.
564
3) The energy use by the HVAC system with the wristband control was almost the
565
same as that with the constant set point. Coupling with occupancy-based control,
566
by means of lighting occupancy sensors or Bluetooth, reduced the heating and
567
cooling loads by 90% and 30%, respectively, in the interior offices.
568
569
570
Acknowledgments
571
The authors would like to thank Dr. Orkan Kurtulus of the Center for High Performance
572
Buildings at Purdue University for his kind assistance in setting the building automation
573
system in the HLAB building. We would also like to thank all the occupants in the offices
574
for their participation and assistance in obtaining the data reported in this study.
575
576
Conflict of Interest
577
None.
578
579
References
580
[1] Neil E. Klepeis, William C. Nelson, Wayne R. Ott, John P. Robinson, Andy M. Tsang,
581
Paul Switzer, Joseph V. Behar, Stephen C. Hern, William H. Engelmann, “The national
582
human activity pattern survey (NHAPS): A resource for assessing exposure to
583
environmental pollutants,” J. Expos. Sci. Environment. Epidemiol. 11 (3) (2001) 231.
584
[2] International Facility Management Association, IFMA survey ranks top 10 office
585
complaints, 2003. https://www.buildings.com/news/industry-news/
586
articleid/1689/title/ifma-survey-ranks-top-10-office-complaints-.
587
[3] T Peffer, M Pritoni, A Meier, C Aragon, D Perry. "How people use thermostats in
588
homes: A review." Building and Environment 46.12 (2011): 2529-2541.
589
[4] S Karjalainen, "Thermal comfort and use of thermostats in Finnish homes and offices."
590
Building and Environment 44.6 (2009): 1237-1245.
591
[5] S Derrible, and M Reeder. "The cost of over-cooling commercial buildings in the
592
United States." Energy and Buildings 108 (2015): 304-306.
593
[6] Zhipeng Deng, Qingyan Chen. "Artificial neural network models using thermal
594
sensations and occupants’ behavior for predicting thermal comfort." Energy and Buildings
595
174 (2018): 587-602.
596
[7] Bing Dong, Vishnu Prakash, Fan Feng, Zheng O'Neill. "A review of smart building
597
sensing system for better indoor environment control." Energy and Buildings (2019).
598
[8] Edward Arens, Hui Zhang, Charlie Huizenga. "Partial- and whole-body thermal
599
sensation and comfort—Part I: Uniform environmental conditions." Journal of Thermal
600
Biology 31.1-2 (2006): 53-59.
601
[9] Edward Arens, Hui Zhang, Charlie Huizenga. "Partial- and whole-body thermal
602
sensation and comfort—Part II: Non-uniform environmental conditions." Journal of
603
Thermal Biology 31.1-2 (2006): 60-66.
604
[10] PO Fanger, “Thermal comfort. Analysis and applications in environmental
605
engineering,” Thermal comfort. Analysis and Applications in Environmental Engineering,
606
Danish Technical Press, Copenhagen, 1970.
607
[11] Joyce Kim, Stefano Schiavon, Gail Brager. "Personal comfort models–A new
608
paradigm in thermal comfort for occupant-centric environmental control." Building and
609
Environment 132 (2018): 114-124.
610
[12] Ali Ghahramani, Chao Tang, Burcin Becerik-Gerber. "An online learning approach
611
for quantifying personalized thermal comfort via adaptive stochastic modeling." Building
612
and Environment 92 (2015): 86-96.
613
[13] Qianchuan Zhao, Yin Zhao, Fulin Wang, Jinlong Wang, Yi Jiang, Fan Zhang. "A data-
614
driven method to describe the personalized dynamic thermal comfort in ordinary office
615
environment: From model to application." Building and Environment 72 (2014): 309-318.
616
[14] June Young Park, Zoltan Nagy. "Comprehensive analysis of the relationship between
617
thermal comfort and building control research-A data-driven literature review." Renewable
618
and Sustainable Energy Reviews 82 (2018): 2664-2679.
