A Neuro-Fuzzy Sugeno-Style HVAC Control System for Balancing

Thermal Comfort and Energy Consumption

Hoba H. Bakr, Nawzad K. Al-Salihi and Oussama H. Hamid

Department of Computer Science and Engineering, University of Kurdistan Hewl

er, Kurdistan Region, Iraq

Keywords:

Neuro-Fuzzy Controller, HVAC Control Systems, Thermal Comfort, Indoor Environment Quality, Energy

Consumption.

Abstract:

Thermal comfort is an environmental state, in which humans enjoy calefactory conditions while being indoor

and wearing a normal amount of clothing. To achieve this, the indoor environment’s temperature should be

adjusted in accordance with the temperature variations of the outdoor space, taking into account the resulting

energy costs. We studied this problem by designing a neuro-fuzzy HVAC control system that provides a

higher indoor environment comfort while decreasing the corresponding energy consumption. Our controller

utilizes a Sugeno-style fuzzy inference system with two sensory inputs: one for temperature and another for

occupants’ motion. It outputs a signal that represents the mode of the air conditioner and the compressor

speed. Simulation results showed that the air conditioner turns off automatically after 10 minutes of the last

detected motion. Furthermore, running the simulations for the energy consumption and resulting costs, both

variables were shown to fall in the absence of occupants’ motion.

1 INTRODUCTION

‘Thermal comfort’ describes an environmental

state, in which a human individual does not suf-

fer from cold or hot temperatures while being

indoor and wearing a normal amount of clothing

(Canadian Centre for Occupational Health and Safety,

2017). Such a state is important for peoples’ health

and is expected to increase their personal satis-

faction and work productivity. Moreover, out of

practical rationale, it is recommended to set the

temperature in winter between 21

◦

C and 23

◦

C while

in summer it should be set at warmer temperature

to decrease the ﬂowing of heat from outdoors to

indoors so as to save energy. These temperature

settings have been conﬁrmed to meet the needs

of 80% of people and, hence, are recommended

by American Society of Heating, Refrigerat-

ing, and Air Conditioning Engineers (ASHRAE)

(Canadian Centre for Occupational Health and Safety,

2017).

Heating, ventilating, and air conditioning (HVAC)

is a control technology that aims at improving indoor

environment quality by increasing thermal comfort

while decreasing energy consumption (Nowak and

Urbaniak, 2016). Traditional approaches to regulate

the operations of an HVAC system include the use of

‘on-off’ and ‘Proportional Integral Derivative’ (PID)

controllers.

On-off controllers are the most intuitive and sim-

plest of control techniques. They regulate the under-

lying process by adjusting the temperature differential

between two set values ‘low’ and ‘high’ with respect

to the desired room temperature. A control process

that underlies the on-off principle leads to a rise of

room temperature when the heating signal is on (it

falls in case of cooling signal) until it hits the value

‘high’. The control signal then turns off and the tem-

perature starts falling (it rises in case of cooling) un-

til it reaches the value ‘low’. Though easy to imple-

ment, control processes utilizing an on-off controller

display large ﬂuctuations of temperature and are in-

capable to control processes with time delays (Afram

and Janabi-Shariﬁ, 2014).

PID controllers were developed to reduce the im-

pact of ﬂuctuations caused by on-off control pro-

cesses (Song et al., 2015). They achieve accurate

control by utilizing error dynamics related to the con-

trolled variable. Despite their relatively promising

results, PID controllers, however, need to be fre-

quently tuned in accordance with the operating con-

ditions. This could be time consuming (Wang et al.,

2001). Otherwise, performance of the controller will

decrease. Moreover, embedding an auto-tuning com-

H. Bakr H., K. Al-Salihi N. and H. Hamid O.

A Neuro-Fuzzy Sugeno-Style HVAC Control System for Balancing Thermal Comfort and Energy Consumption.

