中国科技论文在线 http://www.paper.edu.

cn

Energy and Buildings 39 (2007) 1115–1122

A neural network evaluation model for individual thermal comfort

Weiwei Liu, Zhiwei Lian *, Bo Zhao
Institute of Refrigeration and Cryogenics, Shanghai Jiao Tong University, Shanghai 200240, China
Received 6 December 2006; accepted 12 December 2006

Abstract
An evaluation model for individual thermal comfort is presented based on the BP neural network. The train data came from a thermal comfort
survey. The evaluation results of the model showed a good match with the subject’s real thermal sensation, which indicated that the model can be
used to evaluate individual thermal comfort rightly. Taken a room air conditioner as an example, the application of the NNEM in creating a
microenvironment for individual was discussed. The result showed that the NNEM can play an important role of connecting individual thermal
comfort with the control on the air conditioner.
# 2007 Elsevier B.V. All rights reserved.

Keywords: Thermal comfort; Neural network; Individual; Room air conditioner

1. Introduction subjective sensation. That means, in a comfortable thermal

environment (based on the thermal comfort evaluation for most
A comfortable indoor thermal environment is favorable for people), a single person may still feel uncomfortable. There-
occupants’ health and helps improve working efficiency, which fore, a question arises: is it possible to establish an evaluation
is attached more and more attention. Human thermal comfort is system of thermal comfort for individual? However, as
defined as ‘‘that condition of mind which expresses satisfaction mentioned above, these existing evaluation models were not
with the thermal environment’’ [1]. In order to establish a fully suitable for individual. This paper attempts to develop an
favorable indoor thermal environment, first of all, human evaluation model of thermal comfort for individual based on the
thermal comfort under an environment should be evaluated neural network theory and discuss about its application in
correctly. In the original studies on evaluation of thermal thermal comfort control for an air conditioning environment.
comfort, the boundaries between the scales of human thermal
response, which are essentially fuzzy, were accurately defined 2. A neural network thermal comfort evaluation model
by traditional math methods [2], which sometimes, however,
led to imprecise results. Therefore, a more scientific evaluation Up to now, neural network models (NNM) were successfully
system, the comprehensive fuzzy evaluation model, was applied in control of HVAC system and prediction of cooling/
developed based on the fuzzy mathematical theory [3]. And, heat load in buildings [5–7]. However, only a limited number of
a new approach based on fuzzy logic to estimate the thermal studies were devoted to application of NNM in thermal comfort
comfort level was presented in reference [4]. The common evaluation. Atthajariyakul and Leephakpreeda proposed a
ground of all these methods was that they were used to evaluate practical approach to determine the PMV index of thermal
a large group of persons’ thermal comfort levels. comfort via neural computing in a wide range of human
In general, different individual has different scale of thermal variables and environmental variables for an HVAC control
comfort, due to the physiological difference and the variety of system [8]. Their thermal comfort model was developed for
colony, not for individual.
In this paper, a neural network evaluation model (NNEM) of
Abbreviations: BP, back propagation; CC, comfort chooser; NNC, neural
indoor thermal comfort is established based on the back
network controller; NNCS, neural network control system; NNEM, neural
network evaluation model; NNM, neural network model propagation algorithm (BP neural network [9]), which can be fit
* Corresponding author. Tel.: +86 21 34204263; fax: +86 21 34206814. for individual thermal comfort. The BP neural network has an
E-mail address: [email protected] (Z. Lian). important characteristic of predicting nonlinear systems and
0378-7788/$ – see front matter # 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.enbuild.2006.12.005
转载
中国科技论文在线 http://www.paper.edu.cn
1116 W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122

values: 0, 0.5, and 1, where 0 stands for cool, 1 for warm, and
Nomenclature 0.5 for comfort. Here the thermal sensation was not divided
into the seven stages proposed by P.O. Fanger, because the
d expected output of the output layer
seven stages were developed based on a large group of human
Hj output parameter of hidden layer
thermal feeling, while the NNEM is expected to evaluate
Wj weight from hidden layer to output layer
individual thermal comfort, and as to each individual, he need
Wij weight from input layer to hidden layer
not distinguish cool or warm from slightly cool or warm.
Xi input parameter of input layer
Certainly, the actual outputs of the NNEM were close to the
Y output parameter
three values (0, 0.5, and 1). For example, an output parameter
Greek symbols of 0.99 represents a warm environment and 0.51 means
u threshold value of input layer unit comfort.
uj threshold value of hidden layer unit According to the output (one parameter) and the input (four
parameters), five hidden layer nodes were selected. The NNEM
Subscript is shown in Fig. 1.
t the tth learning sample
2.2. Arithmetic

