h
Initial guess of specific enthalpy at compressor outlet
3
h
Initial guess of specific enthalpy at condenser inlet
h
a
, h
b
Inlet and outlet specific enthalpies
r
h Mean specific enthalpy in each refrigerant cell
r,in
h
Inlet specific enthalpy in each refrigerant cell
r,out
h
Outlet specific enthalpy in each refrigerant cell
xxiv
h
v,in
Inlet specific enthalpy of the expansion valve
h
v,out
Outlet specific enthalpy of the expansion valve
h
ideal
Ideal energy transferred to the refrigerant flow without any losses
J Momentum of inertia of the compressor
k Thermal conductivity of the fluid
k
d
Empirical constant based on the type of damper blades
f
k Friction coefficient in centrifugal compressor
initial
k Gain for the use of pseudo physical analogy
r
k Thermal conductivity of refrigerant
w
k Thermal conductivity of water
l
k Thermal conductivity of saturated refrigerant liquid
IGV
k Normalized IGV position
spring
k Spring constant
L
cond
Tube length of the condenser
L Length of the pipe or tube
L
d
Leakage when the damper is fully closed
L
tot
Total tube length
m
wall
Wall mass
M
z
Mass of air in the air volume in the conditioned zone
a
M
,
b
M
Mass flow rate of the total water vapor from the people in the conditioned zone
cond
M
Refrigerant mass at condenser
cond
M A Change of refrigerant mass at condenser during the time interval [t
0
, t
2
]
evap
M A
Change of refrigerant mass at evaporator during the time interval [t
0
, t
2
]
r
M Mass in each refrigerant cell
xxv
wall
M Mass in each wall cell
1,nom
m Nominal compressor mass flow rate from design specification
2
m Outlet mass flow rate from the compressor based on Figure 5.7
3
m
Inlet mass flow rate to the expansion valve based on Figure 5.7
4
m Outlet mass flow rate from the expansion valve based on Figure 5.7
analogy
m
Compressor mass flow rate calculated from the pseudophysical analogy
analytical
m Compressor mass flow rate calculated from the nonsurge analytical solution
max
m
Maximum compressor mass flow rate
mean
m
Mean mass flow rate of each refrigerant cell
m Compressor mass flow rate
r,in
m Inlet mass flow rate of each refrigerant cell
r,out
m Outlet mass flow rate of each refrigerant cell
v,in
m Inlet refrigerant mass flow rate of expansion valve
v,out
m Outlet refrigerant mass flow rate of expansion valve
smooth
m
Valve mass flow rate for the smoothing function used in the orifice plate
N Total number of tubes
Nu
w
Water side Nusselt number
Nu
m
Mean Nusselt number
p Pressure
1
p
Initial guesses of the inlet pressure to the compressor based on Figure 5.7
2
p
Initial guesses of the outlet pressure to the compressor based on Figure 5.7
3
p
Initial guesses of inlet pressure to the expansion valve based on Figure 5.7
4
p
Initial guesses of outlet pressure to the expansion valve based on Figure 5.7
/ dp dt Time derivative of pressure
xxvi
in
p Inlet pressure
out
p Outlet pressure
IGV
p A Pressure loss across IGV
c,in
p
Compressor suction line pressure
c,out
p
Compressor discharge line pressure
e
p
Evaporator side pressure
bulb
p
TXV bulb pressure
min
dp Empirical parameter of minimum opening pressure of TXV
Pr
r
Refrigerant side Prandtl number
Pr
w
Water side Prandtl number
human
Q
Total evaporative power from the people in the conditioned zone
r
Q
1b
Fluid velocity angle
2
Fluid velocity angle
2b
Backsweep angle at the impeller tip
p
o
Uncertainty of the pressure measurement
/ dp dt
o Uncertainty of the dp/dt calculation
Expansion factor
is
q Isentropic efficiency of the centrifugal compressor
p
q Polytropic efficiency
k
Heat capacity ratio
Friction coefficient
Dynamic viscosity
l
Dynamic viscosity of the saturated refrigerant liquid
s
Dynamic viscosity evaluated at the average value of the mean temperature
Kinetic viscosity
Density
3
Inlet refrigerant density to expansion valve based on Figure 5.7
xxix
c,in
Inlet refrigerant density of the compressor
c,out
Outlet refrigerant density of the compressor
cond
Mean refrigerant density of the condenser based on Figure 5.7
evap
Pressure rise ratio
Rotation speed of the rotor
xxx
ACKNOWLEDGEMENTS
This dissertation was facilitated by tremendous effort and help from other people and
resources. I would like to express my deep and sincere gratitude to my Ph.D. advisor Dr.
Yaoyu Li for his great help and effort. It was my pleasure and honor to be his first Ph.D.
student working on building HVAC research. I truly appreciate all his support and
patience to teach and guide me through the turbulence of research and study smoothly. I
am also sincerely indebted to Dr. John E. Seem, from whom I am fully convinced of the
importance to always master the fundamentals first. His serious attitude towards research
is always inspiring me to never overlook details and never give up. The way he presents
research outcomes, is a persistent reminder for me to always pursue high quality.
I gratefully acknowledge the continuous funding support from Johnson Controls, Inc.
during my Ph.D. study, from which I have all the resources to concentrate on my research
and have the opportunity to attend many conferences, travel to Germany and visit York
China for meaningful collaborations. I am also grateful for the experimental test
conducted at YorkChina, Wuxi to validate the key modeling assumption in my research.
I am thankful for the rest of my committee members: Dr. TienChien Jen, Dr. Ronald
A. Perez and Dr. Dexuan Xie, for their precious time and support on my Ph.D. work. I
am also thankful for the model library developers in TLK/ifT. It was my great experience
and joyful trip to attend the workshop and stay there for a month as a visiting researcher.
Finally, my deepest gratitude goes to my mom, dad, and my wife, for all your endless
support. It is your unconditional love that keeps me moving forward.
1
Chapter 1 Introduction
Buildings are a major sector for energy consumption. As of 2010, commercial and
residential buildings consume nearly 40% of the primary energy supply in the United
States, even more than the industry and transportation sectors [1] (see Figure 1.1).
Considerable amount of carbon emissions from buildings have a large impact on man
made global warming effect. In addition to economical savings, improving energy
efficiency in buildings can also help to reduce carbon emissions. Building energy
efficiency can be improved in a variety of aspects. In commercial buildings, the major
sources of energy consumption in buildings are identified as lighting (17.4%), space
heating and cooling (23.8%), ventilation (8.7%), and refrigeration (6.7%) [1]. Thus,
improving the energy efficiency of heating, cooling, ventilation and airconditioning
(HVAC) systems would have a significant impact on the overall building efficiency.
Figure 1.1: U.S. primary energy consumption [1]
2
The performance and energy efficiency of building HVAC system may be improved
with better control strategies, better control setpoints, and better controller tuning. To
serve these needs, academia and industry have invested tremendous efforts for
developing various modeling and simulation techniques to advance controller design and
validation before actual implementation for field tests. This dissertation is concerned with
both dynamic modeling of building HVAC systems and the development of self
optimizing control strategies for efficient operation of such systems.
1.1 Modeling and Simulation of Building HVAC Systems
Building HVAC simulation programs can be roughly classified as wholebuilding
energy performance simulation and equipment level
1
simulation. A major difference
between these two simulation tasks lies in their respective purposes and the dominant
time scales in the simulation. The time scale for the former usually spans from hours to
days, while that for the latter usually ranges from several minutes to several seconds, or
even down to milliseconds (e.g., compressor failure protection mechanism usually
responses in milliseconds).
In general, the wholebuilding level simulation aims at performance assessment,
building commissioning and diagnostics [24]. For potential savings of applying whole
building simulation tools in building diagnostics, Roth [4] said: The energysaving
potential of retrocommissioning represents the upper boundary on automated whole
building diagnostics (AWBD) energy savings. Thus, AWBD tools could reduce building
energy consumption by 5% to 20%.
1
The equipment level simulation defined here includes component level simulation such as a single heat exchanger, and system
level simulations such chillers, chiller + tower, chiller + air handling units, etc.
3
In control applications, the equipment level simulation focuses on studying the
transients of a single HVAC component or subsystems coupled with local controllers.
Optimal controllers designed in this stage target at minimizing energy losses due to
improper operations on a cyclebycycle
2
base, thus the time scale considered would be
much less than the wholebuilding simulation programs.
For control and diagnostics study, there is a strong need to integrate dynamic
equipment level simulation with the dynamic wholebuilding simulation to solve
challenging problems for building HVAC systems. Currently, such combination
3
can be
realized by running cosimulation with different simulation platforms. The simulation
data is synchronized between different platforms based on a predefined sample rate for
data synchronization. An effort is also being made to unify the capabilities of simulating
both equipment level and wholebuilding level in a same modeling and simulation
platform [5].
Another important area of study is automated fault detection, diagnostics, and
prognostics of building HVAC systems. According to Katipamula and Brambley [6],
Poorly maintained, degraded, and improperly controlled equipment wastes an estimated
15% to 30% of energy used in commercial buildings. Fault detection and diagnostics
schemes can be applied to both wholebuilding level and equipmentlevel simulations to
monitor and correct the performance of building HVAC systems and their controls.
One major focus of this dissertation study is to address the problems in equipment
level modeling and simulation. In the past several decades, there has been tremendous
2
Cycle refers to thermodynamic cycle of a given refrigeration system to start to from its initial state to its final state
3
The purpose for cosimulation is to take the advantage of domainoriented language and put together their respective capability,
e.g., Modelica + EnergyPlus is suitable for dynamic equipment and detailed building envelop simulation, Modelica + Matlab/Simulink
is suitable for dynamic simulation of HVAC system with advanced controllers.
4
progress achieved for modeling and simulation of building HVAC systems. However,
there remain research needs and challenges in the dynamic modeling and advanced
control of building HVAC systems.
There are three modeling approaches that are often employed in control studies: 1)
detailed physicsbased (or firstprinciple) modeling, 2) reducedorder modeling, and 3)
datadriven modeling.
For HVAC controls, both detailed physicsbased modeling and reducedorder
modeling are needed. Physics based model reduction has been described in Tummescheit
[7]. For control applications, the reducedorder models are generally developed based on
the firstprinciples, through linearization to certain operating conditions, sometimes
followed by some model reduction schemes. The states of the obtained models can be
minimized by retaining the key dynamics. Such modeling techniques are sometimes
called controloriented modeling.
Datadriven modeling often results in linear or nonlinear blackbox models. Pairs of
input and output data are used to determine the parameters in predefined mathematical
model structures used to represent the actual HVAC models. Due to the nonlinearities
typically presented in HVAC systems, certain techniques such as neural networks,
support vector machine, and boosting tree are often applied to determine the inputoutput
mappings [8].
A major objective of this dissertation study is the development of detailed physics
based models. While the reducedorder models are generally needed for modelbased
controller design. The physicsbased detailed modeling can provide highfidelity virtual
5
plant to facilitate the development of control and fault detection techniques,
performance assessment, and design optimization.
1.2 Building HVAC Controls
Another theme of this dissertation study is to develop selfoptimizing controllers for
building HVAC systems. For HVAC systems, one typical control objective is setpoint
regulation, e.g. temperature and humidity control. Meanwhile, optimal control or setpoint
optimization is needed for energy saving operation, i.e. minimizing the energy
consumption while satisfying the demand for setpoint regulation. With the decades of
development of feedback and adaptive control strategies, the regulation control has
reached maturity for most occasions.
The setpoint optimization or optimal control HVAC systems, as comparison, is still
an open challenge in general. The building HVAC systems present a dilemma for control
engineers. On one hand, such systems feature complex, nonlinear and timevarying
processes; on the other hand, building operation has been a lowinvestment practice due
to economical reasons. Accurate modeling or detailed/frequent calibration is a luxury in
general. Such reality lends tremendous difficulty for applying the modelbased control
and optimization techniques and many practices that have been successful for other
industrial sectors. Therefore, selfoptimizing control strategies, which may achieve
optimal operation (e.g. in terms of efficiency) with least model information, are desirable
from the standpoint of building HVAC practice. This dissertation study, in addition to
dynamic modeling, also investigates the design of selfoptimizing control strategies for
building HVAC systems using the extremum seeking control.
6
These two research themes may present an apparent paradox: we try to build high
fidelity dynamic HVAC models, while the pursuit of selfoptimizing control strategy
implies the least dependency on the model. In fact, such paradox reflects the two parallel
and somewhat related needs from the practice of HVAC industry. Accurate simulation
models are beneficial for technology development, while operations have to adapt to high
uncertainties.
Among the large set of building HVAC equipment, the chilledwater system is
chosen as the target application. Such system is an important class of equipment for
commercial building ventilation and air conditioning systems [9]. The physics and system
configuration of such system are in many ways similar to many other HVAC systems,
such as rooftop and heat pump systems, etc. Therefore, the research outcomes of this
dissertation research may be easily extended to other types of HVAC systems and render
greater impact.
The remainder of this chapter will introduce the chilledwater system first, followed
by a description of the current status of dynamic modeling for building HVAC systems.
Then, a brief summary of control and optimization for building HVAC systems will be
described. Finally, this chapter will be concluded with a summary and organization of the
whole dissertation.
1.3 ChilledWater System in Buildings
In the central HVAC system of commercial buildings, the chilledwater system
supplies chilledwater to meet the cooling loads in buildings [9]. As shown in Figure 1.2,
a chilledwater cooling system typically consists of three subsystems: 1) cooling tower 2)
7
chiller and 3) airhandling units (AHU). Among these three subsystems, chiller can be
regarded as the heart of the overall system, which provides the cooling source, i.e., the
chilledwater to the conditioned space via AHU. The chiller system typically includes a
water pump to circulate the chilledwater generated from chillers evaporator to the
buildings AHU for cooling. Another water pump is usually placed in the chillers
condenser side to reject heat, which circulates the heated water to the cooling tower and
cooled back.
Figure 1.2: Schematic diagram of a chilledwater system in commercial buildings [10]
8
1.4 Dynamic Modeling of Building HVAC Systems
Dynamic modeling of HVAC system has received significant attention due to the
needs from control system design and fault detection and diagnostics scheme. Such
modeling problems are usually complicated by strong interactions among multiple
physical domains, such as thermodynamics, heat transfer, fluid mechanics, rigid body
dynamics, electromagnetics, electric circuits and control systems. Transient behaviors of
these physical phenomena are in principle very complicated. There are some challenges
in both the modeling and simulation perspectives. First, from the modeling viewpoint, the
mass, energy, and momentum balance are by default described as partial differential
equations (PDEs). For implementation to a modeling platform, such as Modelica [11],
these PDEs need to be transformed mathematically to ordinary differential equations
(ODEs) or differential algebraic equations (DAEs). Second, from computer simulation
viewpoint, simulation of the HVAC equipment dynamics may end up with solving a
mixture of ordinary differential equations (ODEs), linear/nonlinear algebraic equations
and possibly discrete equations, which in turn imposes great challenges to the robustness
of the numerical solvers. Therefore, in spite of extensive study in the past 30 years,
dynamic modeling of HVAC equipment remains an active research subject.
1.5 Control and Optimization of Building HVAC Systems
This section will address the topics of setpoint tracking and optimal control in