DOI: 10.5220/0006558904190426

ponent (e.g. the Ziegler-Nichols method (Ziegler and

Nichols, 1942)) might be helpful for moderating the

time complexity given a realistic modeling of the un-

derlying process (

Astr

om and H

agglund, 1984), (Bi

et al., 2000). But it doesn’t work in all applications

(Salsbury, 2005).

Because of the different inﬂuences on the over-

all quality of indoor environment, an HVAC control

system has to interrelate several input variables into

another set of outputs (Mirinejad et al., 2012). This

requires the system to treat numerous elements of un-

certainty. Historically, uncertainty has been dealt with

by probability theory. Though powerful, probability

theory, however, serves well in modeling situations

where the primary source of uncertainty is random-

ness (Jaynes, 2003). Other sources of uncertainty

such as vagueness, similarity, or preference as op-

posed to ambiguity can be dealt with adequately by

emulating human cognitive and decision-making pro-

cesses. The theory of ‘fuzzy logic’ provides a natu-

ral framework to handle uncertainty in a natural way.

Here, instead of asking whether something is true, we

ask how much it is true (Keller et al., 2016).

Research on HVAC control systems has correlated

the effectiveness of the PID controllers with the nat-

uralness of the fuzzy logic approaches. For exam-

ple, Mohindru and Sharma found that a fuzzy con-

troller with two input signals, one for error and an-

other denoting the rate of change along with seven

membership functions, performed better than a ﬁne-

tuned Ziegler-Nichols-style PID controller (Mohin-

dru and Sharma, 2015). This result was supported

by other studies that showed the fuzzy logic con-

trollers’ ability to adapt to longer time delays (Kobersi

et al., 2013), reduce energy consumption (Dash et al.,

2012), and/or provide thermal comfort (Collotta et al.,

2014).

In the present work, we set out solve the prob-

lem of achieving thermal comfort with possibly re-

duced energy consumption. To this end, we designed

a neuro-fuzzy controller that combines neural net-

works’ ability of generalization (Gl

uge et al., 2010)

and the human-like inference within of the fuzzy logic

framework (Keller et al., 2016). Combining informa-

tion from multiple sensory sources has been shown to

facilitate human decision making (Hamid et al., 2010)

and model’s performance (Hamid, 2015). Our con-

troller achieves a higher indoor environment quality

by balancing thermal comfort and energy consump-

tion. The underlying model of the proposed controller

utilizes a Sugeno-style fuzzy inference system with

two sensory inputs: one for temperature and another

for motion. It outputs a signal that represents the

mode of the air conditioner and the compressor speed

Figure 1: Room geometry.

for each mode. The testing of the controller showed

that the air conditioner of the controlled HVAC sys-

tem turns off automatically 10 minutes after the last

detected motion of room occupants. Accordingly,

simulations of the cost levels and energy consump-

tion were shown to fall when the room was empty as

justiﬁed through the absence of occupants’ motion.

The remainder of the paper is organised as fol-

lows. Section 2 introduces the designed neuro-fuzzy

controller and its underlying mathematical model for

regulating room temperature. In section 3, we discuss

the considered simulation scenarios. Results are pre-

sented and discussed in section 4. We ﬁnally conclude

and portray our plans for future work in section 5.

2 METHODS AND ANALYSIS

2.1 Mathematical Modeling of Indoor

Air Temperature

Controlling the temperature of an indoor environment

can be affected by various factors including the geom-

etry of the indoor environment as well as the environ-

mental conditions of the corresponding outdoor sur-

roundings. Figure 1 shows the geometric characteris-

tics of the room that served for the control task in our

work. The heating and cooling of the room tempera-

ture is affected by the efﬁciency of the air conditioner,

the volume of the room, the heat loss and heat gain in

winter and summer, respectively. The rate of temper-

ature change of the monitored room can be computed

by:

room

room−air

air



gain

−

loss



(1)

where dT

room

/dt denotes the rate of thermal variation,

which is transferred to the room from the air condi-

tioner, in terms of the discrepancy between the rate

of energy gain Q

gain

and that of energy loss Q

loss

(in

joules per seconds). According to Eq. (1), the amount

of transferred thermal energy will cool down or heat

up the room depending on the constant mass of the

air (in kilogram) of the monitored room (m

room−air

)