some tangible merits, such as fast convergence rate and strong In the NNEM, there were 4 input layer units and 5 hidden
false compatibility. Therefore, when applied to the thermal layer units. The transfer from input layer to hidden layer is as
comfort domain, it can automatically learn according to follows:
variations of indoor environment parameters and evaluate X 
4
thermal comfort levels. Hj ¼ f W i j X i  u j ; J ¼ 1; 2; 3; 4; 5 (1)
i¼1
2.1. Mathematical modeling
1
f ðsÞ ¼ (2)
Numerous factors, such as physical factors, physiological 1 þ es
factors and mental factors, affect the human thermal comfort. The transfer from hidden layer to output layer is as follows:
However, it is hard to consider all these factors when evaluating X 
5
the human thermal comfort. Studies indicated that the thermal
Y¼ f W jH j  u (3)
comfort evaluation index comprises of two human factors:
j¼1
human activity and clothing insulation and four environmental
factors: air temperature, air humidity, air velocity and mean After entering the four environmental factors, Y is output by the
radiant temperature [10–13]. Because of the uncertainties of the NNEM. There is an error between Y and the target value before
human factors, the environmental factors are usually con- modifying the weights Wij and Wj, which is used to modify the
sidered as the main factors affecting the human thermal weights. The calculation of weights was repeated until the error
comfort. Therefore, the NNEM in this paper chose the four fell into the range of the desired tolerance.
environmental factors: air temperature, air humidity, air The gradient descent method is adopted as the learning
velocity and mean radiant temperature as input parameters. algorithm, and the square error function is applied as,
There was one output parameter in the NNEM, indicating 2
the thermal comfort level. This parameter had three different Et ¼ 12ðdt  Y t Þ (4)

Fig. 1. Diagrammatic sketch of neural network evaluation model of indoor thermal comfort.
中国科技论文在线 http://www.paper.edu.cn
W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122 1117

3.1. Subjects and instruments

Sixty five male and 48 female volunteers with the following

characteristics (mean) accepted the survey: age: 26.7 years,
height: 165 cm, weight: 55 kg. All the volunteers were in good
health.
Four basic environmental parameters (air temperature, air
velocity, relative humidity and mean radiant temperature) were
measured while the subjects were completing their ques-
tionnaires. The measurement instruments are shown in Table 1.
The accuracy of the instruments met the requirement of
ASHRAE Standard 55 [10] and ISO 7726 [14].