building HVAC system. For setpoint tracking purpose, the proportionalintegral
derivative (PID) controllers are by far the most widely used control algorithms in
industry. strm and Murray [12] said: More than 95% of all industrial control
problems are solved by PID control, although many of these controllers are actually
9
proportionalintegral (PI) controllers because derivative action is often not included.
However, PID controllers may be poorly tuned in practice. In basic control books, these
theories of designing and evaluating PID controllers are typically based on linear systems,
which may not work well in a highly nonlinear system such as buildings HVAC system.
To achieve better performance, PID controller has to be linearized based on certain
operating conditions and then switch different controller parameters based on different
operating ranges. However, such techniques may not work well for building HVAC
systems which typically has a wide range of operations [6]. In addition, even if the
control systems are well designed, building operators may not be well trained for good
controller tuning. Thus, there is a strong needs for automated controller tuning, i.e.,
adaptive control methods, for proper operation and maintenance of buildings. Therefore,
for building HVAC control, adaptive control methods seem desirable in the sense of
smart buildings or intelligent buildings.
For optimal control of building HVAC system, the optimal setpoints are typically
determined at the supervisory control level and then regulated by local controllers in the
HVAC systems, such as PID controllers described above [6]. However, majority of the
buildings today do not use control and optimization method to improve their energy
efficiency. Rather, heuristic approaches are typically applied [13]. Other industry, such as
the process industry, has much more applications for optimal controls than the building
industry [13]. Realtime optimization (RTO), a popular approach in process control, is an
online strategy for determining the setpoint of system operation via minimizing certain
cost function while observing to certain operating constraints [14]. Figure 1.3 shows the
block diagram of RTO and regulatory feedback control in process control. The outer loop
10
is the RTO and the inner loop is the regulatory control. The RTO layer handles relatively
slow dynamics and steadystate optimization is performed to determine the setpoint for
the innerloop control when some of the key measurements reach the steady state [14].
The regulatory control loop deals with dynamic models that have faster dynamics.
Figure 1.3: Block diagram for modelbased RTO and regulatory feedback control
(reproduced with permission from [14])
Traditionally, the RTO is achieved via steadystate optimization only, and is used at
supervisory level to assign optimal setpoints to local controllers. At the local controller
level, model predictive control (MPC) is typically used for setpointtracking purpose
based on the dynamic process model. The main disadvantage [15] for such configuration
is the possible model mismatch (process gain) between the steadystate model used in the
optimization and the dynamic model used in the regulation. Also, because of its
decoupled nature, the RTO layer may render delayed decision and miss the actual
optimum during the dynamic process.
Currently, many research efforts aim at combining the RTO and MPC together, and
trying to solve the economicoriented optimization problem in real time based on linear
11
or nonlinear dynamic models. Recently, such methods have been actively studied for
building HVAC controls [16, 17]. However, there may be several issues need to be
addressed for modelbased optimal control for building HVAC systems.
First, the time and effort for model development are often costly. The key rationale
for MPC to work well is built upon reliable dynamic models. However, for building
HVAC systems, reliable dynamic models may be difficult to obtain. The performance
map and design information for all HVAC equipment in a given building may not be
fully available. Even if all the design information and nominal operating conditions are
known when building up the models, the actual performance for the installed HVAC
equipment may drift far away from their manufacturers rated performance. Thus,
extensive time and effort would also need to spend on model calibration and validation
with different operating conditions to ensure that the models can predict reasonably well
for the dynamics and the process gains.
Second, the cost for instrumentation is often high for model validation and those
measurements needed for control design. Reliability of sensors is an addon issue and
uncertainty for modelbased design. Large number of measurements and highresolution
sensors are desirable for modelbased design, while the cost induced by sensor
acquisition, installation, calibration and maintenance may impose a heavy burden for
building development and operation. In addition, some measurements, such as the
relative humidity, are hard to measure accurately. Sensor redundancy may be needed to
reduce the risk of errors due to poorly calibrated sensor or problematic sensor locations,
which in turn further increases the cost. Future efforts on modelbased control design
shall address this issue and narrow the gap between technically sound solutions and
12
building industrys actual needs and concerns, i.e., reducing the number of sensors while
keeping the associated risks to low enough.
Finally, the online computational load may be an issue for implementation of large
scale nonlinear MPC. The online computational load in general depends on the fidelity
of the models, the optimization algorithms employed, the sampling rate, and the length of
the predicting horizon. Usually it is quite difficult to foresee the actual computational cost
until the control designer finishes all the required steps in modeling and controller design
and then initiate the tests. If the optimal solutions cannot be obtained within the desired
sampling interval, then the control designer may need to switch back and iterate on the
modeling and controller design process until the online computational load can be met.
In this dissertation study, the above limitations of modelbased optimal control are
addressed by considering the approach of selfoptimizing control which requires less or
little information from the process models in the design phase. Thus, the risks associated
with modeling uncertainties, sensor errors and costs are mitigated. Figure 1.4 illustrates
the block diagram of applying selfoptimizing control as the dynamic RTO layer for
building HVAC systems. The sensor output is the performance metric of interest, e.g.,
power or flow rate, etc. The selfoptimizing control strategy would then learn the gradient
of the performance metric w.r.t. the tuning parameters or setpoints during the dynamic
processes. The updated tuning parameters or setpoints are fed back to the building HVAC
system and then repeat the gradientlearning process until the optimum can be achieved.
13
Figure 1.4: Block diagram of applying selfoptimizing control as the dynamic RTO layer
for building HVAC system (modified from Figure 1.3)
The selfoptimizing control strategy employed in this study can be applied to
nonlinear timevarying HVAC systems with unknown parameters. In particular, a dither
based extremum seeking control strategy, is considered to learn the gradient of the actual
HVAC systems in realtime with potential tradeoffs among the regulations of different
tuning parameters or control setpoints. Such approach is often referred as modelfree
approach due to its selflearning and optimizing feature. In practice, the key measurement
needed is the performance metric to be optimized. More detailed descriptions of this
approach and its implementation are described in Chapter 2 and Chapter 5, respectively.
1.6 Organization of the Dissertation
The remainder of this dissertation is organized as follows.
Chapter 2 presents a comprehensive literature review on dynamic modeling of HVAC
systems and extremum seeking control. Success and limitation of the existing work will
be described.
Building HVAC Systems
Data
Reconciliation
RealTime
Gradient Learning
& Optimization
Updated Tuning
Parameters/Setpoints
Sensor
Output
Updated
Performance Metric
14
Chapter 3 describes the modeling of dynamic AHU model. The modeling details of
each component will be given, such as damper module, fans, ducts, zone and the chilled
water cooling coil. The emphasis is on the dynamic cooling coil model, which is
considered to be the most energyconsuming device in an AHU. The dynamic coil model
developed in this study is capable of predicting cooling performances under fullydry,
partiallydrypartiallywet, and fully wet conditions for both transient and steadystate
conditions. The experiment validation of the proposed coil model is given in Chapter 7.
Chapter 4 presents the dynamic modeling of centrifugal chiller system. Two
centrifugal chiller models are developed with the differences in compressor modeling. A
benchmark compressor model is developed following the previous ASHRAE sponsored
project, which is based on a performance map, while the new compressor model is
developed based on detailed physical and geometric parameters, and thus the resulting
chiller model is named as detailed chiller. For both chiller models, the condenser and
evaporator are modeled based on finitevolume method with detailed heat transfer
correlations and geometric configurations. For condenser and evaporator modeling, a new
concept, called variable refrigerant level based modeling, is studied and compared with
the model developed based on finitevolume method. Finally, the two dynamic chiller
models with different compressor models are compared based on a comprehensive
simulation study. For the detailed chiller model, the physical connections of these
dynamic components are considered quite complicated in terms of the differential
algebraic equations (DAE) to be solved. Consistent initialization of such DAE system is
difficult to obtain and special techniques should be applied for a successful simulation.
Such difficulties are further addressed in Chapter 5.
15
Chapter 5 presents the direct method and preprocessing scheme for consistent
initialization of the detailed centrifugal chiller system described by DAE system. First,
reasonable initial guesses for the chiller system components are computed once the
geometric parameters and design conditions are available. Then, a direct initialization
method is proposed for the consistent initialization of dynamic centrifugal compressor.
The mass and energy conservation during the initialization phase is further checked and
the computation efficiency of the proposed initialization method is also studied.
Chapter 6 presents the selfoptimizing operation of airside economizers using
Extremum Seeking Control (ESC). The motivation and principal operation of airside
economizer will be given first. Followed by the review of existing economizer controls
strategies and illustration of the proposed threestate selfoptimizing economizer control.
The overview of ESC is described next. The design guidelines and details of standard
ESC and antiwindup ESC design are further addressed. Finally, simulation studies are
conducted to demonstrate the potential of using ESC to achieve the minimal mechanical
cooling load in a selfoptimizing manner as well as to validate the effectiveness of the
proposed antiwindup ESC.
Chapter 7 presents the experimental validation of the cooling coil modeling and an
important assumption generally made in vapor compression cycle modeling. The
benchmark results of the developed chilledwater cooling coil model with experiment
data will be given. In addition, comparison results are presented to evaluate the transient
and steadystate predictions of the proposed model with the dynamic model developed by
Zhou and Braun as an ASHRAE sponsored project 1194RP conducted at Purdue
16
university [18]. Finally, experimental validation of the dp/dt assumption in heat
exchanger modeling for vapor compression cycle simulation will be given last.
Finally, conclusions and suggested future work are summarized in Chapter 8.
1.7 Summary
This chapter begins with the discussions about energy crisis in buildings. The energy
efficiency of buildings can be greatly improved by increasing the efficiency of building
HVAC systems. Modeling and simulation techniques are essential to preform building
energy analysis as well as design and validate advanced control strategies that may
improve the performance and efficiency of existing or new HVAC systems.
This dissertation study focuses on equipment level simulation and control. Chilled
water system, a common HVAC system in commercial buildings, is selected as the target
system in this study. The first theme of this dissertation study is to develop detailed
physicsbased dynamic HVAC models. These detailed dynamic HVAC models are built
based on firstprinciples and contain detailed formulations for heat and mass transfer as
well as pressure drop calculations. Based on the highfidelity dynamic models developed
in this study, another theme of this dissertation study is to apply selfoptimizing control
strategy to optimize the energy use for such systems. Such control strategy requires little
information from the models as in the control design phase, but its performance and
effectiveness should be validated with highfidelity dynamic models.
17
Chapter 2 Literature Review
This chapter presents a comprehensive literature review on modeling and simulation
techniques, previous work on dynamic modeling of HVAC systems, previous work on
Modelica based dynamic modeling of HVAC systems, and extremum seeking control.
The achievements and limitations of existing work will be discussed in detail and the
motivation and objective of this dissertation study will be justified accordingly.
2.1 Modeling and Simulation Techniques
Modeling and simulation tools are critical for controller design and fault detection
and diagnostics (FDD) of HVAC systems. Conventional HVAC modeling and simulation
programs are typically created based on imperative languages, and the physical models
are mixed with numerical solvers [19]. Block diagram modeling methods, such as
Simulink
A
=
A
(3.7)
where A = (D
2
)/4 is the reference area, C = (DN)/60, vel P A = (v
2
)/2, ex / v Q A = ,
total stat vel P P P A = A + A ,
fan
is the flow factor,
fan
is the pressure factor, Q is the flow rate,
A
ex
is the exhaust area, D is diameter of the impeller, v is the velocity of the outflow air,
N is the rotation speed in rpm, P
stat
is the static pressure, P
vel
is the velocity pressure,
and P
dynam,periph
is the peripheral dynamic pressure. a
1
, a
2
and a
3
are coefficients of the
polynomials relating the flow and pressure factors, which are fitted to the manufactures
fan performance data by the leastsquare estimation. A limited proportionalintegral (PI)
controller was used to regulate the rotation speed of the supply fan to maintain the static
pressure of the supply air duct at its setpoint. In addition, the rotation speed of the return
fan is synchronized by another limited PI controller, with the reference setting satisfying
the steadystate equilibrium of overall flow rate.
3.4 Zone
The zone model was adopted from the cabin model in the ACL. It consists of an air
volume and a cabin mass. The dynamic behavior of the zone is determined mainly by two
factors: 1) the thermal resistance, which is used to determine the heat flux between the air
volume and cabin mass, and 2) the heat capacity, which indicates the heat storage of the
total mass. Thus, the zone modeling is very analogous to the concept of resistance +
capacitance (RC) in electrical circuits. If multizone modeling is needed, the concept of
resistance + capacitor (RC) can still be applied by duplicating the instances of
48
resistances and capacitors based on the actual configurations of difference zones. In the
singlezone case shown below, the heat port is connected to the cabin mass and
represents external energy flow into the zone. The mass and energy balance equations in
the air volume are written as:
human
z
a b
dM
M M M
dt
= + +
(3.8)
human
z
a a b b z
dU
M h M h Q Q
dt
= + + +
(3.9)
where M
z
is the mass of air in the air volume,
a
M
and
b
M
is the approximate flow rate of the total water vapor from the
people in the conditioned zone, h
a
and h
b
are the inlet and outlet specific enthalpies,
respectively, U
z
is the total internal energy of the air, and
human
Q
is given by:
wall z
z
T T
Q
R
(3.10)
where T
wall
and T
z
are the mean temperatures of the cabin mass and the air volume,
respectively. R is the estimated thermal heat resistance from the cabin mass to the air
volume.
49
3.5 Cooling Coil
Cooling coil is the most important component between the primary plant (e.g. chiller)
and the air distribution system. It is among the fastest responding components in air
handling units. Therefore, the transient behavior of cooling coil may have significant
effect on the closedloop control performance [109].
Since Version 1.4, the AirConditioning Library has included a group of heat
exchanger models that are capable of simulating both transient and steadystate
operations [110]. The dynamic energy and mass balances are formulated based on the
finitevolume method [111]. The number of discretization at the refrigerant side is
proportional to that for the solid wall and the air side. The heat conduction in the solid
wall is modeled as a onedimensional problem perpendicular to the fluid flow direction.
