Sie sind auf Seite 1von 9

BUILD SIMUL (2014) 7: 237–245

DOI 10.1007/s12273-013-0138-3

Modeling and optimization of HVAC systems using artificial neural


network and genetic algorithm

Research Article
Nabil Nassif ()

North Carolina A&T State University, Department of CAAE Engineering, 455 McNair Hall, 1601 E. Market Street, Greensboro, NC 27411, USA

Abstract Keywords
Intelligent energy management and control system (EMCS) in buildings offers an excellent means HVAC systems,
of reducing energy consumptions in HVAC systems while maintaining or improving indoor self tuning models,
environmental conditions. This can be achieved through the use of computational intelligence artificial neural network,
and optimization. The paper thus proposes and evaluates a model-based optimization process for energy management and control
HVAC systems using evolutionary algorithm for optimization and artificial neural networks for systems,
modeling. The process can be integrated into the EMCS to perform several intelligent functions optimization

and achieve optimal whole-system performance. The proposed models and the optimization
process are tested using data collected from an existing HVAC system. The testing results show
Article History
Received: 23 August 2012
that the models can capture very well the system performance, and the optimization process can
Revised: 19 April 2013
reduce cooling energy consumption by about 11% when compared to the traditional operating
Accepted: 3 May 2013
strategies applied.

© Tsinghua University Press and


Springer-Verlag Berlin Heidelberg
2013

1 Introduction genetic algorithm (GA) inspired by natural evolution which


is successfully applied to a wide range of applications
Great efforts have been invested in minimizing the energy including HVAC systems (Deb 2001; Goldberg 1989; Xu et
costs associated with the operation of HVAC systems. al. 2009; Mossolly et al. 2009). The HVAC optimization
Intelligent Energy Management and Control Systems (EMCS) problems are dynamic and the problem changes over the
can provide an effective way of decreasing energy costs in course of the optimization. Thus, the GA with proper
HVAC systems while maintaining indoor environmental enhancement and the ability of continuously track the

Building Systems and


conditions (ASHRAE 2011). This EMCS can include several movement of the optimum over time are developed and
Components
intelligent functions such as optimum set points and used. Components models are required for the optimization
operating modes (ASHRAE 2011; Nassif 2012; Nassif et al. process and for any other functions. Depending on the type
2005; Wang and Jin 2000; Zheng and Zaheer-Uddin 1996) of functions and the accuracy required, the models can vary
and fault detection and diagnosis (Seem 2007; Lee et al. from simple to more sophisticated calculations. However, it
1996). The intelligent EMCS can be achieved through the use is of practical importance to develop a simple, yet accurate
of the computational intelligence and optimization (Hagras and reliable model to better match the real behavior of the
2008; Kusiak and Xu 2012). The paper thus proposes model- subsystems and overall system over the entire operating
based optimization process for HVAC systems using genetic range. Models can be developed by two distinct methods
algorithm and artificial neural networks. The process can (ASHRAE 2009): forward models and data-driven models.
be integrated into the EMCS to perform such intelligent Forward models may need detailed physical information that
functions and achieve optimal whole-system performance. is not always available. Moreover, it is almost impossible to
The HVAC optimization problems are solved using the develop a model based on physical knowledge that perfectly

E-mail: nnassif@ncat.edu
238 Nassif / Building Simulation / Vol. 7, No. 3

simulates the real system behavior. In addition, the mainly of (i) the return and supply fans, (ii) the outdoor,
characteristics of the real system can change over time; the discharge, and recirculation dampers, (iii) the air handling
physical model with fixed parameters may no longer be unit (AHU) with components such as filter and cooling
able to track correctly the real performance. Alternatively, and heating coils, (iv) the pressure-independent VAV
data-driven models are another approach to be used for terminal boxes, and (v) the local-loop controllers (i.e., C1,
existing systems. In this paper, we will explore the use of C2, and C3). The supply air temperature is controlled by the
artificial neural network as data-driven and self tuning controller (C1). The duct static pressure is controlled by the
models in HVAC applications, focusing on optimization controller (C2). The zone air temperature at any particular
and control. The proposed models are tested evaluated using zone n is controlled by the controller (C3 (n)).
data collected over three summer months from a typical The EMCS collects the measured data (real data) from
existing HVAC system. However, the optimization process components or subsystems. The ANN models are con-
was evaluated by simulations applied to the same system. tinuously trained using the real data to better match the real
behavior of the subsystems and overall system. The training
2 HVAC and EMC system configuration algorithm used is Levenberg-Marquardt, which is a built-in
algorithm in MATLAB. As recommended in this paper for
Figure 1 shows a schematic diagram of a typical HVAC optimal control strategy, at each time interval (e.g. 10 min),
system and the optimization process integrated into EMCS. the ANN models provide optimal whole system performance
Included in this schematic are the proposed artificial neural through determining optimal set points and operation
network (ANN) models and how they interact within the sequences. A genetic algorithm as described in next section
overall system. A typical HVAC system that uses variable is used to solve the optimization problem.
air volume (VAV) control is illustrated in Fig. 1. It consists