3.2. Thermal comfort questionnaire

Subjects were asked to complete a questionnaire during the

survey. The questionnaire was based on four sections: personal
information, clothing and activity, thermal sensation and
thermal comfort level, environmental parameters. The personal
information included gender, age, height and weight. The
section of clothing and activity required the subjects to
Fig. 2. Arithmetic process of the NNEM. carefully record what they dress and what they are doing during
the survey. The subjects’ thermal sensation was tested using 7-
The variety in the weights from hidden layer to output layer point ASHRAE scale, as shown in Fig. 3, and the thermal
is: comfort was only divided into two levels: comfortable and
uncomfortable. The environmental parameters near the subjects
@E @E @Y
DW j ¼ h ¼ were also recorded in the questionnaire.
@W j @Y @W j
X
5  3.3. Protocol
0
¼ hðd  YÞ f W j H j  uÞH j (5)
j¼1
The survey was carried out in 20 offices and 10 computer
And the variety in the weights from input layer to hidden rooms in Shanghai of China during June–October 2004. In the
layer is: survey, the subjects’ thermal sensation and thermal comfort
level was inquired under various combinations of air
@E @E @Y @H j temperature and velocity. The air temperature and the air
DW i j ¼ h ¼ h
@W i j @Y @H j @W i j velocity were not restricted in some given values, but changed
X
5  in a range, randomly, in order to bring the subjects different
¼ hðd  YÞ f 0 W jH j  u W j f 0 thermal sensation. The air temperature was controlled in 20–
j¼1 35 8C by room air-conditioners (Sometimes, nature ventila-
X
4 
tion was chose). At the same time, the air velocity was
 W i j X i  u j ÞX i (6) regulated in almost 0–0.8 m/s by the room air-conditioners or
i¼1
fans.
Fig. 2 depicts the arithmetic process of the NNEM. Usually, thermal comfort experiments require fixed clothing
and activity level. However, the survey allowed the two human
3. Human thermal comfort survey factors to change, which means that the subjects can wear
different clothes and do different work under various thermal
In order to provide data for training the NNEM, an environments. By doing so, the survey can provide training data
experiment was carried out in the manner of thermal comfort of human thermal comfort with different clothing and activity
survey. for the NNEM.
Table 1
Location of measurements point and accuracy
Parameter Accuracy ASHRAE or ISO standard Measurement instrument
Air temperature 0.2 8C (25 to +75 8C) 0.2 8C TESTO 110
Air velocity 2% full scale 0.05 m/s EY3-2A
Relative humidity 0.5 8C 0.5 8C WS508D
Mean radiant temperature 0.1 8C 0.2 8C 150 mm black-bulb
中国科技论文在线 http://www.paper.edu.cn
1118 W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122

Table 2
The training samples of the NNEM
No. Air Air Air relative Mean radiant Thermal
temperature velocity humidity temperature comfort
(8C) (m/s) (%) (8C) level
Fig. 3. Seven grades of human thermal sensation. 1 27 0.25 50 27 0.5
2 22 0.38 64 22 0
3 29 0.28 45 30 1
The subjects were asked to learn the questionnaire before the
4 25 0.06 44 27 0.5
survey. And, after staying for at least 30 min in an experimental 5 26 0.2 80 27 0.5
condition, the subjects began to fill the questionnaire. Every 6 33 0.09 75 33 1
subject accepted 20–40 tests under different combinations of air 7 25 0.12 40 28 0.5
temperature and velocity. 8 29 0.18 90 32 1
9 30 0.3 60 32 1
10 27 0.1 50 33 1
4. Realization of the NNEM 11 21 0.03 32 20 0
12 25 0.02 37 26 0
In this section, trained by data from the human thermal 13 22 0.04 50 23 0.5
comfort survey, the NNEM can be used to evaluate individual 14 22 0.32 60 26 0
15 26 0.08 60 26 0.5
thermal comfort.
16 34 0.05 80 35 1
17 27 0.13 75 29 1
4.1. Method 18 26 0.14 70 26 0.5
19 27 0.11 60 30 1
Considering the change of the environmental factors and the 20 30 0.21 80 33 1
human factors in actual circumstances, the NNEM adopts the
following method to deal with the experimental data. First, the comfortable. The output of the NNEM was 0.5065. The two
NNEM learns one subject’s thermal comfort scale with one values agreed pretty well. So, according to the output of the
kind of clothing and activity level. And then, adapt the NNEM NNEM, 0.5065, the subject’s thermal comfort can be obtained.
to the same subject’s new thermal comfort scale with another In actual application, with rational initialization of weights
kind of clothing and activity level. Finally, make the NNEM fit and thresholds, the error can be properly broadened, so that the
different subjects’ thermal comfort scale. NNEM can converge rapidly.
As mentioned above, subjects’ thermal sensation was
assessed using 7-point ASHRAE scale in the thermal comfort 4.3. Fixed subject, variable clothing and activity
survey. However, when training the NNEM, only three levels
need to be considered: uncomfortable cool (including cool or As described in Section 4.2, the subject’s thermal comfort
cold), comfort (including cool, slightly cool, slightly warm and level with one kind of clothing and activity level (circumstance
warm) and uncomfortable warm (including warm or hot). A) can be rightly evaluated by the NNEM. Here, the NNEM
was required to be effective for another kind of clothing and
4.2. Fixed subject, clothing and activity activity level (circumstance B). Therefore, the samples under
circumstance B was input to the NNEM one by one. Modified
Taking a subject as an example, his thermal comfort levels by n samples, the NNEM can also rightly evaluate the subject’s
(fixed clothing and activity) in 20 different indoor circum- thermal comfort level under circumstance B. The number n of
stances were chosen as the training data of the NNEM, as samples was determined in the learning process. For example,
shown in Table 2. The subject wore short-sleeved clothing, thin n = 3 means that the NNEM can fit the subject’s thermal
trousers, silk stockings and slippers. During the experiment, he comfort scope under circumstance B after being modified by
did office work, such as writing or reading. three samples.
In Table 2, every row is a training sample. The four Fig. 4 shows the curvy of number of new samples—veracity,
environmental factors of each sample were linearly mapped to as error was 0.01 and 0.1 respectively. The abscissa of Fig. 4
zone (1,1) as the input of the NNEM. The number of the denotes the number of the new samples (under circumstance B)
training samples was 20. The learning rate took 0.01. When the added to the NNEM after the NNEM learned all the 20 samples
error was 0.1 and the weights and thresholds were set randomly, under circumstance A. When a new sample was added, an old
the NNEM converged after 3000 operations. By learning, the sample under circumstance A was deleted. Therefore, the total
weights were modified, and then the NNEM can evaluate the number was always 20. The ordinate denotes the veracity of
subject’s thermal comfort scale. evaluation on the subject’s thermal comfort level under
Table 3 illustrates the comparison between the output values circumstance B via the NNEM, which varied with the number
of the NNEM and the actual values, from which it can be seen that of new samples.
the evaluation results matched the subject’s real thermal comfort As shown in Fig. 4, the veracity of evaluation rose with the
level very well. For example, in the sample of No 21, the subject’s increasing number of new samples. The NNEM had a veracity
actual thermal comfort level was 0.5, which means that he felt of about 80% when the number of new samples reached 20.
中国科技论文在线 http://www.paper.edu.cn
W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122 1119