In particular, the simulation results of a crosscounter flow evaporator model for an
R134a based automotive air conditioning system had been validated in an experiment
conducted by Chrysler [112]. The measured data were compared with the simulation
results of the medium properties and the steadystate heat transfer rates, for three sets of
boundary conditions in terms of the mass flow rate, the inlet temperature, the inlet
enthalpy, and the relative humidity of the ambient air. The heat transfer rates had good
consistency while the refrigerantside pressure drop and the airside water condensation
needed improvement. However, to the authors best knowledge, there has not been
experimental validation conducted for a chilledwater cooling coil model built upon
Modelica.
There were some challenges to directly use the heat exchanger models from the ACL
as the cooling coils in buildings. In the ACL Version 1.5, the choices of state variable
50
pairs include pressurespecificenthalpy (p, h), densitytemperature (, T), and mass
internal energy (M, U). Such choices are suitable for the air flow and twophase
refrigerants in the automotive refrigeration systems. However, for building HVAC
systems, especially for the cooling coils in air handling units, the working medium is
typically singlephase, i.e. water. Also, the temperature range is limited to the ambient
temperature variation. The current heat exchanger model in the ACL is based on two
phase flow with various types of the refrigerants, which covers a very broad range of
temperature variation for all the mediums to be used in a single heat exchanger model. It
is thus necessary to reformulate the existing heat exchanger model in the ACL to
accommodate the specific needs in building HVAC systems.
Finally, to apply the heat and mass transfer analogy to the dehumidifying process, it is
necessary to clarify and distinguish the definitions of Lewis number (Le) and Lewis
relation (N
Le
), which are often ambiguously stated in some heat and mass transfer
literature. According to [113], this problem was probably created and overlooked in air
conditioning literature due to the approximation of Le = 1 or N
Le
= 1. However, in order
to be consistent with other mass transfer literature and to allow for Le or N
Le
different
from unity in the design of the coil model, the use of a consistent definition is required.
The overall structure of the heat exchanger model in ACL is relatively complex,
many multiple inheritances and aggregation are applied. Figure 3.4 (a) shows the
hierarchical structure of the HXFluidFlow, a key component of the waterside heat
exchanger modeling. This figure follows the standard of Unified Modeling Language
(UML) [114], which is a standard graphical language for visualizing and specifying the
artificial structure involved in the software analysis and designs. Note that the shaded
51
blocks represent our efforts made to modify the original components from the ACL.
Modelica.Media, a fluid property Library that is part of the Modelica Standard Library,
contains IAPWSIF97 based water models, but this Library tries to combine all possible
fluid models from ideal gases to twophase fluids into a single framework to maximize its
generality. The resulting objectoriented structure is rather complex and the Library is
hard to understand for new developers and even users with a good knowledge of object
orientation [40]. In addition to the complex structure, it is not straightforward to
implement the IAPWSIF97 model defined in Modelica.Media Library to the heat
exchanger model in ACL due to different structures of the two models. Therefore, a new
medium model, called CoolWater, was created based on the IAPWSIF97 formulation
but with much simpler structure that was easily fitted to the existing heat exchanger
modeling. As can be observed in Figure 3.4 (a), the use of CoolWater as the medium
model can greatly simplify the waterside modeling structure (upper part of Figure 3.4 (a))
by decoupling its nested relationship to other packages from the Modelica.Media Library.
Similarly, Figure 3.4 (b) shows our efforts made to the airside heat exchanger modeling.
The details will be further explained in the later section.
52
(a): UML class diagram of waterside heat exchanger modeling
(b): UML class diagram of airside heat exchanger modeling
Figure 3.4: UML class diagrams of the water and airside heat exchanger modeling. The
shaded blocks represent the efforts made in this dissertation research to modify the
corresponding components from ACL
53
3.5.1 Medium Model Design and Implementation
An accurate water medium model is critical for the transient and steadystate
simulation of cooling and heating coils in the AHU. For the waterproperty calculation,
there are two major formulations as international standard, known as the IAPWS95 [115]
and the IAPWSIF97 [107, 115]. The former was developed for scientific computation,
while the latter was developed for industrial applications. Prior to the release of Version
1.4, the ACL had included a large set of medium models for many refrigerants, but not
the water medium. Since Version 1.5, the Library has adopted an LUT based
incompressible fluid (water) medium model for heat exchanger modeling. However, it
may have the following drawbacks. First, in the control volumes, the pressure responses
are decoupled from the thermal responses, which may lead to inaccurate mass
distribution predictions. Second, the incompressible water model will also result in
inaccurate pressure drop calculations, which will in turn affect the calculation of heat
transfer related properties.
To validate the accuracy of different formulations of water property model, the IF97
formulae based model (abbreviated as IF97 model later) and the LUT based
incompressible water model (abbreviated as LUT model later) were compared with the
IAPWS95 standard. Since the IAPWS95 model is generally considered as the scientific
standard, it can be considered as the true model. The FLUIDCAL program developed by
the Wagner group was used to obtain the IAPWS95 based water properties [116]. For
Dymola 6.1, the water medium in Modelica_Media follows the IF97 model, while the
water medium of ThermoFluidPro in the ACL Version 1.5 adopts the LUT model. The
comparison was conducted in the temperature range from 274.15 K to 373.15 K with an
54
increment of 5 K, and the pressure input was set at 5 bars for all cases. Table 3.1
summarizes the maximum errors of several properties based on the IF97 and LUT
models relative to those derived from the IAPWS95 standard. It was observed that for all
cases, IF97 is more accurate than the ACL Version 1.5. Notice that the maximum
relative errors for the specific entropy and C
v
are significantly larger. Figure 3.5 through
Figure 3.8 compare the relative errors of the IF97 and LUT models in density, specific
entropy, C
p
and C
v
, respectively. Note that C
p
and C
v
are assumed identical in the LUT
model. More discrepancies were observed for entropy and C
v
.
Table 3.1: Relative Errors of IF97 and LUT Water Models to IAPWS95 Standard in
Property Calculations
Water Property
Maximum Relative Error (%)
IF97 ACL Version 1.5
Density 0.0015 0.09
Specific Entropy 0.018 28.223
C
p
0.052 0.189
C
v
0.075 11.833
55
Figure 3.5: C
p
errors of the IF97 and LUT models relative to the IAPWS95 standard.
Figure 3.6: C
v
errors of the IF97 and LUT models relative to the IAPWS95 standard
56
Figure 3.7: Density errors of the IF97 and LUT models relative to the IAPWS95
standard
Figure 3.8: Specificentropy errors of the IF97 and LUT models relative to the IAPWS
95 standard
57
Within the Modelica_Media Library, a group of waterIF97 models have been well
defined to compute the physical properties for water in the liquid, gas and twophase
regions based on the IF97 formulae. However, there are several technical issues with
using these waterIF97 medium models directly in the functions of ACL. First, the
waterIF97 medium models contain both single and multiplephase calculations, but the
multiplephase portion is not needed for building chilledwater cooling system. Earlier
development in the ACL was well compatible with the automotive air conditioning
systems whose working medium are various kinds of refrigerants. The composition is a
critical argument contained in most functions. For cooling and heating coils in the AHU,
the singlephase water is the only working medium to deal with. The composition
argument in the existing functions results in significant inconvenience. For the single
phase water medium used in the heating/cooling coils, it would be more convenient to
remove the composition argument.
Second, in order to improve computational efficiency, it is necessary to remove the
multiphase portion of medium property computation in the ACL. In the ACL, such
calculation covers both single and multiplephase processes, involving not only the
balance equations of the dynamic control volumes, but also the calculations of various
thermodynamic states, such as density, enthalpy and specific heats, which are irrelevant
to the dynamic states of the control volumes. Also, there are many other computations
related to multiplephase processes, e.g. the twophase fraction calculation and the
moving boundary method, which are not useful in a singlephase computation. The
heating and cooling coils in the AHU involve only the singlephase water medium, which
58
is a much simpler case. If all the multiphase computations can be removed, the resultant
computational efficiency can be greatly improved.
Thirdly, the state variable pair needs to be updated to fit the water property
calculation in the AHU modeling. The refrigerants used by typical automotive air
conditioning systems are modeled on the basis of the Helmholtz functions with density
temperature (, T) as the pair of state variables. In many HVAC applications, it would be
more convenient if the water properties are based on the pairs of pressuretemperature (p,
T) or pressurespecificenthalpy (p, h). In addition, for physical property calculations in
the control volumes, the users can access the medium functions only at hierarchically
higher levels, which limits the customization or reformulation of these functions for
particular applications, especially when the userpreferred pair of state variables is not
supported in the existing package.
To address the above issues, a simpler and more efficient water model, named as
CoolWater, was developed based on Modelica_Media.Interfaces.PartialMedium. The
basic formulation of the CoolWater model was obtained from [117]. In particular, all
redundant and conflicting variables and options in the original waterIF97 model were
either removed or modified, e.g. the BaseProperties code. To be consistent with the
coding style and physical property calculations preserved in the ACL, several IF97
based lowlevel medium functions and utilities were adopted from the Modelica_Media
Library.
A heat exchanger model was developed based on the CoolWater medium described
above. Heat exchanger model is generally considered the most computationally intensive
59
entity in a refrigeration system [118]. To properly adapt the CoolWater model to the
refrigerant side, equations in the dynamic control volumes should be rewritten, but the
change should not degrade the overall inheritance structure and accuracy of the heat
exchanger model. Since the uppermost hierarchical structure of the heat exchanger is
composed of only a few lines of code, the work of implementing singlephase water
model should begin from the most rudimentary control volumes.
Different statevariable pairs were evaluated in order to find suitable choice to
achieve both engineering convenience and numerical efficiency. It was stated in [119]
that the massinternalenergy (M, U) pair could decrease the numerical efficiency. The
densitytemperature (, T) pair was considered by [7] a bad choice in the liquid region for
compressible fluids due to the amplification of numerical error.
The IF97 formulation divides the thermodynamic property calculations into five
regions [107]. Figure 3.9 presents the five regions that cover the entire range of validity
of the IF97 formulation. In particular, the fundamental equations for the Gibbs free
energy g(T, p) are applied to regions 1, 2, and 5. Since region 3 includes the critical point,
a fundamental equation for the Helmholtz free energy f(, T) that uses density and
temperature as the independent variables, is applied instead of the fundamental equation
for the Gibbs free energy. Region 4 is defined as the saturation curve [107].
60
Figure 3.9: Regions and equations of IAPWSIF97 (reproduced with permission from
Figure 1 in [107])
For the operation of building HVAC systems, only region 1 in Figure 3.9 is involved.
The basic equation for region 1 of IF97 is a fundamental equation of the specific Gibbs
free energy, expressed in a dimensionless form, g = g/(RT) [107]:
34
1
g( , )
( , ) (7.1 ) ( 1.222)
i i
i
I J
i
p T
n
RT
t t t t
=
= =
(3.11)
where the dimensionless pressure and temperature are p = p/p* and t = T*/T, respectively,
with p* = 16.53 MPa, T* = 1386 K, and R = 0.461526 kJ kg
1
K
1
. The coefficients n
i
and
f(,T) g(p,T)
T(p,h)
T(p,s)
p
s
(
T
)
T
s
(
p
)
g(p,T)
1 3 2
4
5
273.15 623.15 1073.15 2273.15
Temperature (K)
g(p,T)
T(p,h)
T(p,s)
61
exponents I
i
and J
i
of Eq. (3.11) are listed in Table 2 of [107]. The backward equation
T(p, h) for region 1 is formulated as a dimensionless form [107]:
34
*
1
( , )
( , ) ( 1)
i i
i
I J
i
T p h
n
T
u t q t q
=
= = +
(3.12)
where q = T/T
*
, p = p/p*, and h = h/h* with T
*
= 1 K , p* = 1 MPa, and h
*
= 2500 kJkg
1
[107]. The coefficients n
i
and exponents I
i
and J
i
of Eq. (3.12) are listed in Table 6 in
[107]. The numerical consistency of such calculation has also been studied in [107], For
10 million random pairs of p and h covering the entire region 1, the difference between
the temperatures (AT) from Eq. (3.11) and from Eq. (3.12) was calculated and the
absolute maximum difference (,AT,
max
) and the rootmeansquare difference (AT
RMS
) were
calculated as [107]:
max 25 mK; 13.4 mK; 23.6 mK
RMS
tol
T T T A = A = A = (3.13)
If the pressuretemperature (p, T) pair is selected as states, the fundamental equation
for the specific Gibbs free energy will be calculated explicitly in pressure and
temperature. If pressurespecificenthalpy (p, h) pair is adopted, the backward equation
will be first used to find the temperature first, and then the calculated temperature will
serve as an explicit input to the fundamental equation for the specific Gibbs free energy.
Currently, the statevariable pairs of pressuretemperature and pressurespecificenthalpy
have both been formulated into the heat exchanger model for comparison purpose. The
techniques of state variable transformation were performed in the dynamic balance
equations for pressuretemperature and pressurespecificenthalpy, respectively [7, 35].
To ensure consistent and convenient initialization, the pressuretemperature pair
62
(compared to the pressurespecificenthalpy pair) has been added into the initialization
options, since temperature is easier to set for HVAC operation rather than some other
variables such as specific enthalpy.
3.5.2 Dynamic Control Volumes
The dynamic control volumes were modeled based on a finite volume method (FVM)
that allows a numerically robust simulation of the cooling coil model. The cooling coil
model is composed of two fluid models, i.e. the air and water, and a wall element in
between. Heat transfer between the air, wall and water is modeled by a heat connector
model. Figure 3.10 shows the object diagram of the connections among the air, wall and
water volumes. The transient mass and energy storage is considered at the water and the
wall side but neglected at the air side.
Figure 3.10: Object diagram of the cooling coil composition of air, wall and water sub
models
A single pipe method [120] is used to describe the water path through the heat
exchanger. The flow through all parallel pipes in one flat tube and all flat tubes of one
63
water pass is handled as a single pipe flow with respective dimensions. The water flow
path is then segmented into a discrete number of volumes. The governing equations are
discretized into the socalled staggered grids based on an FVM approach [111, 121].
Each pipe segment is connected to a wall segment and an air channel that represents all
air flow channels in contact with one water segment. Heat conduction in the wall side is
modeled in one dimension and perpendicular to both fluids. Axial conduction is assumed
negligible. The mass of the coil metal is determined with the detailed geometric data
derived from the specifications of the coil tube and fins. One air channel is then modeled
as an air pipe with respective dimensions and segmentation in air flow direction. Since
the fundamental governing equations of a thermofluid system are PDEs, application of
the FVM requires these equations to be integrated over a fixed control volume in order to
obtain the corresponding ODEs.
The following mass and energy equations were written explicitly for pressure
specificenthalpy (p, h) and pressuretemperature (p, T) state pair at water side,
respectively.
 PressureSpecificEnthalpy (p, h) State Pair
For the mass and energy equations of (p, h) state pair, the basic statevariable
transformation is given by [35, 122]:
2
( )
p p
dp dM dU
V h
a dt h dt h dt
c c
= +
c c
(3.14)
2
(1 )
h h
dh dM dU
V h
a dt p dt p dt
c c
= +
c c
(3.15)
64
where
2
1 1
h p p h
a
c c
= +
c c
. Rearranging Eq. (3.14) and (3.15) yields [35, 122]:
2
=
1
p p
h h
dp dM
h
h h
dt dt
V
dh dU a
h
dt dt
p p
  c c
   