Fig. 1 Typical HVAC system and EMCS along with the proposed ANN models and optimization process
Nassif / Building Simulation / Vol. 7, No. 3 239

3 Genetic algorithm search space. This algorithm uses the elite-preserving


operator, which favors elites of a population by giving them
Genetic algorithm GA is used to solve the optimization an opportunity to be directly carried over to the next
problem. GA is inspired by natural evolution which is generation. After two offspring are created using the
successfully applied to a wide range of applications including crossover and mutation operators, they are compared with
HVAC systems. The problem variables are the controller both of their parents to select two best solutions among the
setpoints and operation modes. In this paper, the supply air four parent-offspring solutions. The stochastic universal
temperature, duct static pressure, and outdoor airflow rate sampling—SUS version of proportionate selection is used
are the only problem variables considered. The objective (Deb 2001). A penalty is imposed on the objective function
function is the total energy consumption due to the operation based on the constraints. This is done by adding on the
of fans, pumps, chillers and associated equipment, and objective function (energy use) used by GA an amount of
electric reheat if used. Figure 2 shows the flowchart of 10% (adjustable) of that energy use. Due to elevated energy
genetic algorithm to solve optimization problem. The GA use, this solution has less chance to carry over to next
starts with a random generation of the initial population generation. The constraints are related to the restrictions
(initial solution). The problem variables (controller setpoints) on the operation of the HVAC system and the rate of
are encoded to form a chromosome (a string of variables) variable changes. They cover the lower and upper limits of
that represents an individual solution in the population. variables, the design capacity of components. The fan and
The performance or objective function f of each individual of zone airflow rates, for instance, are restricted within the
the first generation is estimated. The second generation is maximum and minimum limits. The minimum zone
generated using operations on individuals such as selection, airflow rates are restricted to be always higher than 20% of
crossover, and mutation, in which individuals with higher design airflow rates.
performance (fitness) have a greater chance to survive (Deb The maximum available zone airflow rate Vzmax (when the
2001). The performance of each new individual is again VAV damper is fully opened) varies with duct static pressure
evaluated. The process is repeated until the maximum setpoint (Ps) and can simplify by the following:
number of generations (Gmax) is reached. The simulated
binary crossover SBX is used to create two offsprings from Ps
Vz max = Vz design ´ (1)
two-parent solutions. The random simplest mutation Ps design
operator is applied to create a new solution from the entire
When the duct static pressure Ps is equal to the design
value Psdesign, the maximum available zone airflow rate
Vzmax is equal to the design value Vzdesign. To ensure that
every individual zone at any duct static pressure setpoint
(Ps) receives adequate supply air, the “zone airflow rate
constraint” must be respected so that the estimated zone
airflow rates are restricted to be less than the maximum
available zone airflow rate determined by Eq. (1).
The supply air temperature setpoint is limited within
13℃ and 18℃. The duct static pressure setpoint is limited
within 150 Pa and 250 Pa. For the results of this study, the
computations with the optimization process are performed
well within a time length of three minutes on a current
standard computer with CPU of 2.6 GHz. This computation
time allows the optimization process to be implemented
online. The time could be also decreased using a more
modern computer.