and speciﬁc heat capacity (c

air

) in (joule/kilogram

degree). The speciﬁc heat capacity of air (c

air

) in

(joule/kilogram degree) for both 20

◦

C and 25

◦

C as

initial room temperature is 1005 joule/kilogram de-

grees. Moreover, the mass of the room’s air m

room−air

depends on air density ρ (in kg/m

) and the volume V

(in m

) of the room. Hence, the mass of the room’s

air can be computed as follows.

room−air

= ρ.V (2)

We chose ρ = 1.225 kg/m

of the standard atmo-

sphere, which is the value of air density at sea level

and at 15

◦

C (McCormick, 1995), (Cavcar, 2000). The

volume of the room is 36 m

. Lastly, the energy loss,

that is dQ

loss

/dt, is computed from

loss

κA(T

room

− T

outside

)

(3)

Here, κ is the thermal conductivity (in Joule/sec/m

◦

of the insulation materials to conduct energy trans-

fer. It is 0.72 Joule/sec/m

◦

for common brick, 0.78

Joule/sec/m

◦

for glass window, and 0.8 Joule/sec/m

◦

for concrete roof. The term A in Eq. (3) refers to the

area. We considered three areas in our calculations,

the area of the window A

window

, that of the roof A

roo f

and the area of the wall A

wall

. Finally, D in Eq. (3)

represents the thickness of the common brick, dou-

ble glazed window, and concrete roof, which is in our

case 0.3m, 0.01m, and 0.25m, respectively.

Figure 2: Architecture of the neuro-fuzzy system.

2.2 Modeling the Air Conditioner

Selecting the right size of an air conditioner is crucial

for optimizing energy consumption. For example, a

bigger sized air conditioner may cool a room faster

than a smaller one, but it consumes more energy.

On the other hand, choosing the right size of an air

conditioner is bedeviled by some factors such as

the area and the volume of the room, the number of

occupants, the local climate, shading, and the size

of windows inside the room. All the same, there

is a set of rules that can be used to estimate the

proper capacity of an air conditioner. To regulate

the room temperature described in Section 2.1, we

considered a one ton air conditioner, which corre-

sponds to 3516 Joules/second of energy consumption

(Air Conditioning Systems, 2017). The compressor

speed for heating and cooling within the air condi-

tioner is not limited to certain values of electricity

usage. This implies that the working power of the

compressor varies according to the difference be-

tween the actual and the desired temperatures (Song

et al., 2015). This allows adjusting the compressor

to the desired capacity and operating conditions

(Engineering 360 Powered by IEEE GlobalSpec,

2017). Thus, the air conditioner will not work with

its maximum amount of energy when there is a

small amount of differences present between actual

temperature of the room and the desired temperature.

2.3 The Neuro-Fuzzy Controller

We modeled the air conditioning using a neuro-fuzzy

control system as developed by ANFIS toolbox in

Matlab. Figure 2 shows the general architecture of

the devised neuro-fuzzy. It consists of ﬁve layers with

following speciﬁcations. The ﬁrst layer corresponds

to input variables, the second layer represents input

membership functions (inputmf), the third layer refers

to the rule base, the fourth layer is denotes the out-

put membership functions, and the ﬁfth layer refers

to the output variable. Importantly, our model utilizes

a fuzzy inference system (FIS) with two sensory in-

puts: one for temperature and another for motion. It

then produces an output signal that controls the com-

pressor speed. Speciﬁcally, the ﬁrst input, termed as

‘input1’, corresponds to the error that results from the

discrepancy between the desired and the actual room

temperatures. The second input, referred to as ‘in-

put2’, models the room occupancy, that is, whether or

not the sensor detects motion in the room. The out-

put represents the mode of the air conditioner and the

compressor speed for each input combination. The

error input has three trapezoidal membership func-

Figure 3: Modeling the Neuro-Fuzzy Controller with Matlab ANFIS toolbox.