619
[15] Mohamed Abouelenien, Mihai Burzo, Rada Mihalcea, Kristen Rusinek, David Van
620
Alstine. "Detecting human thermal discomfort via physiological signals." In Proceedings
621
of the 10th International Conference on Pervasive Technologies Related to Assistive
622
Environments, pp. 146-149. ACM, 2017.
623
[16] Y Yao, Z Lian, W Liu, and Q Shen. "Experimental study on physiological responses
624
and thermal comfort under various ambient temperatures." Physiology & Behavior 93, no.
625
1-2 (2008): 310-321.
626
[17] Joon-Ho Choi, Dongwoo Yeom. "Development of the data-driven thermal satisfaction
627
prediction model as a function of human physiological responses in a built environment."
628
Building and Environment 150 (2019): 206-218.
629
[18] Tanaya Chaudhuri, Deqing Zhai, Yeng Chai Soh, Hua Li, Lihua Xie. "Random forest
630
based thermal comfort prediction from gender-specific physiological parameters using
631
wearable sensing technology." Energy and Buildings 166 (2018): 391-406.
632
[19] Joon-Ho Choi, V Loftness. "Investigation of human body skin temperatures as a bio-
633
signal to indicate overall thermal sensations." Building and Environment 58 (2012): 258-
634
269.
635
[20] Joon-Ho Choi, Dongwoo Yeom. "Study of data-driven thermal sensation prediction
636
model as a function of local body skin temperatures in a built environment." Building and
637
Environment 121 (2017): 130-147.
638
[21] Zhibin Wu, Nianping Li, Haijiao Cui, Jinqing Peng, Haowen Chen, Penglong Liu.
639
"Using upper extremity skin temperatures to assess thermal comfort in office buildings in
640
Changsha, China." International Journal of Environmental Research and Public Health 14,
641
no. 10 (2017): 1092.
642
[22] Bo Yi, Joon-Ho Choi. "Facial skin temperature as a proactive variable in a building
643
thermal comfort control system." In Sustainable Human–Building Ecosystems, pp. 117-
644
125. 2015.
645
[23] Andrei Claudiu Cosma, Rahul Simha. "Thermal comfort modeling in transient
646
conditions using real-time local body temperature extraction with a thermographic
647
camera." Building and Environment 143 (2018): 36-47.
648
[24] Joon-Ho Choi, Vivian Loftness, Dong-Won Lee. "Investigation of the possibility of
649
the use of heart rate as a human factor for thermal sensation models." Building and
650
Environment 50 (2012): 165-175.
651
[25] Kizito N. Nkurikiyeyezu, Yuta Suzuki, Guillaume F. Lopez. "Heart rate variability as
652
a predictive biomarker of thermal comfort." Journal of Ambient Intelligence and
653
Humanized Computing 9.5 (2018): 1465-1477.
654
[26] J. P. Carvalho, B. I. L. Barroso, L. B. da Silva, A. I. A. Neves, M. G. L. Torres, C. A.
655
Falcão, J. C. F. Siqueira, A. G. L. Souza, E. L. Souza, J. F. da Silva. "Students’ blood
656
pressure and heart rate in learning environments with thermal changes." International
657
Journal of Occupational and Environmental Safety 2, no. 1 (2018): 29-37.
658
[27] Syed Ihtsham-ul-Haq Gilani, Muhammad Hammad Khan, Muzaffar Ali. "Revisiting
659
Fanger’s thermal comfort model using mean blood pressure as a bio-marker: An
660
experimental investigation." Applied Thermal Engineering 109 (2016): 35-43.
661
[28] Yunchan Shin, Gwanghyun Im, Kwanjong Yu, Honghyun Cho. "Experimental study
662
on the change in driver’s physiological signals in automobile HVAC system under full load
663
condition." Applied Thermal Engineering 112 (2017): 1213-1222.