4 HVAC component models

The HVAC component models based on artificial neural


networks with “self-learning” capability are developed and
Fig. 2 Flow chart of genetic algorithm GA for the optimization utilized in the EMCS to perform the advanced and intelligent
process functions. An infinite variety of network architectures can
240 Nassif / Building Simulation / Vol. 7, No. 3

be used for this purpose, but in the interest of conserving setpoint), and the output is cooling load. The inputs of the
computer time the simplest structure needs to be considered. chiller model are the cooling coil load, chilled water supply
Each model has only one hidden layer with twenty neurons temperature, and the condenser water temperature, and the
and with hyperbolic activation function (tanch). The model output is the compressor power. Additional basic calculations
has the output layer with one neuron and an activation are also required for the optimization, including a zone
function based on the sum of the weighted hidden layer model. The zone model is to determine the zone airflow
neurons. Each neuron also has bias. rates, local heating energy use, and return air conditions
The models described here are the fan, cooling coil, based on thermal loads. The calculations are based on the
and chiller models. Those models can be used for various steady state heat balance equation for each zone in which
applications but the inputs and the outputs have to be clearly the sensible load is a function of airflow rate and the
defined. The inputs of the models for a specific application differences between the space and supply air temperatures.
may become outputs in other application. For instance, in Similarly, the humidity is determined using the latent load.
the control application, the controlled variables such as duct The loads are determined from the same model but with an
static pressure and supply air temperature are the outputs inverse form using measured data of the previous period
of the fan model and cooling coil model, respectively. For and then the loads are assumed to be constant during the
the optimization, the model outputs are estimates of the current optimization period. The electric reheat is considered
energy consumptions (the objective function) such as fan here and it turns on only when the airflow rate reaches its
power and compressor power (see the second and third rows minimum level (e.g., 20% of design airflow rate) and the
in Table 1). Although the models were validated and tested space temperature becomes lower than heating setpoint.
for both cases, the detailed discussions were only made for The system airflow rate used as an input for cooling or fan
the optimization application as our main focus in this paper model is equal to the sum of zone airflow rates found from
is to test the proposed optimization process. Such process the zone model. An iteration process should be applied to
needs the models as shown in Fig. 3. The inputs of the fan estimate the return air conditions, the initial cooling coil
model are the system airflow rate and static pressure (duct leaving air humidity ratio is assumed, and the new value
static pressure setpoint) and the output is fan power. The is calculated and reused. This iterative process continues
inputs of the cooling coil model are fan airflow rate, entering calculating through the loop several times until the values
liquid temperature (chilled water supply temperature set- of cooling coil leaving air humidity ratio stabilize within
point), entering air dry bulb temperature and humidity ratio, a specified tolerance. The calculation presented above
supply air dry bulb temperature (supply air temperature is similar to that for a VAV model in the HVAC toolkit

Table 1 Model testing results


Results
Models Application Model output Model test Samples MaxE MAE CV(%)
Training 34560 × 3 0.475 0.19 1.13
Optimization Power (kW) Validation 6912 × 3 0.487 0.20 1.2
Testing 10080 × 3 0.950 0.227 2.8
Fan model
Training 34560 × 3 4.55 2.14 1.61
Control Static pressure (Pa) Validation 6912 × 3 5.08 2.22 2.12
Testing 10080 × 3 6.29 3.07 3.1
Training 34560 × 3 8.56 3.12 4.10
Optimization Cooling load (kW) Validation 6912 × 3 8.72 3.20 4.23
Testing 10080 × 3 11.10 5.12 7.8
Cooling coil model
Training 34560 × 3 0.57 0.32 2.11
Control Supply air temperature (℃) Validation 6912 × 3 0.72 0.41 3.66
Testing 10080 × 3 0.87 0.52 5.7
Training 80 0.81 0.41 2.71
Optimization Power (kW) Validation 20 0.84 0.42 2.12
Chiller model
Testing 18 1.23 0.58 3.18
Control Chilled water temperature (℃) Not tested
Nassif / Building Simulation / Vol. 7, No. 3 241