+

 
c c

 

c c
 
 

\ . \ .
c c
\ .
(3.16)
Finally, the equations in the dynamic control volume can be written as:
2
1
=
( )
h p
h p
h p
p h
dM dp
h h
p h V
dt dt
dU dh a
p h dt dt
  c c

c c


c c
   
+ + 
 
c c
\ .
 
c c
 
+
 
c c \ . \ .
(3.17)
 PressureTemperature (p, T) State Pair
For the pressuretemperature (p, T) state pair, the mass and energy in the control
volume can be written as:
( )
p T
p
T
dM dp dT
V
dt dt dt
c c
= +
c c
(3.18)
( )( ) ( )
p T p T
p p
T T
dU dp dT p u dp u dT
V h V
dt dt dt dt dt
c c c c
= + + +
c c c c
(3.19)
where M is the mass, U is the total internal energy, V is the volume, p is the pressure, is
the density, T is the temperature, u is the specific internal energy and h is the specific
65
enthalpy. The four partial derivatives
p
T
c
c
p
T
c
c p
T
u c
c
p
u
T
c
c
could be computed using
rudimentary IF97 property functions implemented in the CoolWater model.
3.5.3 Friction Factor
As part of this dissertation study, a new explicit form of the friction factor calculation
for smooth pipes is developed and used in the heat transfer calculation in later sections
and other heat transfer applications. This piece of work is systematically documented in
[123]. For completeness, the following reiterates the contents and results described in
[123].
3.5.3.1 Introduction
Calculation of the friction factor for smooth pipes in fully developed flow is
important for heat transfer applications. Many different coefficients in the Prandtl
equation [124], or PKN correlation [125, 126], are cited in the literature. This report
reviews these coefficients and presents the original Prandtl equation based on the Fanning
friction factor. In addition, a new explicit equation is presented to approximate the PKN
correlation. This equation yields a more accurate approximation than the existing ones.
Two types of friction factors are often cited in the literature. i.e., friction factors for
rough pipes and for smooth pipes, respectively. The friction factor for rough pipes is
often useful for detailed pressure drop calculation. There are already some good explicit
equations [127129] to approximate the implicit ColebrookWhite equation [130] for
rough pipes. For heat and mass transfer applications, the friction factor for smooth pipes
is often involved. Reynolds [131] was probably the first to use friction factor in the heat
66
and mass transfer calculations, in which he formulated the wellknown Reynolds analogy.
The friction factor has played an important role in the heat and mass transfer literatures.
For example, the Reynolds analogy and the Colburn and Chilton jfactor analogy [132,
133] both depend on the friction factor as the sole parameter to interconnect the heat and
mass transfer characteristics. Specific to heat transfer applications, many other
researchers also applied the friction factor for smooth pipes in their correlations, such as
Prandtl and Taylor [134, 135], Von Krmn [136], Prandtl [124] and many others [137
142]. The Gnielinski correlation [143], which is considered the most popular correlation
for computing heat transfer characteristics of internal flow in the transition and turbulent
regions, also needs the friction factor for smooth pipes as the independent variable.
The Prandtl equation [124], which has also been known as the KrmnNikuradse
Relation [144, 145] and the PKN correlation [125, 126], is regarded as an accurate
equation for determining the friction factor [125]. However, this equation is not explicit
in the Reynolds number, and thus makes it inefficient for computer simulations. Many
researchers have developed explicit approximations to the Prandtl equation. Blasius [146]
was the first to apply the similarity theory and to formulate that the friction factor is a
function of the Reynolds number in turbulent flow. His formula is recommended for 4000
Re 10
5
. For higher Re, the maximum errorF
4
F was about 30%. The prediction for the
higher Re range of 10
5
Re 10
7
was later proposed by Nikuradse [145], and the errors
for this particular Re range was improved to be within 2%. However, this correlation may
not be suitable for low and midrange Re (4000 Re < 10
5
), the maximum error was
about 14%. Colebrook [130] presented a simple explicit equation valid for the whole Re
4
For convenient notation, the error mentioned is referred to 100 Cf, PKNCf, Model/Cf, PKN
67
range. This equation was also recommended by Haaland [128], with prediction error
within 1.7% for 4000 Re < 10
4
and about 0.6% for 10
4
Re 10
7
. The prediction
results from other researchers will be detailed in section 3. Among all the existing
explicit equations, the most accurate approximation [125] is from Techo et al. [147].
Chen [148] presented a systematic study on the Prandtl equation and formulated two
approximation equations. The first equation was obtained by using the Blasius equation
as the initial value and substituting it into the righthand side of the Prandtl equation. To
iterate, the obtained equation was then substituted into the righthand side of the Prandtl
equation again until the desired accuracy was achieved. The second equation was
obtained by reducing the number of significant figures from the first equation. However,
as Chen [148] pointed out, the two equations are only as good as, if not better than the
approximation equation from Techo et al. [147]. In addition, with the advance of
computational power, the number of significant figures has now become less of an issue.
Recently, [149] derived a representation of the PKN correlation using the Lambert W
function. The derived equation is presented explicitly in Reynolds number. However,
since the Lambert W function is inherently implicit, this solution often requires iteration
processes. The author compared the percent differences between the result of the PKN
correlation and those from Jain [150], Techo et al. [147] and [151], respectively. The
relative difference of the derived approximation equation to the PKN correlation was not
reported. More recently, Avci and Karagoz [152] proposed a novel explicit friction factor
based on a new logarithmic velocity profile. The derived equation was claimed to be
indistinguishable from the solution of the PKN correlation. The maximum errors of this
equation are around 0.5% for 4000 Re < 10
4
and 1% for 10
4
Re 10
7
.
68
Another issue is that the PKN equation has been cited differently in terms of the
coefficients in many heat transfer literatures. The reasons may be twofold. First, different
authors have used different definitions for the friction factor, i.e., Darcy friction factor or
Fanning friction factor. For example, the original Gnielinski equation [143] was derived
based on the Darcy friction factor. However, to calculate the friction factor with better
accuracy, the readers often need to search heat transfer references for answers. A
common mistake could be substitution of the wrong friction factor, i.e., the correlation
for the Fanning friction factor, which would lead to incorrect Nusselt number. Second,
the Prandtl equation [124] was expressed in the base10 logarithm, but some authors [125,
126] preferred natural logarithm representations. Consequently, many rearranged PKN
correlations appeared in the heat transfer literatures with different coefficients. The
rearranged PKN correlations are summarized in Table 3.2.
Table 3.2: Summary of CitedPKN Correlations from Some Heat Transfer References
Sources CitedPKN Correlations
Handbook of SinglePhase Convective
Heat Transfer [125]
1
1.7372ln(Re ) 0.3946
f
f
C
C
=
Handbook of Heat Transfer [126]
1
1.7272ln(Re ) 0.3946
f
f
C
C
=
Handbook of Heat Transfer [153]
1
1.74ln(Re ) 0.3939
f
f
C
C
=
Convective Heat and Mass Transfer [142]
1
1.7395ln(Re ) 0.3907
f
f
C
C
=
69
In this study, an exact transformation to natural logarithm form is derived in the
following subsection.
3.5.3.2 Explicit Equation for Friction Factor Calculation
The original Prandtl equation [124] is given by:
10
1
2log (Re ) 0.8 f
f
= (3.20)
where f is the Darcy friction factor. Substituting f = 4C
f
, where C
f
is the Fanning friction
factor, into Eq. (3.20) yields:
10
1
2log (Re2 ) 0.8
2
f
f
C
C
= (3.21)
Converting the base of logarithm in Eq. (3.21) from 10 to e, and rearranging the outcome
yields:
( ) 10
1 4
ln Re 4log 2 1.6
ln(10)
f
f
C
C
= + (3.22)
which can be further simplified into:
1
1.73718ln(Re ) 0.39588
f
f
C
C
~ (3.23)
The KrmnNikuradse Relation cited from Bejan and Kraus [154] gives:
1
1.737ln(Re ) 0.396
f
f
C
C
= (3.24)
70
It is important to note that Eq. (3.22) is same as Eq. (3.23) except for the roundoff
errors after the third decimal digit. Therefore, it is recommended that Eq. (3.22) and Eq.
Eq. (3.23) are the correct forms of the PKN correlation and are reliable for the benchmark
of other approximation equations.
The relationship of the Fanning friction factor and the Reynolds number is implicit in
Eq. (3.22) and needs to be solved numerically. To obtain an explicit equation, the
software package EES
. (3.27)
where j is the index of test points, Re
init
is the initial Reynolds number tested, Re
final
is the
final Reynolds number tested, and N is the total number of test points which was set to
1000. The Re
step
was thus computed to be 0.003401. For the proposed model, the
optimization methods described above were applied to determine the unknown
coefficients in Eq. (3.25) as [123]:
2 3
0.001570232 0.394203137 2.534153311
ln(Re) ln(Re) ln(Re)
f
C
= + + . (3.28)
One disadvantage of many existing correlations is that they often give acceptable