Table 3
Validation of the NNEM
No. Air temperature Air velocity Air relative Mean radiant Subject’s thermal Output of
(8C) (m/s) humidity (%) temperature (8C) comfort level the NNEM
21 25 0.18 60 27 0.5 0.5065
22 29 0.15 67 30 1 0.9904
23 25 0.5 50 26 0 0.0218
24 22 0.25 45 24 0 0.0132

Fig. 4. Curvy of number of new samples—veracity with variable clothing and activity.

4.4. Variable subjects evaluates individual thermal comfort, it seems that the NNEM
can achieve good effects in creation of comfortable environ-
When evaluating different subjects’ thermal comfort, ment for individual.
subject B’s samples were added to the NNEM one by one,
like dealing with the problem of variable clothing and activity 5.1. A neural network control system for a room air
in Section 4.3. At the same time, the veracity of evaluation on conditioner
the subject B’s thermal comfort level via the NNEM was
verified. It can be predicted that the condition of variable In China, room air conditioners are broadly used in personal
subjects exerted the same influence on the NNEM as that of offices and families. Generally, a room air conditioner only
variable clothing and activity. Fig. 5 illustrates the curvy of controls indoor temperature, which can be easily realized, but
number of new samples—veracity based on two subjects’ always lead to unsatisfied indoor environment. If the room air
samples, whose thermal feelings were distinct. The curvy in conditioner can automatically judge the owner’s thermal
Fig. 5 demonstrates the same trend as in Fig. 4, which indicates comfort and determine the comfortable environment para-
that the NNEM can also adapt to the new subject’s thermal meters, it will be more significative to improve thermal comfort
comfort scope very well. level for individual. In this section, the room air conditioner was
taken as an example to discuss the application of the NNEM.
5. Discussion on application of the NNEM Here, a neural network control system (NNCS) of a room air
conditioner was designed, which is shown in Fig. 6. For the
When creating a microenvironment for individual, such as room air conditioner, indoor air temperature and velocity can be
room air conditioners, how to control the environment factors to controlled by regulating frequency of the compressor, angle of
bring an occupant comfortable feelings is especially important. vanes in the airflow outlet and speed of the fan. Therefore, the
As mentioned before, the comfortable conditions obtained three parameters were adopt as control variables. The neural
according to a large group of persons’ thermal feelings may not network controller (NNC) was desired to output a reasonable
be suitable for every person. Because the NNEM directly combination of the three controlled variables according to the

Fig. 5. Curvy of number of new samples—veracity with variable subjects.