664
[29] Y. Yao, Z. Lian, W. Liu, C. Jiang, Y. Liu, H. Lu. "Heart rate variation and
665
electroencephalograph-the potential physiological factors for thermal comfort study."
666
Indoor Air 19, no. 2 (2009): 93.
667
[30] JK Sim, S Yoon, Y-H Cho. "Wearable sweat rate sensors for human thermal comfort
668
monitoring." Scientific Reports 8, no. 1 (2018): 1181.
669
[31] Chin-Chi Cheng, Dashang Lee, Bi-Song Huang. "Estimated thermal sensation models
670
by physiological parameters during wind chill stimulation in the indoor environment."
671
Energy and Buildings 172 (2018): 337-348.
672
[32] Wenjie Ji, Bin Cao, Yang Geng, Yingxin Zhu, Borong Lin. "Study on human skin
673
temperature and thermal evaluation in step change conditions: from non-neutrality to
674
neutrality." Energy and Buildings 156 (2017): 29-39.
675
[33] Yiwen Jian, Xiaoyan Chang, Yimin Wu, Meng Gao, Yuanquan Tian. "Study on
676
dynamic change of skin temperatures in actual air-conditioned environment and its effects
677
on air conditioning OFF behavior." Procedia Engineering 205 (2017): 3389-3396.
678
[34] Shichao Liu. "Personal thermal comfort models based on physiological parameters
679
measured by wearable sensors." (2018). Proceedings of 10th Windsor Conference:
680
Rethinking Comfort.
681
[35] Soo Sim, Myung Koh, Kwang Joo, Seungwoo Noh, Sangyun Park, Youn Kim, Kwang
682
Park. "Estimation of thermal sensation based on wrist skin temperatures." Sensors 16, no.
683
4 (2016): 420.
684
[36] Wei Li, Jili Zhang, Tianyi Zhao. "Indoor thermal environment optimal control for
685
thermal comfort and energy saving based on online monitoring of thermal sensation."
686
Energy and Buildings 197 (2019): 57-67.
687
[37] Da Li, Carol C. Menassa, Vineet R. Kamat. "Personalized human comfort in indoor
688
building environments under diverse conditioning modes." Building and Environment 126
689
(2017): 304-317.
690
[38] Andrei Claudiu Cosma, Rahul Simha. "Machine learning method for real-time non-
691
invasive prediction of individual thermal preference in transient conditions." Building and
692
Environment 148 (2019): 372-383.
693
[39] Wei Li, Jili Zhang, Tianyi Zhao, Ruobing Liang. "Experimental research of online
694
monitoring and evaluation method of human thermal sensation in different active states
695
based on wristband device." Energy and Buildings 173 (2018): 613-622.
696
[40] ASHRAE, Handbook Fundamentals, Atlanta (2017).
697
[41] Yanfeng Liu, Lijuan Wang, Yuhui Di, Jiaping Liu, Hao Zhou. "The effects of clothing
698
thermal resistance and operative temperature on human skin temperature." Journal of
699
Thermal Biology 38, no. 5 (2013): 233-239.
700
[42] Jonathan A. Green, "The heart rate method for estimating metabolic rate: review and
701
recommendations." Comparative Biochemistry and Physiology Part A: Molecular &
702
Integrative Physiology 158, no. 3 (2011): 287-304.
703
[43] ISO 8996:2004 Ergonomics of the thermal environment — Determination of
704
metabolic rate.
705
[44] Jörn von Grabe. "Potential of artificial neural networks to predict thermal sensation
706
votes." Applied Energy 161 (2016): 412-424.
707
[45] Jian Liang, Ruxu Du. "Thermal comfort control based on neural network for HVAC
708
application." In Proceedings of 2005 IEEE Conference on Control Applications, 2005.
709
CCA 2005., pp. 819-824. IEEE, 2005.