results limited to a certain Re range. Furthermore, among all the existing correlations
being studied, there is still some room for improvement in the 4000 s Re s 10000, which
is an important range for many heat transfer applications, e.g., chilledwater cooling coils.
The accuracy of the friction factor calculation will influence not only the pressure drop
calculations, but also the heat transfer characteristics. To give a comprehensive
comparison, the maximum errors and the RSS were compared based on three Re ranges,
i.e., 4000 s Re s 10
4
, 10
4
< Re s 10
5
and 10
5
< Re s 10
7
. As previously mentioned, the
PKN model in Eq. (3.22) is considered as the true value in the comparison. Table 3.4
and Table 3.5 compare the maximum errors and RSS from the PKN model (Eq. (3.22)),
respectively. For better illustration of the comparison results, Table 3.4 and Table 3.5 are
graphically shown as Figure 3.11 and Figure 3.12, respectively. Notice that the results
from the proposed model (Eq. (3.28)) are bolded. The proposed model yields better
results with smaller maximum errors and RSS.
73
Table 3.4: Comparison of Maximum Errors of Existing Equations to the PKN Correlation
Model
Maximum Error (%)
4000 Re 10
4
10
4
< Re 10
5
10
5
< Re 10
7
Blasius [146] 2.418993945 2.817661587 30.56774613
Nikuradse [145] 14.36715796 8.920393954 1.908100776
Drew et al. [158] 3.078669632 3.084158127 4.610963449
Colebrook [130] 1.724498244 0.638647729 0.641283717
McAdams [159] 12.24738174 5.590932348 9.605374156
Filonenko [160] 3.82086276 1.91228246 0.277528134
Techo et al. [147] 0.291991201 0.076751118 0.076396283
Jain [150] 1.460591397 0.808104805 0.808695302
Swamme and Jain [127] 3.820862760 1.912282460 0.277528134
Chen [148] 0.420950938 0.158386623 0.076274694
Bhatti and Shah 1 [161] 18.80284197 11.17457982 3.558076151
Bhatti and Shah 2 [161] 0.963440367 1.043152319 2.595920413
Avci and Karagoz [152] 0.484376421 0.981200423 0.983332497
Proposed 0.036754682 0.036759109 0.036756982
Table 3.5: Comparison of RSS with Existing Equations to the PKN Correlation
Model
Residual Sum of Squares (RSS)
4000 Re 10
4
10
4
< Re 10
5
10
5
< Re 10
7
Blasius [146] 1.8108110
6
5.4590710
6
1.0337410
4
Nikuradse [145] 1.2783310
4
3.4906710
5
6.120810
7
74
Drew et al. [158] 6.8206610
6
6.8191410
6
6.8058110
7
Colebrook [130] 1.1344110
6
1.5544310
6
1.302910
7
McAdams [159] 7.7979610
5
7.9464510
6
5.6352610
6
Filonenko [160] 7.7765110
6
1.1770510
6
1.4091910
8
Techo et al. [147] 3.0202510
8
2.6880310
9
1.5515610
9
Jain [150] 7.1343410
7
2.8245910
7
2.0641710
7
Swamme and Jain [127] 7.7765110
6
1.1770510
6
1.4091910
8
Chen [148] 7.9336110
8
5.9907310
9
4.301910
10
Bhatti and Shah 1 [161] 2.1459910
4
4.4734510
5
1.793510
6
Bhatti and Shah 2 [161] 2.8603110
7
6.6524810
7
2.0458710
6
Avci and Karagoz [152] 6.8855410
8
6.4368210
7
3.6980910
7
Proposed 3.6178210
10
9.0021910
10
3.7529110
10
75
Figure 3.11: Comparison of maximum errors of existing equations to the PKN correlation
(Table 3.4), plot created on logarithmic scale with base 10
76
Figure 3.12: Comparison of RSS with existing equations to the PKN correlation (Table
3.5), plot created on the logarithmic scale with base 10
In addition, the computation efficiency of all existing correlations was investigated.
The test of computation time was conducted by first running all the test points defined in
Eq. (3.26) and Eq. (3.27) for the operating range of 4000 Re 10
7
and then repeating
the computation process for 10 times. Table 3.6 and Figure 3.13 summarize the
comparison results.
77
Table 3.6: Comparison of Computation Time among Existing Equations
Model Computation Time (sec)
Blasius [146] 0.005956
Nikuradse [145] 0.004528
Drew et al. [158] 0.004591
Colebrook [130] 0.004255
McAdams [159] 0.004424
Filonenko [160] 0.004095
Techo et al. [147] 0.004614
Jain [150] 0.033555
Swamme and Jain [127] 0.035141
Chen [148] 0.004681
Bhatti and Shah 1 [161] 0.004804
Bhatti and Shah 2 [161] 0.004531
Avci and Karagoz [152] 1744.969836
Proposed 0.00536
78
Figure 3.13: Comparison of computation time among existing equations (Table 3.6), plot
created on the logarithmic scale with base 10
Furthermore, Techo et al.s model [147] gives better accuracy over other existing
equations as shown in Table 3.4 and Table 3.5. The following comparison was conducted
between Techo et als approximation and the proposed model based on the whole range
of Reynolds number of interest, i.e.
7
4000 Re 10 s s , as shown in Figure 3.14. Over the
whole Re range, the RSS of the proposed model and Techo et al.s are 1.63710
9
and
3.44410
8
, respectively. As can be observed from Figure 3.14, a major improvement was
79
made in 4000 Re 10
4
. The proposed model also predicts better results in other
operating ranges by reducing the RSS and the maximum errors. For the convenience of
comparison, the model proposed by Techo et al. is:
1 Re
1.7372ln
1.964ln(Re) 3.8215
f
C
 
=

\ .
(3.29)
Figure 3.14: Comparison of percent errors to the PKN correlation (Eq. (3.22)) between
the proposed model (Eq. (3.28)) and Techo et al.s approximation (Eq. (3.29)).
3.5.4 Water Side Heat Transfer
For the air and waterside convection, the heat transfer coefficients are needed to
determine the heat transfer rate. For the water side, the fluid is assumed to be Newtonian
and only singlephase is concerned. As stated in [162], for laminar flow, none of the
80
available correlations is capable of predicting the coil performance accurately at water
side when the Reynolds number is less than 2300. In that case, the use of a laminar,
developing flow correlation at these Reynolds numbers might be considered reasonable.
The SiederTate correlation [163] was adopted to serve this need:
1/3 0.14
Re Pr
Nu 1.86
/
w w
w
i s
L D
   
=
 
\ . \ .
(3.30)
where Nu
w
is the waterside Nusselt number, Re
w
is the waterside Reynolds number,
Pr
w
is the waterside Prandtl number, L is the tube length, D
i
is the inner pipe diameter,
is the dynamic viscosity, and
s
is the dynamic viscosity evaluated at the average value of
the mean temperature. Note that for tube length L, the length of one tube pass (twice of
the tube length in the general sense) is adopted in order to compensate for the
performance underprediction. It is believed that although the flow is laminar at such low
Reynolds numbers, heat transfer is enhanced by secondary flows generated by buoyancy
forces and the tube bends. These effects are not taken into consideration by the Sieder
Tate correlation [162]. For transition flow, the Gnielinski correlation [143] is applied:
1/ 2 2/3
( / 2)(Re 1000) Pr
Nu
1 12.7( / 2) (Pr 1)
f w w
w
f w
C
C
=
+
(3.31)
The Fanning friction factor C
f
is developed in [123] (see Eq. (3.28)) by approximating
the Implicit Prandtl Equation [164], which is also known as the Krmn Nikuradse
Relation [144, 165, 166] and PKN correlation [125, 126]. The details for the friction
factor are described in Section 3.5.3.
81
The transition between the SiederTate and the Gnielinski correlations were smoothed
out by adopting a specialized hyperbolic tangent function, the socalled spliceFunction,
from the Modelica Standard Library:
*
* * *
w
Eq. 3.29 Re 1
Nu(Re ) (Re )Eq. 3.30 [1 (Re )]Eq. 3.29 1 Re 1
Eq. 3.30
w
w w w
 
s
= + < <
*
Re 1
w
>
(3.32)
{ }
* *
1 1
(Re ) tanh tan[arcsin(1) Re ]
2 2
w w
 = + (3.33)
where
*
Re (Re 2000) / 300
w w
= . The heat transfer coefficient is then continuously
determined by the above two correlations with the Nusselt number covering the entire
region of the Reynolds number of interest. The following equation is used to calculate the
heat transfer coefficient from the mean Nusselt number:
wat
Nu
m
h
k
h
D
= (3.34)
where k is the thermal conductivity of the fluid, D
h
is the hydraulic diameter of the pipe,
and Nu
m
is the mean Nusselt number.
3.5.5 Water Side Pressure Drop
Pressure drop needs to be considered due to its impact on the driving temperature
difference [61]. For the water side, the pressure drop is given by:
Loss,wat
v  v 
2
h
L
P
D
A = (3.35)
82
where D
h
is the hydraulic diameter, is the density, v is the mean velocity, L is the length
of the pipe, and is the Darcy friction factor which depends on the Reynolds number and
the relative roughness of inner pipe wall. Notice that if the flow is laminar, the analytical
solution to the friction factor is given by 64/Re. However, it is impossible to directly
implement this expression into the computer program. For example, if the Reynolds
number goes to zero in the middle of the simulation process, then divisionbyzero would
occur and the computation would be terminated. A remedy for this singularity problem
was introduced in Elmqvist et al. [117] with a redefined friction factor
2
. The waterside
pressure drop could then be computed from the new friction factor:
2
Loss,wat 2 3
2
h
L
P
D
A = (3.36)
where
2
Re  Re  = and Re = vD
h
/. The characteristics of the pressure drop could be
described respectively in three regions. The critical Reynolds number Re
1
is determined
based on Idelchik [167]:
1
0.0065/
745 0.0065
Re
745 0.0065
e
e
A
A s
=
A >
(3.37)
where is the relative roughness. For the laminar flow in region 1 (Re < Re
1
), the
friction factor is given by [117]:
2
64Re = (3.38)
For the transition flow in region 2 (Re
1
< Re < 4000),
2
depends on more factors than
just the Reynolds number and the relative roughness, and thus only rough approximations
83
are available for this region. The approximation equation is given by [117]:
2
1 1
2 1
1
Re Re
1 log log
Re Re
Re
64Re
Re
A B
(    
+ +
(  
\ . \ .
=
 