中国科技论文在线 http://www.paper.edu.cn
1120 W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122

Fig. 6. A neural network control system of a room air conditioner.

four basic environmental parameters. The neural network was The four environmental parameters (air temperature, air
trained according to the deviation of output values between the velocity, air humidity, and mean radiant temperature) measured
NNC and the training part. By learning, the neural network can by sensors are input to the NNEM. The output of the NNEM
reflect the mapping relation between the environmental means the thermal comfort levels, which is in the range of 0–1.
parameters and the controlled variables. During the training process of the NNEM, the output is given a
Finished training the NNC, the NNCS was an open loop factual meaning that 0.5  e stands for comfort, 0 + e for cool,
control system without error, which can achieve better and 1  e for warm, with e being the set error. The output value
controlling results. is compared with the preset value of 0.5 to get deviation 1.
Deviation 1 is input to the comfort chooser (CC), which works
5.2. Training part based on the NNEM like a trigger.
Signal 1 has three values: 0, 0.5 and 1, which corresponds to
According as Fig. 6, the training part directly provided the the occupant’s actual thermal sensation of cool, comfort and
controlled parameters for the NNC, which can determine the warm, respectively. This signal is input by the occupant. In fact,
indoor thermal comfort level. Considering the advantage of if three keys (cool, comfort and warm) are set as the input of the
evaluating individual thermal comfort, the NNEM was applied signal 1 in the remote device of the room air conditioner, the
in the training part as the core. The NNEM learnt the occupants’ occupant only needs to select the corresponding key according
thermal comfort scope based on their thermal sensation and the to his thermal feeling.
four environmental parameters. The thermal comfort scope When the occupant inputs signal 1, the difference between
obtained by the NNEM provided a basis for the NNC to control signal 1 and the preset value 0.5 is deviation 2, and the
the environmental parameters. Fig. 7 depicts the structure of the difference between signal 1 and the output of the NNEM is
training part. deviation 3. Like deviation 1, deviation 2 is also input to CC.

Fig. 7. Training part of the indoor air-conditioner control system.

中国科技论文在线 http://www.paper.edu.cn
W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122 1121