Fig. 3 Flow chart of component models

(Brandemuehl et al. 1993). The minimum airflow rate 5 Modeling testing


based on the ASHRAE standard 62.1 2010 (ASHRAE 2010)
is included in the optimization calculations. The outdoor Data from an existing VAV system are collected over three
air is determined by the multi-zone procedure of ASHRAE summer months (June, July, and August) at one-minute
62.1 standard based on the actual zone airflow rates. The intervals. The first 23 days data of each month (34560 data
advantage of including the minimum outdoor standard points) were randomly divided into two types of samples:
procedure in the whole optimization process is to minimize 80% for training (27648 data points) and 20% for validation
the energy use while respecting the ventilation requirements by (6912 data points). The last 7 days of each month (10080
the current standard. The standard prescribes two ventilation data points) are kept for the model testing and to define the
rates, one intended to dilute the contaminants generated by model accuracy. During the training, the network is adjusted
occupants (Rp) and the other for building-related sources according to its error using Levenberg-Marquardt optimiza-
(Ra). The required minimum breathing zone outdoor air rate tion. The validation samples are used to measure network
is as a function of the number of zone occupants Pz and the generalization, and to halt training when generalization
zone floor area Az. When the economizer is not activated, stops improving. The testing data have no effect on training
the minimum outdoor airflow rate Vot is found: and provide an independent measure of network performance
after training. The model performance is measured by the
Vou
Vot = (2) mean absolute error MAE, maximum error MaxE, and the
Ev
coefficient of variation CV, which is defined as the ratio of
The uncorrected outdoor air intake flow Vou and system the standard deviation to the mean.
ventilation efficiency Ev are given: Figure 4 shows the results of the ANN fan model
training, validation, and testing. The straight line is a one
to one line, indicating agreement between the measured
Vou = å (Rpi ´ Pzi + Rai ´ Azi ) (3)
(target) and simulated (output) fan power. The design
fan capacity and power input are 23000 L/s and 60 kW,
Vou Rpi ´ Pzi + Rai ´ Azi respectively. The model slightly overestimates the power at
Ev = min 1 + ( Vs
-
Vzi ) (4)
elevated values during the training period and underestimates
the power during the testing period. The CV, MAE, and
The zone airflow rates Vzi and system flow rate Vs (fan airflow MaxE during the testing period are 2.8%, 0.227 kW, and
rate) in Eq. (4) are optimally found by the optimization 0.950 kW, respectively, comparing to 1.13%, 0.19 kW, and
process. The term inside the parenthesis is calculated for 0.475 kW during the training period. The results in term of
each zone i and the Ev is equal to the minimum value. CV, MAE, and MaxE are shown in Table 1. The results are
242 Nassif / Building Simulation / Vol. 7, No. 3

Fig. 4 Testing of the ANN fan model

averaged for a period of three months (one week for each possible other applications when the supply air temperature
month). The inputs of the ANN fan model are the measured is the output as shown in Table 1. The model is able to
total pressure difference across the fan and airflow rate. In estimate accurately the supply air temperature with the
the optimization problem, the energy use by the fan needs to CV of 5.7%.
correlate with the duct static pressure setpoint not with the The chilled water and condenser water supply tem-
total pressure difference. Thus, the pressure drops between peratures collected from the existing system are always
the duct static sensor and fan outlet and between the fixed, and testing of the chiller model is not valid over a
outdoor air damper and fan inlet need to be considered wider range of operation. Thus, the ANN chiller model is
and can be simplified as proportional to the square of the evaluated against EnergyPlus’s electric chiller model based
flow rate. The model is also tested for the other possible on condenser leaving temperature, developed by Hydeman
applications when the duct static pressure is the control et al. (Hydeman et al. 2002). The chiller performance curves
output as shown in Table 1. The model is able to estimate are generated by fitting manufacturer’s catalog data. The
accurately the duct static pressure with the CV of 3.1%. cooling coil load, chilled water temperature, and condenser
In the cooling coil model, the required output for the water temperature are then the inputs for the chiller model.
optimization purpose is cooling coil load that is in turn the The output is the compressor power. The design compressor
input to the chiller model. The measured cooling coil loads power is 150 kW. Figure 6 shows the testing results of the
required for training and testing are calculated from the chiller model. The chiller energy use as a function of part
measured airflow rate and difference between the measured load ratio PLRr (cooling load to rated one) is illustrated at
inlet and outlet enthalpies. Figure 5 shows the results of the two different values of chilled water supply temperature (6℃
ANN cooling coil model training, validation, and testing. By and 10℃) (42.8℉, 50℉) and an entering condenser water
comparing between the measured (target) and simulated temperature of 33℃ (91.4℉). Under these two conditions
(output) cooling load for a testing period of three weeks and with an interval of 10% PLRr (18 operating conditions),
(10080×3), the coefficient of variance CV, MAE, MaxE are the accuracy of the model in terms of the coefficient of
7.8%, 5.12 kW, and 11.10 kW, respectively. The model variation CV, MAE, and MaxE are 3.18%, 0.58 kW, 1.23 kW,
slightly overestimates the outputs at elevated values and respectively. These testing results show that the models
underestimates at relatively lower values during the training capture very well the system performance and can be used
period. However, the outputs are somewhat scattered for the calculations required for the optimization process or
during the training period. The model is also tested for the any other applications.
Nassif / Building Simulation / Vol. 7, No. 3 243