\ .
(3.39)
where
2
2 3 1 1
0.9
2 1 2 3 2
2 1
log0.25(Re / ) log64Re 4
3 4
log(Re / Re ) Re
log(Re / Re )
a a
a a
A
=
( )
2
1 2 3 1
0.9
2 3 2 2 1
2
2 1
4 log0.25(Re / ) log64Re
3 2
Re log(Re / Re )
log Re / Re
a a
a a
B
+
=
where Re
2
= 4000, a
1
= 0.55.740.9/ln(10), a
2
= /3.7 + 5.74/Re
2
0.9
, a
3
= log(a
2
).
For the turbulent flow in region 3 (Re > 4000), given the mass flow rate (i.e. the
Reynolds number),
2
can be determined by an approximation of the inverse of the
ColebrookWhite equation [127]:
2
2
0.9
Re
0.25 sgn(Re)
5.74
lg( )
3.7  Re 
(
(
= (
A
(
+
(
(3.40)
where sgn() is the signum function, and is the relative roughness of inner pipe wall.
3.5.6 Air Side Heat and Mass Transfer
For the airside, the energy storage is assumed to be negligible. The total heat transfer
rate is the sum of sensible and latent heat transfer. The dominant resistance for an air
84
cooled heat exchanger is generally on the air side, which may account for 85% or more of
the total resistances. When the dew point of the inlet air temperature is above the surface
temperature, or equivalently, when the humidity ratio of the inlet moist air is greater than
the humidity ratio evaluated at the saturated wall surface, condensation will occur. Since
water condensate has a high contact angle on the aluminum fins, the water may adhere as
droplets causing bridging between the fins, and thus increasing the airside pressure drop
[168]. On the other hand, as described in [169], the tubes in some cooling coils are
spaced quite apart so that the water condensed from the air on the coil surface readily
drains off without obstructing the air flow. In that case, the airside pressure drop is still
quite small. Therefore, in general, these uncertainties make it very difficult to determine
the air flow pattern. Many investigators have been studied the airside heat transfer
characteristics of typical finandtube heat exchangers. McQuiston [170, 171] developed
the wellknown heat transfer and friction correlations based on five plate finandtube
heat exchangers. [172] developed heat transfer and friction correlations based on nine
sample finandtube heat exchangers. These correlations were compared to McQuistons
work and improvement was reported. This work was later extended by [173] into a tube
bytube approach to determining the heat transfer and friction coefficient for 17 finand
tube heat exchangers. Additional work was done to determine the mass transfer
characteristics. In our study, the heat transfer correlation developed by [173] is adopted:
row row
0.002061 0.625 0.001921 0.068 0.0575
row Dc row
1.49 Re (0.00583 0.825)
N N
c
j N N c
+
= + (3.41)
p,o
o
A
A
c = (3.42)
85
2/3
air pa
(Pr)
c c
h j m C
=
(3.43)
where j
c
is the sensible Colburn factor, Re
Dc
is the Reynolds number based on the outside
diameter (include collar), N
row
is the number of tube rows, A
o
is the total surface area, and
A
p,o
is the outside surface area of tubes,
Pr is the Prandtl number and is assumed to be a
constant value of 0.72, h
c
is the sensible heat transfer coefficient,
air
m is the air flow rate,
and C
pa
is the specific heat capacity of air.
3.5.7 Lewis Number and Lewis Relation
Klinkenberg and Mooy [174] appeared to be the first to define the ratio /D as Lewis
number (Le). Although this notation was challenged by Arnold at the discussion section
of their paper, the definition of Le = /D has been accepted by widely cited literatures in
heat and mass transfer [175178]. In addition, particular attention is needed for the mass
transfer coefficient. Since different heat and mass transfer literature may prefer different
notations. To make it clear in the following analysis, h
D
denotes the mass transfer
coefficient using the mass fraction of water vapor per unit mass of moist air as the driving
potential, denotes the mass transfer coefficient using the mass fraction of water vapor
per unit mass of dry air as the driving potential, k denotes the mass transfer coefficient
using partial pressure difference as the driving potential. Unlike the convective heat
transfer coefficient, there are very few available correlations to determine the mass
transfer coefficient. Many researchers have sought to discover the relationship between
the heat and mass transfer. Lewis was probably the first to formulate the following
approximate relation [179]:
86
h
s
k
= (3.44)
where h is the convective heat transfer coefficient, k is the mass transfer coefficient, and s
was termed as humid heat, which has the same meaning as specific heat capacity. This
relation was validated by Lewis through experiments. For moist air, h/k and s were shown
to be 0.236 and 0.238, respectively. From the above relation, one may infer that the ratio
h/(ks) is approximately unity. Note that the ratio h/(ks) is essentially the same expression
as the Lewis relation termed by Kusuda [180] except for the definition of the mass
transfer coefficient.
Figure 3.15 shows an element of the cooling coil examined in the heat and mass
transfer analysis. In this analysis, the dry surface occurs near the inlet section of the air
flowing in the cooling coil. The transition between dry and wet surface is driven by the
humidityratio difference of the moist air between the bulk flow and the saturated wall
surface.
Figure 3.15: Cooling and dehumidifying of moist air over the cooling coil surface
The total heat transferred to the whole fin and tube surface region in each control
volume is thus given by:
87
sen lat
dQ dQ dQ = + (3.45)
The sensible heat dQ
sen
transferred between the air and the condensed water film
surface is given by:
sen bulk wall
( )
c
dQ h dA T T = (3.46)
where h
c
is the sensible convective heat transfer coefficient, dA is the differential heat
transfer area, T
bulk
is the bulk flow temperature of the moist air, and T
wall
is the moist air
temperature at the fin and tube wall surface. The latent heat dQ
lat
transferred is given by:
lat cw fg
dQ m h dA = (3.47)
where h
fg
is the latent heat of vaporization. To determine the condensing flow rate
cw
m ,
the mass transfer coefficient h
D
and the driving potential are needed. The complete
equation to solve for the mass transfer rate of water vapor may be given by [181]:
sat,cw bulk sat,cw
( ) / (1 )
cw D
m h X X X = (3.48)
where X
bulk
is the mass fraction of water vapor in bulk flow moist air, and X
sat,cw
is the
mass fraction of water vapor of saturated moist air at the condensed water film. Note that
the presence of 1/(1X
sat,cw
), known as the blowing effect [181], is a correction term for
induced flow, which is associated with a convective velocity required to counterbalance
the diffusion of air to the condensing water surface and due to the impermeability
condition at this surface. However, as pointed out by Prata [182], this term seemed a
defect of constructing the heat and mass transfer analogy. Redefining the mass transfer
coefficient h
D
is one remedy suggested by Prata [182] in order to formulate a tight
88
correlation between the ratio of the heat and mass transfer coefficients and the Lewi4s
number. The maximum deviation of using this correction method was shown to be 0.59%
from our own calculation and less than 0.6% as reported by Prata [182].
Next, the driving potential needs to be determined. The following two equations
frequently appear in heat and mass transfer literature:
cw bulk sat,cw
( )
D
m h X X = (3.49)
cw bulk sat,cw
( ) m W W o = (3.50)
where h
D
is the mass transfer coefficient defined based on dryair, X
bulk
is the vapor mass
fraction defined based on moistair, X
sat,cw
is the vapor fraction of saturated moist air at
the water film defined based on moistair, is the mass transfer coefficient defined based
on dryair, W
bulk
is the humidity ratio of bulk flow moist air, and W
sat,cw
, is the humidity
ratio of saturated moist air at the condensed water film.
In this study, the dryair based humidityratio difference (Eq. (3.50)) is selected as the
driving potential over the moistair based approach (Eq. (3.49)), which has the following
advantages. First, the total mass of the moistair varies continuously along the airflow
path, which makes it very difficult to accurately determine the total mass at any given
location and time. In contrast, the mass of dryair is always invariant to condensation.
Second, the humidity ratio is widely used in the air conditioning literatures, and it is very
convenient to visualize the state of the moist air by simply examining the psychrometric
chart. As a result, the condensation will occur when the humidity ratio of the bulk flow is
greater than that of the saturated wall surface. The transition between dry and wet
89
surfaces will affect the total heat transfer rates. Similar to Eq. (3.33), the spliceFunction
is used to interpolate the region where the humidity ratio driving potential is positive,
negative, and close to zero. For the dry surface conditions, the driving potential is
negative. The output of the spliceFunction will be zero, and thus only the sensible heat
transfer will be calculated. For the wet surface conditions, the driving potential is positive,
the condensing flow rate will be determined based on the driving potential and hence the
latent heat transfer rate. The implementation form of the condensing flow rate is given by:
*
* *
bulk sat,cw
*
bulk sat,cw
0 1
( ) ( ) ( 1,1)
( ) 1
cw
W
m W W W W
W W W
 o
o
A s
= A A e
A >
(3.51)
{ }
* *
1 1
( ) tanh tan[arcsin(1) ]
2 2
W W  A = A + (3.52)
where
* 7
bulk sat,cw
( ) / (10 ) W W W
A = . Finally, the condensing flow rate can be solved by
applying the heat and mass transfer analogy.
The socalled Lewis relation had been systematically studied by Kusuda [180], who
described the following empirical Lewis relation of heat and vapor over the wet
evaporating surface of many geometrical configurations (such as flatplate, cylinder,
sphere, and packed beds):
2/3
Le
N
c c
D p p
h h
h C C D
o
o
 
= ~ ~

\ .
(3.53)
According to Kusuda [180], the Lewis relation is not unity and is nearly equal to (/D)
2/3
for forced convection and (/D)
0.48
for natural convection. With the Lewis relation, the
90
condensing flow rate could be determined by combining Eqs. (3.50) and (3.53).
3.5.8 Air Side Pressure Drop
For the air side, the friction factor is calculated based on the following empirical
correlations given in [172]:
1.3405
0.5653 0.1026 1.3343
Dc row
28.209Re
p
C
F
f N
D
c
 