tions, the motion input has two trapezoidal member-

ship functions, and the output has six membership

functions. We tested the devised FIS with different in-

put combinations, ranging from -9.531

◦

C to 3.242

◦

for ‘input1’ and from 0 to 1 for ‘input2’ as illustrated

in the upper left and upper right plots of Fig. 3, re-

spectively). The range of output membership func-

tion lies within the interval [-65, 65] as shown in the

bottom left plot of Fig. 3. Finally, the rule base of a

Sugeno-style FIS is demonstrated in the bottom right

plot of Fig. 3.

2.4 Cost Calculation

Total energy cost was calculated as the product of the

amount of energy consumed and the cost for one en-

ergy unit. The amount of energy consumed in time

unit is computed from

= Q.t (4)

with Q representing the energy in Joule unit and t re-

ferring to time in seconds.

3 SIMULATION SCENARIOS

Once the devised neuro-fuzzy controller is setup, the

simulation for regulating room temperature can start.

We set our target temperatures at 22

◦

C for winter and

◦

C for summer. Our choice of these temperature

values was motivated by research results on ‘thermal

comfort’ (Caldo, 2015) and international standards

along with a sense of practical rationale. A place with

relatively high temperature will cause its occupants to

feel tired and exhausted, whereas they will become

restless and distracted, if the temperature is relatively

low. Consequently, the simulation runs with the tar-

get temperatures of 22

◦

C for winter and 24

◦

for sum-

mer. We used different initial room temperature at the

start of the simulation for summer and winter. Tem-

perature variation is generated from the differences

between the current and the desired room tempera-

tures. Once the rate of temperature differences are fed

into the controller, the system outputs the mode of the

air conditioning, which will be selected automatically,

and the amount of energy given to the air conditioner

Figure 4: Simulation results. The ﬁrst column corresponds to the presence of motion, whereas the second simulates the

absence of motion. In each column, the rows from up to down show the neuro-fuzzy controller’s output response and the

controller’s speed.

to control the compressor speed. The second input,

which represents the signal provided by the motion

sensor, detects whether there is moving occupants to

ensure that the air conditioner turns on only if there

is motion and off 10 minutes after the last observed

motion.

4 SIMULATION RESULTS OF

THE NEURO-FUZZY

CONTROLLER

4.1 In The Presence of Motion

Simulation results of the neuro-fuzzy controller in

the presence of motion are shown in the ﬁrst subplot

columns of ﬁgures 4 and 5. Subplot A of ﬁgure 4

represents the indoor temperature change according

to the outdoor temperature and the target temperature.

The starting indoor temperature is 20

◦

C for simulat-

ing the outdoor temperature in winter. The system

starts at 08:00 AM and the indoor temperature raised

fast by 70 seconds and was brought to around the set

point of 22

◦

C as it is shown in subplot C of ﬁgure 4.

This time for reaching the set point will be shorter if

the initial indoor temperature is higher or the target

temperature is lower. The lower target temperature is

easier to achieve since it is closer to the outdoor tem-

perature whenever the system is starting to work. This

means that the room temperature is well controlled

by the proposed neuro-fuzzy controller. Then the in-

door temperature remains relatively steady and keeps

around the set point. However, there is a tempera-

ture ﬂuctuation which is because of the outdoor tem-

perature reaches to the lowest temperature during this

time periods or evening time. As a result of that the

conduction of heating ﬂow from indoor to outdoor is

increasing but the controller has the capability of reg-

ulating room temperature. Therefore, there are some

variations due to outdoor temperature during 24 hours

or 86400 seconds. Although the proposed intelligent

temperature controller is able to rise the indoor tem-

perature back to the set point, there is a small ﬂuc-

tuation occurred which is variant between 21.98 and

21.99.