Deviation 3 returns to the NNEM to modify the weights of the especially air temperature and air velocity. Thus, it is
neural network, so that the NNEM can evaluate the occupant’s necessary to study how to design the optimal combination
thermal comfort scope better. with the maximum benefit in terms of individual thermal
Under most time, the NNEM can rightly evaluate the comfort and energy utilization.
occupant’s thermal comfort level in current environment, (2) Control principles of an individual air conditioner. Once the
therefore the occupant does not need to input signal 1. At this optimal combination of environmental parameters is
time, the output of the CC is only deviation 1, meaning that determined, how to control the air conditioner to realize
there is not deviation 2. However, when the NNEM does not the combination, with high efficiency and low energy
entirely learn the occupant’s thermal comfort scope or the consumption, is another problem.
occupant’s thermal comfort scope changes due to different
clothing or activity, there is an error between the output of the 6. Conclusions
NNEM and the occupant’s actual thermal sensation. Therefore,
he needs to input signal 1 via the remote device, and the output In the paper, a neural network evaluation model for
of the CC is deviation 2 (deviation 1 does not act). individual thermal comfort was developed based on the back
The signals of the four environmental parameters from the propagation algorithm. Compared with the experimental data
sensors are input to the register. However, the register only from the human thermal comfort survey, the evaluation results
stores the latest favorable environmental signals. Hence, the of the NNEM showed a good match with the subject’s real
register instruction is decided by the output value of the CC. thermal sensation, which indicated this model can be used to
Once the output of CC means comfort, the register will evaluate individual’s thermal comfort, rightly.
store the current four environmental signals and then output The application of the NNEM in creating a microenviron-
them. ment for individual was discussed. Taken the room air
Four deviations (deviation 4–7) are obtained by subtracting conditioner as an example, a neural network control system
the output of the register from the signals of the four current was designed. In the system, the NNEM plays an important role
environmental parameters. Thus, the four deviations represent of connecting individual thermal comfort with the control on
the difference between the current four environmental the air conditioner. Therefore, with the application of the
parameters (measured by the sensors) and the latest four NNEM, the room air conditioner can be more significative to
favorable values (stored in the register). They are input to the improve thermal comfort level for individual.
NNC. Here, considering the individual’s diversity and simplify
control, the four deviations instead of the current four Acknowledgements
environmental parameters are used as the input of the NNC.
If the current four environmental parameters are directly input This project is financially supported by National Natural
to the NNC, a control principle must be designed for each room Science Foundation of China (No. 50478018). And the authors
air conditioner based on every owner’s thermal sensation. also want to thank Mr. S.G. Tan for his improvement of the
However, when the four deviations used, a universal control English of the paper.
principle can be designed.
At last, based on the four deviations (deviation 4–7), the References
NNC adopts appropriate control arithmetic to get right values of
frequency of the compressor, angle of vanes in the airflow outlet [1] ASHRAE, Standard 55. Thermal Environmental Conditions for Human
Occupancy, American Society of Heating, Refrigerating and Air-Con-
and speed of the fan, and regulate them.
ditioning Engineers, Atlanta, GA, 2004.
[2] P.O. Fanger, Thermal Comfort—Analysis and Application in Environ-
5.3. Key problems mental Engineering, Danish Technical Press, Copenhagen, 1970.
[3] Z.W. Lian, H.Y. Feng, A fuzzy evaluation model on indoor thermal
With the application of the NNEM, the room air conditioner environment inside a building, Journal of Shanghai Jiao Tong University
36 (2) (2002) 169–172.
can be treat as ‘‘individual air conditioner’’. Obviously, in the
[4] M. Hamdi, G. Lachiver, F. Michaud, A new predictive thermal sensation
NNCS, the NNEM plays an important role of connecting index of human response, Energy and Buildings 29 (1999) 167–178.
individual thermal comfort with the control on the air [5] O. Ahmed, J. Mitchell, S. Klein, Application of general regression neural
conditioner. network (GRNN) in HVAC process identification and control, ASHRAE
However, following two key problems must be solved in the Transactions 102 (1996) 1147.
[6] G.P. Henze, R.E. Hindman, Control of air-cooled chiller condenser fans
future work, before the whole NNCS can be successfully
using clustering neural networks, ASHRAE Transactions 108 (2002)
realized: 4574.
[7] S.A. Kalogirou, M. Bojic, Artificial neural networks for the prediction of
(1) Determination on the optimal combination of environ- the energy consumption of a passive solar building, Energy 25 (2000)
mental parameters. It is known that different combination of 479–491.
[8] S. Atthajariyakul, T. Leephakpreeda, Neural computing thermal comfort
air temperature, air velocity, air humidity, and radiant
index for HVAC systems, Energy Conversion and Management 46 (2005)
temperature can lead to different thermal sensation, and 2553–2565.
the same thermal comfort level can also be achieved [9] R. Hechi-Nielsen, Theroy of the back propagation Neural Network,
with various combination of environmental parameters, Proceedings of the IJCNN (1989).
中国科技论文在线 http://www.paper.edu.cn
1122 W. Liu et al. / Energy and Buildings 39 (2007) 1115–1122

[10] ASHRAE, ASHRAE Fundamentals 2001, Thermal Comfort, American [13] X.J. Ye, H.L. Lu, L. Dong, B.Y. Sun, Y.M. Liu, Thermal comfort and air
Society of Heating, Refrigerating and Air-Conditioning Engineers, quality in passenger rail car, International Journal of Ventilation 3 (2)
Atlanta, GA, 2001. (2004) 183–191.
[11] J. Toftum, Human response to combined indoor environment exposures, [14] ISO, International Standard 7730, Moderate Thermal Environments—
Energy and Buildings 34 (2002) 601–606. Determination of the PMV and PPD Indices and the Specification of
[12] D.A. McIntyre, Indoor Climate, Applied Science Publishers Ltd., Lon- Conditions for Thermal Comfort, 2nd ed., International Standards
don, 1980. Organization, Geneva, 1994.