=

\ .
(3.54)
where f is the friction factor, Re
Dc
is the Reynolds number based on the outside diameter
(include collar), N
row
is the number of tube rows, F
p
is the fin pitch, D
c
is the outside
diameter of the tube, A
o
is the total surface area, and A
p,o
is the outside surface area of
tubes.
3.6 Summary
This chapter presents the development of dynamic models in an AHU. The modeling
details of the dynamic models for airmixing box, air duct, fan, zone and cooling coil are
presented in sequence. Major work was carried out to build a highfidelity chilledwater
cooling coil. Compared to the existing work, the major improvement of this work is
summarized as follows:
1) detailed latent heat transfer calculation based on Lewis relation at the air side;
2) highaccuracy calculation of thermodynamic and transport properties at the water
side;
3) convenient and efficient statepair formulation for the dynamic mass and energy
balance equations at the water side; and
91
4) detailed and accurate heat transfer and pressuredrop correlations at both the air
and water sides.
The benchmark results of the proposed dynamic cooling coil model with
experimental data and the dynamic model developed in ASHRAE RP1194 [18] will be
described in Chapter 7. Dynamic simulation of the AHU will be described in Chapter 6
for the study of the proposed selfoptimizing control scheme.
92
Chapter 4 Dynamic Modeling of Centrifugal Chiller for
Building Air Conditioning Systems
Chapter 4 presents component level dynamic modeling of watercooled centrifugal
chillers. An addon package TIL.Chillers is developed based on TIL. The TIL.Chillers
contains: 1) centrifugal compressors, 2) condensers and evaporators, and 3) expansion
devices. This package includes simpletocomplex level component models to facilitate
modeluse for different simulation purposes. This chapter focuses on describing the
modeling details of each individual component. System level integration of chiller
components and chiller simulation with a proposed scheme of consistent initialization
will be presented in Chapter 5.
4.1 Introduction
In watercooled buildings, chiller system plays a critical role in regulating chilled
water loop and condenser water loop throughout buildings HVAC systems. Thus, in
HVAC literature, the chiller system is commonly referred to as the primary loop. In a
broad sense, chiller systems can be classified as absorption and vaporcompression
refrigeration systems. A major difference is that the absorption system is driven by heat
to achieve the cooling effect, while the vaporcompression system is driven by
mechanical work from compressors. The chiller system described below refers to vapor
compression refrigeration cycle, which typically includes compressor, condenser,
expansion device, and evaporator. The type of chiller is commonly determined from the
type of its compressor. ASHRAE [183] broadly defines the compressors into two
93
categories: 1) positivedisplacement compressors, which includes reciprocating, rotary
(rolling piston, rotary vane and screw) and orbital (scroll and trochoidal), and 2) dynamic
compressors such as centrifugal compressors. According to ASHRAE [183], the positive
displacement compressors increase the pressure and temperature of vapor by reducing the
volume of the compression chamber through mechanical work. By comparison, the
dynamic compressors continuously convert the kinematic energy from vapor into
pressure lift through mechanical work. For vapor compression refrigeration systems,
Stoecker and Jones [184] list four most common types of compressors: reciprocating,
screw, centrifugal and vane.
The U.S. Department of Energy (DOE) has conducted studies on energy consumption
of HVAC equipment in commercial buildings [9]. It was stated that watercooled chiller
systems can achieve higher efficiency than other types of equipment such as packaged
AC units. They also discovered that centrifugal chillers are responsible for the largest
share of primary cooling energy use over other type of chillers. Table 4.1 shows the
equipment breakdown of chillers for primary cooling energy use [9].
Table 4.1: Equipment breakdown of primary cooling energy use [9]
Type of Chiller Energy Use
Rotary Screw Chillers 3%
Reciprocating Chillers 12%
Centrifugal Chillers 14%
94
This dissertation study focuses on the dynamic modeling of component models in
watercooled centrifugal chiller systems. Figure 4.1 shows the schematic diagram of a
centrifugal chiller system interacting with the condenser water loop and the chilledwater
loop, respectively.
Figure 4.1: Schematic drawing of watercooled centrifugal chiller system interacted with
cooling tower and air handling unit
Figure 4.2 is a schematic of a centrifugal chiller and Figure 4.3 shows the cyclic
operation of an R134abased centrifugal chiller with each thermodynamic state point
labelled on a ph (pressureenthalpy) diagram. In its principal operation, the circuiting
refrigerant enters the centrifugal compressor as vapor and is compressed to a higher
pressure and temperature. The compressed refrigerant is superheated and then passes
through a condenser where heat is removed and rejected to the cold water from the
cooling tower side. Ideally, the superheated refrigerant will be condensed into the
95
saturated liquid state with high pressure, i.e. the subcooled refrigerant. The subcooled
refrigerant is then routed through an expansion valve where it undergoes a sudden drop in
pressure, resulting in a twophase refrigerant flow with low pressure. Finally, the two
phase refrigerant is routed through an evaporator, where heat transfer occurs and heat is
removed by the return water from the air handling unit (AHU) side. Consequently, the
refrigerant with twophase state is again heated to the vapor state, and the chilled water is
produced by rejecting heat to the refrigerant flow in the evaporator. To complete the
chiller cycle, the refrigerant vapor is again fed into the centrifugal compressor and the
above operations repeat in sequence.
Figure 4.2 Schematic drawing of centrifugal chiller
96
Figure 4.3: ph diagram of cyclic operation in R134abased centrifugal chiller system
4.2 ObjectOriented Modeling of Centrifugal Chiller System
There has been significant amount of work done on steadystate modeling for
centrifugal chillers [185190]. However, quality dynamic models are needed for
centrifugal chiller in order to develop advanced control techniques and fault detection and
identification (FDI) schemes. The most recent work on dynamic modeling of centrifugal
chiller is from Bendapudi [118]. This dissertation study is intended to improve the
existing work by introducing:
1) detailed physicsbased centrifugal compressor model for chiller simulation;
2) heat exchanger modeling with variable refrigerant level; and
97
3) integration of detailed physicsbased compressor model with finitevolume
(FV) based heat exchanger models and method to overcoming convergence
issue for such systems (see Chapter 5).
Other contributions of this study include: 4) implement FV based shellandtube heat
exchanger models based on existing components in TIL, and 5) implement the constant
speed centrifugal compressor model and thermal expansion valve model developed by
Bendapudi [118] into TIL.Chillers.
This section presents the modeling details of each chiller component developed in
TIL.Chillers. The calculations of thermal and fluid transport properties in TIL.Chillers
are handled by the TILMedia Library developed by TLKThermo [42].
4.2.1 Centrifugal Compressor
The centrifugal compressor, while being the most energy consuming equipment in the
chiller system, is difficult to model accurately because of its complicated geometry,
highly nonlinear behavior, and unstable operation (surge) regime. For efficient operation,
the capacity of the compressor is generally manipulated by adjusting the inlet guide vane
(IGV) position and the compressor speed. The inlet guide vane (IGV) (see Figure 4.4),
also known as prerotation vane, is generally designed to change the entering flow angle
of the impeller and thus control the amount of vapor flow entering the compressor. Figure
4.5 is a schematic of the centrifugal compressor model connected with the condenser and
evaporator as the boundary conditions.
98
Figure 4.4: Crosssectional view of a centrifugal compressor (reproduced with permission
from Figure 47 in 2008 ASHRAE Handbook: HVAC Systems and Equipment, p. 37.28)
Figure 4.5: Schematic drawing of a centrifugal compression system in chiller with inlet
guide vane (IGV) and speed control
For compressor modeling, there are generally two modeling methods: 1)
performancemap based semiempirical modeling and 2) physicsbased modeling. The
performancemap based method is easy to implement when relevant experiment data are
available. The limitation of this method lies in the fact that the performance map obtained
99
by regression analysis is suitable mostly for steadystate analysis; it may not be suitable
for studying transient behaviors such as load changes, startup and shutdown processes.
In addition, the performance map thus obtained is highly specific to the machine design
and conditions for which the experimental data are obtained. If there is any change in
design geometry or variation of component characteristics through the operating life, new
data should be obtained and the associated regression analysis shall be performed again.
Performancemap based (or datadriven) modeling relies on little knowledge about
system design, which facilitates system/component characteristics at the user end, since
design details are normally unknown and/or proprietary. On the other hand, the
disconnection of design parameters from the input/output data becomes a disadvantage
for design optimization, accurate simulation of system transient, fault detection and
diagnostics, and rootcause analysis.
The physicsbased method takes into account mass, momentum, and energy balance
equations as well as geometric configurations of the actual devices, which may avoid the
disadvantages of the performance map method aforementioned. With quality modeling
and proper numerical solution, the physicsbased method can well capture the transient
behaviors of compressor characteristics. Again, this approach is constrained by the full
knowledge of all relevant physical parameters required to establish the models, which
may pose as a difficulty under certain circumstances. As a remedy, if experimental data
are available to calibrate the models, tuning factors can be introduced to minimize
modeling errors using optimizationbased methods.
100
4.2.1.1 PerformanceMap based Modeling
In this class of methods, compressor performance maps can be normalized based on
certain nominal operating conditions and then scaled for actual compressor operation, or
can be piecewisely fitted to experiment data for each operating condition of interest [7].
The TIL.Chillers package adopts the centrifugal compressor developed by Bendapudi
[118]. The key modeling equations are described as follows. For more details, refer to
Bendapudi [118].
For compressor modeling, there are two important classes of performance maps: 1)
efficiency map, and 2) pressureflow map. The compressor efficiency can be studied
based on isentropic or polytropic analysis. In Bendapudis work, the polytropic analysis
is adopted and the efficiency map is modeled as [118]:
2 2
0 1 2 3 4 p p p
a a V a V a W a W q = + + + +
(4.1)
where V
(
(
= = +
(
(
(
(4.5)
where is defined as
1
c,in 1 1
cot
b
Ar

=
.
(4.6)
The definitions of fluid velocity angle in Eq. (4.5) and Eq. (4.6) are given in Appendix A.
In centrifugal air compression systems, such as the system in [197], the compressor inlet
density
c,in
can be assumed as a constant due to the openloop structure of the system.
Thus, would be a constant in such systems. However, for vapor compression
refrigeration systems such as chillers, the compressor inlet density
c,in
cannot be treated
as constant, rather, it is determined by chillers actual operations and the corresponding
responses at the evaporator side. Note that in Eq. (4.5), is defined as,
2
c,in 1 1
cot
1
b
m
Ar

o
e
 
=


\ .
, (4.7)
In the current development, a radially vaned impeller is considered with no backsweep
and thus
2
90
b
 =
[194]. Equation (4.7) can thus be simplified to
o = (4.8)
103
 Inlet Guide Vane (IGV)
Two simplified IGV models are implemented in the current study: 1) linear
characteristic and (2) quadratic characteristic. The models can be improved when
experiment data are available. The IGV is assumed to be fully open when
IGV
1 o = , and
fully closed when
IGV
0 o = .
1) Linear Characteristic
This is the simplest model and the idea comes from Bendapudis work [118]. Note
that in Bendapudis work, the linear relationship is established between the IGV position
and the actual mass flow rate based on the maximum mass flow rate obtained from
regression analysis (see Eq. (4.4)); while in this study, the linear relationship is
established between the IGV position and the pressure loss across IGV based on a
resistance
linear
R , which is an empirical constant to be determined with experimental data.
IGV IGV linear
(1 ) p R o A = (4.9)
2) Quadratic Characteristic
In this model, the nonlinear relations of pressure loss and mass flow rate are
described by a quadratic equation:
2
IGV IGV nonlinear
(1 ) p R m o A = (4.10)
Note that
nonlinear
R is an empirical constant. There could a variation of Eq. (4.10) if
experimental data support: similar to the damper model developed in Chapter 2, the term
104
( )
IGV nonlinear
1 R o in Eq. (4.10) can be replaced by an exponential profile over certain
range of IGV openings. If high accuracy is needed,
nonlinear
R can be fitted piecewisely as
long as smooth transitions can be ensured between different fitted models using suitable
numerical techniques.
Due to the lack of experimental data in the present study, the IGV models proposed as
above are intended for ideaproof purpose in order to complete the modeling of the
centrifugal compressor. With detailed experimental data available, more accurate model
can be obtained in terms of the guide vane angular positions, the pressure differentials
across the IGV, and the refrigerant flow rates.
 Isentropic Efficiency
The isentropic efficiency can be determined based on the ideal energy transfer and
various losses at different locations of the compressor. Based on the original definition by
Cumpsty [198], Gravdahl and Egeland [194] enhanced the calculation of isentropic
efficiency q
is
by considering more loss terms described in Watson and Janota [196] and
Cumpsty [198].
ideal
is 1 bf
ideal loss
( , )
c v d
h
mU
h h
q q q q q
A
= A A A A
A + A
(4.11)
The details about the calculation of isentropic efficiency are summarized in Appendix A.
105
Code Segment 4.1 shows the Modelica implementation of the isentropic efficiency.
Code Segment 4.1: Modelica implementation of isentropic efficiency
In the above implementation, the isentropic efficiency is fixed as constant for
negative mass flow rates in order to avoid computing the isentropic efficiency when the
flow direction is reversed,
The tuning mechanism of the centrifugal compressor is characterized by the spool
dynamics, which is modeled as a firstorder differential equation [197]:
d c
d
J
dt
e
t t = (4.12)
4.2.2 Condenser and Evaporator
4.2.2.1 Finite Volume Based Modeling
The modeling of condenser and evaporator requires quality heat exchanger models
since they can capture the major transients in the chiller system. There are mainly two
partial model PartialCentrifugalCompressor
parameter SI.MassFlowRate mdotLimit = 1e4 "Transition flow rate for reversed flow";
equation
// Compute isentropic efficiency when flow rate is positive
isEff_pos = dh_Ideal/(dh_Ideal+dh_Loss)etaBackFlowetaClearanceetaVolute
etaDiffusion;
// Compute isentropic efficiency over fullrange of mass flow rate
isEff = spliceFunction (isEff_pos, isEff_zero, portA.m_flow, mdotLimit) isEff_zero is
the isentropic efficiency when mass flow rate equals zero;
106
methods for heat exchangers modeling: 1) finite volume (FV) method and 2) moving
boundary (MB) method. The FV method is based on discretizing the heat exchanger into
a number of fixed control volumes and integrating the balance equations over each
volume. The MB method is based on dividing the heat exchanger into variablelength
control volumes that correspond to each phase region. The number of control volumes
and the length of each control volume depend on the actual system operation. For shell
andtube heat exchanger modeling in centrifugal chiller systems, the comparison between
the FV and MB methods has been extensively studied by Bendapudi et al. [199]. For
simulation of both individual components and the whole chiller system, the FV method
was found to be more robust for scenarios of startup and transient load change. The MB
method was also found to be less accurate for charge prediction. Although the FV method
amounted to about 20% increase in computation time, the tradeoff is worthwhile
considering the improvement in the modeling accuracy. Therefore, the FV method was
adopted in this study.
 Shell and Tube Sides Geometry
According to Bendapudi [118], assumptions of concentric, tubeintube and counter
flow can be applied to the shellandtube heat exchanger modeling in chillers. For both
condenser and evaporator, the water flows at the tubeside and the refrigerant flows at the
shellside. The shellside heat transfer area is computed from the outer surface area of the
water tubes and a surface enhancement factor. In this study, the condenser and evaporator
were developed based on the TubeAndTube heat exchanger model in TIL. Figure 4.6
shows the Modelica model layout of the condenser in TIL.
107
Figure 4.6: Modelica model layout of the condenser model (reproduced and modified
with permission from TILs TubeAndTube heat exchanger model)
To adapt the above model to the flooded type shellandtube heat exchangers, some
modifications are needed. The difference in geometric dimensions between a shelland
tube heat exchanger and a tubebytube heat exchanger mainly lies at the shell side. As
the refrigerant flows at the shell side, the face flow area and the volume of the refrigerant
can be computed as [118]:
2 2
1
( )
4
f s o
A D Nd t = (4.13)
tot tot f
V A L = (4.14)
The Reynolds number in each refrigerant cell is determined based on the average
velocity and is given by:
mean
mean
r f
m
v
A
=
(4.15)
108
mean r,in r,out
( ) / 2 m m m = + (4.16)
mean
Re
o
r
v d
v
= (4.17)
 Mass and Energy Balance Equations
With the finite volume method, both the refrigerant and water sides are discretized
into a fixed number of control volumes (called cells in TIL). For the refrigerant side, a
singlecell model named as RefrigerantCell
5
is modified from TIL.Cells; while for the
water side, the same model of LiquidCell is adopted from TIL.Cells. Both the mass and
energy balance equations are formulated in these two singlecell models, respectively. An
integer parameter nCells (number of cells) can be used to duplicate multiple instances
of the RefrigerantCell and LiquidCell and construct a finite volume based heat exchanger
model.
In RefrigerantCell, the mass balance equation is given as [40]:
r,in r,out
r
r
d
V m m
dt
= + (4.18)
The energy balance is derived by applying the first law of thermodynamics on an
open control volume [40]:
r,in r,in r,out r,out
1
( ) ( )
r
r r r r
r
dh dp
m h h m h h Q V
dt M dt
 