While providing a comfortable indoor tempera-

ture, it is important to consume less energy. The more

Figure 5: Simulation results. The ﬁrst column corresponds to the presence of motion, whereas the second simulates the

absence of motion. In each column, the rows from up to down show the amount of energy consumed by the controller and the

corresponding total cost.

energy consumed will result in increasing the emis-

sion of greenhouse gases like carbon dioxide. The

validation of the neuro-fuzzy controller according to

energy consumption is shown in subplot A of ﬁgure

5. The measurements indicate that the maximum of

energy is consumed when there is a huge difference

between indoor temperature which is affected by out-

door temperature and set temperature. On the other

hand, the energy consumed decreases when the sys-

tem is relatively steady, since the variation of temper-

ature has decreased. This is applied in the ﬁrst period

of the day which is from 08:00 AM to 08:00 PM or

every twelve hours of a day. Therefore, the variation

of the amount of energy consumed will results in the

variation of electric power consumed. The amount of

consumed electricity over a speciﬁc time period is the

cost of consumed energy. As we can see in subplot C

of ﬁgure 5, the cost of electric power is stable during

the period of (08:00 AM) to (08:00 PM) which indi-

cates the period of saved energy. On the other hand, in

the period between the (08:00 PM) to (08:00 AM) the

cost increases because of the increasing energy con-

sumption. This results in a total cost of energy con-

sumption of $3.412.

4.2 In the Absence of Motion

Analogously, the simulation results in the absence of

occupants’ motion are given in the second subplot

columns of ﬁgures 4 and 5. Speciﬁcally, the system

is tested for assuming that there is no motion in the

room in the period from 2938 seconds to 29978 sec-

onds while the set point temperature is 22

◦

C (subplot

B of ﬁgure 4). The HVAC system will turn off after

10 minutes or 600 seconds at point 3538 seconds of

no occupants in the room. On the other hand, as soon

as the room is occupied at 29978 seconds, the air con-

ditioning system will start to work and it controls the

indoor temperature to be back on track of rising up

to reach the set point. The measurements shown in

subplot D of ﬁgure 4 illustrate that the system will

cool down or reach around the outdoor temperature

of about 16

◦

C by 867 seconds. Nevertheless, there

are some small differences between the outdoor and

indoor temperatures because of the wall and window

insulations.

Importantly, one of the aims of our controller’s

design is to automatically turn off the air conditioner

when there is no occupants, so as to save energy and

decrease the cost of electricity used. The result of sav-

ing energy can be seen clearly in subplot B of ﬁgure

5. When there is no room occupancy, the amount of

energy consumed becomes zero. As a result,the total

cost of electricity decreases to $3 as shown in subplot

D of ﬁgure 5.

5 CONCLUSIONS

We have designed a neuro-fuzzy HVAC control sys-

tem for regulating room temperature. Our controller

achieves a higher indoor environment quality by bal-

ancing thermal comfort and energy consumption. The

underlying model of the proposed controller utilizes

a Sugeno-style fuzzy inference system with two sen-

sory inputs: one for temperature and another for mo-

tion. It outputs a signal that represents the mode of

the air conditioner and the compressor speed for each

mode. The testing of the controller showed that the air

conditioner of the controlled HVAC system turns off

automatically 10 minutes after the last detected mo-

tion of room occupants. Accordingly, simulations of

the cost levels and energy consumption were shown

when the room was empty as justiﬁed through the ab-

sence of occupants’ motion.

For the future, we aim at expanding the number

of environmental factors to be considered compared

with only one variable, that is, the room temperature

in the current analysis. Also, we shall deploy the con-

troller to monitor a whole house rather than merely

one room.

ACKNOWLEDGEMENTS

The authors would like to thank the anonymous re-

viewers for their valuable time and helpful comments.

REFERENCES

Afram, A. and Janabi-Shariﬁ, F. (2014). Theory and ap-

plications of hvac control systems–a review of model

predictive control (mpc). Building and Environment,

72:343–355.

Astr

om, K. J. and H

agglund, T. (1984). Automatic tuning

of simple regulators with speciﬁcations on phase and

amplitude margins. Automatica, 20(5):645–651.