Fig. 5 Testing of the ANN cooling coil model

operating modes. The inputs are the controller setpoints


(problem variables) and the output is the energy use (objective
function). As shown in Fig. 2, the genetic algorithm GA
sends a set of individual solutions containing trial controller
setpoints, and the models then estimate the objective function
(total energy use) and send it back to the GA to eliminate,
evolve, and pass this solution to the next generation. This
process continues until optimal/or near optimal solutions
are reached. The supply air temperature, duct static pressure,
and outdoor airflow rate are only considered, and the
optimization is done only for the summer period. Figures 7–9
show optimal and non-optimal (from exiting system) supply
Fig. 6 Testing of the ANN chiller model air temperature, duct static pressure, and outdoor airflow
rate for four days. The non-optimal setpoints are collected
6 System optimization from the actual operation of the existing system. In the
existing system, although the supply air temperature setpoint
The optimization process including the static ANN models is automatically reset based on the supply airflow rate
as shown in Fig. 1 is evaluated using data from an existing and outdoor air dry bulb temperature, the actual supply
VAV system serving class rooms and offices with a total of temperature setpoint seems to be always constant at 13℃
61 zones. Detailed information on the system can be found overridden by the operator. In addition, no control strategy
in the reference (Nassif 2005). The optimization process is applied to reset the duct static pressure setpoint and it is
predicts the system performance over a period of 10 min always constant at 250 Pa. As shown in Figs. 7 and 8, the
(optimization period). During this short optimization period, optimal values of the supply temperature and static pressure
the loads and outdoor air conditions are assumed to be vary with the operating conditions. The system is normally
constant and estimated from the measured data collected operating from 6 AM to 10 PM on weekdays and from 9 AM
during the previous period. The ANN models are used to to 8 PM on weekends. The optimal duct static can be as low
find the energy use by each component and then the total as 120 Pa (a lower limit is 100 Pa) compared to 250 Pa. A
energy use in response to the controller setpoints and significant energy saving in fan can be achieved due to
244 Nassif / Building Simulation / Vol. 7, No. 3

the operation at low duct static pressure setpoint. The


optimization process runs with three constraints related to
the duct static pressure: (1) maximum duct static pressure
is 250 Pa based on the design condition, (2) minimum
duct static pressure is 100 Pa based on fan performance
specifications to avoid the instability region, and (3) zone
airflow rate is restricted to be less than the maximum
available zone airflow rate determined by Eq. (1). The
optimal supply air temperature is slightly higher than 13℃.
Elevated temperature will increase the airflow rate and then
the fan power but it will improve slightly the ventilation
efficiency Ev. In addition, elevated supply air temperature Fig. 9 Optimal and non-optimal outdoor airflow rates
increases the chilled water return temperature and
consequently better chiller efficiency. The whole system
optimization finds the solution that produces the least total
energy use. As shown in Fig. 9, the optimal outdoor airflow
rate is lower than the actual outdoor airflow rate, which is
based on constant minimum damper position (approximately
constant fraction percentage) to provide 7.5 L/s per person
on design conditions. The optimal outdoor airflow rate is
found based on the multi-zone ventilation procedure of
ASHRAE Standard 62.1 2010 (Eqs. (2)–(4)) and using the
design occupancy but the optimal zone airflow rates.
Figure 10 shows the optimal and non-optimal total rate
of energy use for four days. The total energy use is equal to Fig. 10 Optimal and non-optimal energy uses
the sum of fan, electric reheat, and chiller powers. In the
existing system, several air handling units receive chilled
water from one common chiller equipped with a constant-
speed chilled water pump. As only one air handling unit is
considered in this study, one dedicated chiller is assumed
and the power is simulated based on EnergyPlus’s electric
chiller model (Hydeman et al. 2002) with the rated cooling
capacity according to the investigated AHU capacity. As a
result, by applying the optimization process, the total cooling
energy saving is 12.5% for those four days and 11% for the
three summer months (June, July, and August). The savings
could vary depending on the system types, building type and
locations, existing energy efficiency opportunity and current
Fig. 7 Optimal and non-optimal supply air temperatures control strategies.