710
[46] Weiwei Liu, Zhiwei Lian, Bo Zhao. "A neural network evaluation model for
711
individual thermal comfort." Energy and Buildings 39, no. 10 (2007): 1115-1122.
712
[47] Standard, ASHRAE, 55 (2013), Thermal environmental conditions for human
713
occupancy (2013).
714
[48] Katarina Katić, Rongling Li, Wim Zeiler. "Thermophysiological models and their
715
applications: A review." Building and Environment 106 (2016): 286-300.
716
[49] Farzin Davoodi, Hassan Hassanzadeh, Seyed Alireza Zolfaghari, George Havenith,
717
Mehdi Maerefat. "A new individualized thermoregulatory bio-heat model for evaluating
718
the effects of personal characteristics on human body thermal response." Building and
719
Environment 136 (2018): 62-76.
720
[50] Mark Hudson Beale, Martin T. Hagan, Howard B. Demuth. Neural net- work toolbox
721
getting started guide. Matlab, Mathworks, R2017a. https://www.
722
mathworks.com/help/nnet/gs/fit- data- with- a- neural- network.html.
723
[51] Manolis IA Lourakis. "A brief description of the Levenberg-Marquardt algorithm
724
implemented by levmar." Foundation of Research and Technology 4.1 (2005): 1-6.
725
[52] Wanni Zhang, Chenlu Zhang. "Maximize Thermal Comfort in Open-plan Offices by
726
Occupant-oriented Control based on Individual Thermal Profile." ASHRAE Transactions
727
125 (2019) : 167-175.
728
[53] Zhipeng Deng, Qingyan Chen. "Simulating the Impact of Occupant Behavior on
729
Energy Use of HVAC Systems by Implementing a Behavioral Artificial Neural Network
730
Model." Energy and Buildings (2019). 216-227
731
[54] Taku Inoue, Saori Oshiro, Kunitoshi Iseki, Masahiko Tozawa, Takashi Touma,
732
Yoshiharu Ikemiya, Shuichi Takishita. "High heart rate relates to clustering of
733
cardiovascular risk factors in a screened cohort." Japanese Circulation Journal 65, no. 11
734
(2001): 969-973.
735
[55] Taku Inoue, Kunitoshi Iseki, Chiho Iseki, Yusuke Ohya, Kozen Kinjo, Shuichi
736
Takishita. "Effect of heart rate on the risk of developing metabolic syndrome."
737
Hypertension Research 32, no. 9 (2009): 801.
738
[56] Julian S. Huxley. (1932). Problems of Relative Growth. London.
739
[57] Jyrki T. Kuikka. "Scaling laws in physiology: relationships between size, function,
740
metabolism and life expectancy." International Journal of Nonlinear Sciences and
741
Numerical Simulation 4, no. 4 (2003): 317-328.
742
[58] HooSeung Na, Joon-Ho Choi, HoSeong Kim, Taeyeon Kim. "Development of a
743
human metabolic rate prediction model based on the use of Kinect-camera generated visual
744
data-driven approaches." Building and Environment (2019): 106216.
745
[59] Rodney K. Dishman, Janet Buckworth. "Increasing physical activity: A quantitative
746
synthesis." Medicine and Science in Sports and Exercise 28, no. 6 (1996): 706-719.
747
[60] Joel W. Hughes, Catherine M. Stoney. "Depressed mood is related to high-frequency
748
heart rate variability during stressors." Psychosomatic Medicine 62, no. 6 (2000): 796-803.
749
750
751
Highlights
752
Accurately predict TSV in offices in real time with physiological signals from wristband
753
754
Develop and validate a smart HVAC control system in multi-occupant offices by using
755
physiological and indoor environmental parameters
756
757
Improve the overall thermal comfort in multi-occupant offices by using the wristband
758
control system
759
760
Energy use of the developed wristband control system was almost the same as using
761
constant set point
762