Bi, Q., Cai, W.-J., Wang, Q.-G., Hang, C.-C., Lee, E.-

L., Sun, Y., Liu, K.-D., Zhang, Y., and Zou, B.

(2000). Advanced controller auto-tuning and its appli-

cation in hvac systems. Control Engineering Practice,

8(6):633–644.

Caldo, R. B. (2015). FLC-based indoor air quality assess-

ment for ashrae standard conformance. In Industrial

Engineering, Management Science and Applications

2015, pages 711–718. Springer.

Cavcar, M. (2000). The international standard atmosphere

(isa). Anadolu University, Turkey, 30:1–7.

Collotta, M., Messineo, A., Nicolosi, G., and Pau, G.

(2014). A dynamic fuzzy controller to meet thermal

comfort by using neural network forecasted parame-

ters as the input. Energies, 7(8):4727–4756.

Dash, S. K., Mohanty, G., and Mohanty, A. (2012). Intelli-

gent air conditioning system using fuzzy logic. In-

ternational Journal of Scientiﬁc & Engineering Re-

search, 3(12):1–6.

uge, S., Hamid, O. H., and Wendemuth, A. (2010). A

simple recurrent network for implicit learning of tem-

poral sequences. Cognitive Computation, 2(4):265–

271.

Hamid, O. H. (2015). A model-based markovian context-

dependent reinforcement learning approach for neu-

robiologically plausible transfer of experience. In-

ternational Journal of Hybrid Intelligent Systems,

12(2):119–129.

Hamid, O. H., Wendemuth, A., and Braun, J. (2010).

Temporal context and conditional associative learn-

ing. BMC neuroscience, 11(1):45.

Jaynes, E. T. (2003). Probability theory: The logic of sci-

ence. Cambridge university press.

Keller, J. M., Fogel, D. B., and Liu, D. (2016). Fundamen-

tals of computational intelligence. Wiley-IEEE Press.

Kobersi, I. S., Finaev, V. I., Almasani, S. A., and Abdo,

K. W. A. (2013). Control of the heating system

with fuzzy logic. World Applied Sciences Journal,

23(11):1441–1447.

Air Conditioning Systems (2017). Air Conditioning Calcu-

lations. (online accessed: 2017, January 27).

Canadian Centre for Occupational Health and Safety

(2017). Thermal Comfort for Ofﬁce Work. (2017,

February 17).

Engineering 360 Powered by IEEE GlobalSpec (2017).

HVAC - Ventilation. (online accessed: 2017, January

27).

McCormick, B. W. (1995). Aerodynamics, aeronautics, and

ﬂight mechanics, volume 2. Wiley New York.

Mirinejad, H., Welch, K. C., and Spicer, L. (2012). A re-

view of intelligent control techniques in hvac systems.

In Energytech, 2012 IEEE, pages 1–5. IEEE.

Mohindru, P. and Sharma, G. (2015). Simulation perfor-

mance of pid and fuzzy logic controller for higher or-

der system. Simulation, 1(7).

Nowak, M. and Urbaniak, A. (2016). Application of predic-

tive control algorithms for thermal comfort and energy

saving in the classroom. In Carpathian Control Con-

ference (ICCC), 2016 17th International, pages 527–

532. IEEE.

Salsbury, T. I. (2005). A survey of control technologies in

the building automation industry. IFAC Proceedings

Volumes, 38(1):90–100.

Song, Y., Wu, S., and Yan, Y. (2015). Control strategies

for indoor environment quality and energy efﬁciencya

review. International Journal of Low-Carbon Tech-

nologies, 10(3):305–312.

Wang, Y.-G., Shi, Z.-G., and Cai, W.-J. (2001). Pid auto-

tuner and its application in hvac systems. In American

Control Conference, 2001. Proceedings of the 2001,

volume 3, pages 2192–2196. IEEE.

Ziegler, J. G. and Nichols, N. B. (1942). Optimum settings

for automatic controllers. trans. ASME, 64(11).