7 Summary and conclusion

Artificial intelligence approaches are proposed for use in


HVAC control and to advance the EMCS. Self-tuning HVAC
component models based on an artificial neural network
were developed and validated against data collected from an
existing HVAC system. The testing results showed that the
models exhibit good accuracy and fit well the input–output
data. An infinite variety of network architectures can be
used for this purpose, but in the interest of conserving
computer time the simplest structure with one hidden layer
Fig. 8 Optimal and non-optimal supply duct static pressures
and twenty neurons is used. The errors of the fan, cooling coil,
Nassif / Building Simulation / Vol. 7, No. 3 245

and chiller models in terms of the coefficient of variation Hydeman M, Webb N, Sreedharan P, Blanc S (2002). Development
(CV) are within 2%–8%. The models could be incorporated and testing of a reformulated regression-based electric chiller
into the EMCS to perform several intelligent functions model. ASHRAE Transactions, 108(2): 1118–1127.
including energy management and optimal control. To that Kusiak A, Xu G (2012). Modeling and optimization of HVAC systems
end, a whole system optimization process based on a genetic using a dynamic neural network. Energy, 42: 241–250.
Lee WY, House JM, Park C, Kelly GE (1996). Fault diagnosis of an
algorithm was developed and tested. The testing results
air handling unit using artificial neural networks. ASHRAE
indicated that the optimization process can provide a cooling
Transactions, 102(1): 540–549.
energy saving of 11%. The saving could vary depending on
Mossolly M, Ghali K, Ghaddar N (2009). Optimal control strategy for
the system types, building type and locations, existing energy a multi-zone air conditioning system using a genetic algorithm.
efficiency opportunity and current control strategies. Energy, 34: 58–66.
Nassif N (2005). Optimization of HVAC control system using two-
References objective genetic algorithm. PhD Thesis, Ecole de technologie
superieur, Montreal, Canada.
ASHRAE (2009). ASHRAE Handbook—Fundamentals, Chapter 19. Nassif N (2012). Modeling and optimization of hvac systems using
Atlanta, GA, USA: American Society of Heating, Refrigerating artificial intelligence approaches. Paper presented in ASHRAE
and Air-Conditioning Engineers. Annual Conference, San Antonio, TX, USA.
ASHRAE (2010). ASHRAE Standard 62-2010, Ventilation for Acceptable Nassif N, Kajl S, Sabourin R (2005). Optimization of HVAC control
Indoor Air Quality. Atlanta, GA, USA: American Society of system strategy using two-objective genetic algorithm. HVAC&R
Heating, Refrigerating and Air-Conditioning Engineers. Research, 11: 459–486.
ASHRAE (2011). ASHRAE Handbook—HVAC Applications, Chapter Seem JE (2007). Using intelligent data analysis to detect abnormal
42. Atlanta, GA, USA: American Society of Heating, Refrigerating energy consumption in buildings. Energy and Buildings, 39: 52–58.
and Air-Conditioning Engineers. Wang S, Jin S (2000). Model-based optimal control of VAV
Brandemuehl MJ, Gabel S, Andersen I (1993). A Toolkit for Secondary air-conditioning system using genetic algorithm. Building and
HVAC System Energy Calculation. Atlanta, GA, USA: American Environment, 35: 471–487.
Society of Heating, Refrigerating and Air-Conditioning Engineers. Xu XH, Wang SW, Sun ZW, Xiao F (2009). A model-based optimal
Deb K (2001). Multi-Objective Optimization Using Evolutionary ventilation control strategy of multi-zone VAV air-conditioning
Algorithms. New York: John Wiley & Sons. systems using genetic algorithm. Applied Thermal Engineering,
Goldberg DE (1989). Genetic Algorithms in Search, Optimization, 29: 91–104.
and Machine Learning. Reading, MA, USA: Addison-Wesley. Zheng GR, Zaheer-Uddin M (1996). Optimization of thermal processes
Hagras H (2008). Employing computational intelligence to generate in a variable air volume HVAC system. Energy, 21: 407–420.
more intelligent and energy efficient living spaces. International
Journal of Automation and Computing, 5: 1–9.