= + + +

\ .
(4.19)
5
The modified RefrigerantCell model includes phasedependent heat transfer calculations for condenser and evaporator.
109
To avoid automatic differentiation in Dymola, the time derivative of density (
r
d dt )
is manually transformed in terms of the state variables [40]:
r
r r r r
r h p
d dh dp
dt h dt p dt
    c c
= +
 
c c
\ . \ .
(4.20)
Note that a Modelica tool such as Dymola includes algorithms to symbolically transform
differential equations in terms of their states. In the Modelica.Media Library, to transform
differential equations into desired state variables, the user needs to set the flag
preferredMediumStates = true when declaring the medium model BaseProperties. Such
option is also available in TILMedia by setting the flag stateSelectPreferForInputs = true
when declaring thermodynamic variables.
It is important to note that / dp dt
6
is assumed invariant for each pressure level
7
. This
is a key assumption originally proposed by Lemke [200] with experimental validation
and later implemented into TIL by Richter [40]. As illustrated in Figure 4.7, the
assumption can be expressed as [200]:
in out
dp dp
dt dt
= (4.21)
According to Richter [20], the key benefit of this assumption is that / dp dt would
become known quantities instead of unknowns in Eq. (4.19) and Eq. (4.20), which would
improve the numerical efficiency and robustness of refrigerant cycle simulations.
6
The subscript r is not placed on dp/dt because dp/dt is only defined at the refrigerant side and neglected at the water side.
7
By neglecting pressure drops, chillers have two pressure levels: 1) condenserside pressure, and 2) evaporatorside pressure.
110
Figure 4.7: Illustrative diagram for the dp/dt assumption
However, there is a limitation in Lemkes experimental validation. The experiments
were conducted on a gas cooler from a transcritical refrigeration system, where the
refrigerant is in the supercritical state only. Thus, it is not appropriate to draw a quick
conclusion that this assumption can also be applied to typical vapor compression
refrigeration systems such as chillers in this study, which would undergo phase
transitions rather than the supercritical state only. This concern is addressed in this study
with experimental investigation on a vapor compression chiller. The details of this study
and experimental results are summarized in Chapter 7.
In LiquidCell, the state variable includes temperature only. Thus, the calculation of
time derivative of pressure (dp/dt) is neglected. Similar to RefrigerantCell, the mass and
energy balance equations are manually transformed to the state variable.
 Water and Refrigerant Sides Heat Transfer
The convective heat transfer coefficients are implemented in the basic cell models of
both RefrigerantCell and LiquidCell. Figure 4.8 shows the schematic for the water and
refrigerant sides heat transfer in the basic cell models.
111
Figure 4.8: Illustration of water and refrigerant sides heat transfer in basic cell models
In Figure 4.8, the convective heat transfer rates on the water and refrigerant sides can
be determined from Newtons law of cooling [40]:
wall
( )
w w w w
Q A T T o =
(4.22)
wall
( )
r r r r
Q A T T o =
(4.23)
Neglecting axial conduction, the energy storage at the wall cell can be modeled as
[163]:
wall
wall p,wall r w
dT
M c Q Q
dt
= +
(4.24)
Water Side Heat Transfer
For the water side, the flow region is turbulent mainly due to high water flow rates.
Thus, the Gnielinski correlation [143] can be applied:
112
1/2 2/3
( / 2)(Re 1000) Pr
Nu
1 12.7( / 2) (Pr 1)
f w w
w
f w
C
C
=
+
(4.25)
The Fanning friction factor developed by in [123] is applied (see Chapter 3 for details):
2 3
0.0015702 0.3942031 2.5341533
ln(Re) ln(Re) ln(Re)
f
C
= + + (4.26)
Finally, the water side connective heat transfer coefficient can be determined as [163]:
Nu
w w
w
i
k
d
o = (4.27)
Refrigerant Side Heat Transfer
Compared to the water side, the heat transfer at the refrigerant side is more complex
due to the twophase flow situation. For condenser, suitable convective heat transfer
correlation is needed for describing condensation process in the twophase region; while
for evaporator, a suitable correlation is needed for the boiling process in the twophase
region. In addition, smooth transitions between the singlephase and twophase heat
transfer correlations are necessary for implementation.
For condensation in the twophase region, extensive study has been reported for
condensation inside tubes. Also, there have been good models for condensation over a
single tube. However, the condensation over tube bundles is much more complicated.
The presence of neighboring tubes brings forth additional complexities. As described in
[201], In the idealized case, the condensation from a given tube is assumed to drain by
gravity to the lower tubes in a continuous and laminar sheet. In reality, depending on the
spacingtodiameter ratio of the tubes and whether they are arranged in a staggered or
inline configuration, the condensate from one tube many not fall on the tube directly
113
below it, but instead may flow sideways. In addition, experimental study has shown that
condensate does not drain from a horizontal tube in a continuous sheet but in discrete
droplets along the tube axis. When these droplets strike the lower tube, considerable
splashing can occur, causing ripples and turbulence in the condensate film. Perhaps most
important of all, large vapor velocities can create significant shear forces on the
condensate, stripping it away, independent of gravity.
For laminar film condensation on a horizontal tube, Dhir and Lienhard [202]
proposed a correlation based on Nusselts analysis [203]. It is also widely used to
correlate the heat transfer for the condenser model with tube bundle on the shell side.
This model does not consider the condensation inundation effect, heat transfer coefficient
increase in a tube bundle, or the surface enhancement effect. Belghazi et al. [204]
conducted a study on condensation on the exterior of a bank of smooth tubes for pure
fluid and for zeotropic binary mixture HFC23/HFC134a. For the test of pure HFC134a,
the experimental data on the first tube deviate from the Nusselt theory [203] by 5%. The
effect of heat flux to the condensation inundation was considered by modifying Chens
correlation [205] based on the exponent in Kerns correlation [206]. However, the
method proposed in this paper cannot predict the increase of heat transfer coefficient in
the tube bundle since it is difficult to determine the pattern of condensate falling from the
upper rows. Later, Belghazi et al. [207] studied the film condensation of downward vapor
flowing on staggered bundles of horizontal finned tubes, using HFC134a and the binary
zeotropic mixture HFC23/HFC134 as the refrigerants. Experimental study was conducted
based on five commercial tubes with dissimilar fin pitches. Experimental results indicated
that the condensation curve method underestimated the heat transfer coefficient for the
114
first row. A modified condensation curve method was then proposed by adding the effect
of Lewis number. This method has good agreement (within 10%) with the experimental
data in terms of the total heat flux prediction. This method also improved the prediction
of heat transfer coefficient for the first row, but still it cannot accurately predict the heat
transfer coefficient all other rows because it is difficult to calculate the increase of heat
transfer coefficient due to the upper rows.
For the condenser side, Dhir and Lienhards correlation [202] is adopted with a
surface enhancement factor to account for the effect of surface enhancement and the
increase of heat transfer coefficient in tube bundles due to particular patterns of the
condensate. For the evaporator side, the boiling heat transfer correlation is adopted from
Bendapudis work [118] based on an empirical correlation from the manufacturer.
The geometries of the condenser and evaporator in this work follow the models from
Bendapudi et als experimental study [199]. The heat transfer coefficients in their model
were determined with some empirical correlations from the manufacturers, and the
surface enhancement factors were determined with experimental data. This study adopts
the empirical heat transfer models and the surface enhancement factors from Bendapudi
et als work [199] for the ideaproof purpose.
In general, refrigerant flow in the evaporator undergoes transitions between two
phase and superheat regions, as illustrated in Figure 4.9. For detailed flow patterns of
forced convection boiling in a tube, refer to [163]. In Figure 4.9, the boiling heat transfer
coefficient
boil
o is given by [118]:
115
boil
boil 1
2
Q
e
e
o
''
= +
(4.28)
Figure 4.9: Illustrative diagram of heat transfer regions in the evaporator
The superheat heat transfer coefficient is given by Bendapudi et al. [199]:
sf,sup
sup
Re Pr
r r r
o
C k
d
o = (4.29)
Figure 4.10 schematically shows the heat transfer regions in the condenser. The heat
transfer coefficient of condensation is given by [208]:
1/4
3
fg
cond sf,cond
sat
( )
( )
l l v l
l t o
g k h
C
T T d
o
' (
=
(
(4.30)
Figure 4.10: Illustrative diagram of heat transfer regions in the condenser
Note that
fg
h' is the modified latent heat of vaporization [209]. The same correlation for
superheat heat transfer described in Eq. (4.29) is adopted except for a different surface
enhancement factor. The subcooling heat transfer coefficient is also given in Bendapudi
116
et al. [199]:
sf,sub
sub
Re Pr
r r r
o
C k
d
o = (4.30)
For condenser, smooth transition of heat transfer coefficients among different phase
regions is reinforced and implemented as follows. Implementation for the evaporator side
is very similar.
Code Segment 4.2: Implementation of heat transfer coefficients at the condenser side
4.2.2.2 Variable Refrigerant Level (VRL) Based Modeling
Recently, TLKThermo proposed a simplified method of heat exchanger modeling
[210], i.e. the socalled variable refrigerant level (VRL) modeling approach. The basic
idea of VRL modeling is to connect a large tank of refrigerant with a bank of water tubes
via a heat port. The liquid level of the refrigerant, or equivalently the volume ratio of
equation
if noEvent(properties.x < 0.5) then
// Handle transitions between twophase and subcooling
alpha=CF_alpha_condenser*spliceFunction(alpha_twophase,alpha_subcool,properties.x 
0.02,0.0001) CF_alpha_condenser is the overall surface enhancement factor of
condenser;
else
// Handle transitions between twophase and superheat
alpha=CF_alpha_condenser*spliceFunction(alpha_superheat,alpha_twophase,properties.x
 0.98,0.0001);
end if;
117
liquid refrigerant to the total volume of the tank, is varied depending on the magnitude of
heat transfer between the water tubes and the refrigerant in the tank.
More specifically, the large tank has two outlet fluid ports, i.e. the liquid port and the
gas port. Same tasks are performed as for a phase separator, i.e., when the liquid port is
connected, the outlet flow will be liquid refrigerant; when the gas port is connected, the
outlet flow will be refrigerant vapor. The bank of water tubes can be realized by
multiplying the heat transfer area of a single liquid tube by the number of tubes (i.e. the
parameter nParallelTubes). Heat transfer from the water tubes to the refrigerant in the
tank is realized by the interface of a heat port placed on the liquid tube and the tank
models.
The volume of the tank is fixed, but the volume ratio of refrigerant in the tank can be
easily monitored. In this modeling approach, the heat transfer coefficient in the liquid
tubes can be tuned to match the experimental results for the outlet refrigerant and liquids.
Compared to detailed formulation such as the finite volume method, which requires
extensive experimental data as well as detailed geometric information to calibrate, this
approach has the advantage in model validation when there are not enough information
and experimental data available to calibrate the model with details.
During this dissertation study, this approach is implemented with existing component
models in TIL. For the modeling of shellandtube heat exchangers, the component model
IdealSeparator can be adopted together with LiquidTube model to formulate dynamic
models of condenser and evaporator. Figure 4.11 shows the implementation of condenser
model with variable refrigerant level in TIL. The construction of an evaporator model is
118
very similar. In Figure 4.11, a Pressure State Element is placed at the refrigerant side to
compute the time derivative of pressures (dp/dt). The Ideal Separator is referred to the
aforementioned big tank with two outlet fluid ports. For the case of condenser, the liquid
port is connected to the outlet refrigerant flow. The Heat Port between the Liquid
Tube and the Ideal Separator is connected to transfer thermal energy. Upon clicking
the liquid Tube model, it is convenient to set up the number of liquid tubes to be
included into the condenser via the parameter nParallelTubes, and the corresponding
heat transfer area can be automatically updated.
Figure 4.11: Modelica model of a VRLbased condenser in TIL
119
4.2.3 Expansion Device
The expansion valve is a key component for the refrigeration cycle, which is typically
used to control the superheat at the outlet of the evaporator by adjusting the valve
opening. The key functionality of the expansion valve is to regulate the refrigerant
pressure to a lower level while roughly maintaining the enthalpy of the refrigerant. In
centrifugal chillers, orifice plates or float valves are often used to throttle the pressures
and thus control the flow rates [211]. In the current model development, an orifice plate
model is adopted from TIL by assuming a quadratic relation of the pressure difference
across the valve and the corresponding mass flow rate. The flow rate through the
expansion valve can be adjusted by the effective flow area (A
eff
). The TIL.Chillers
package also adopts the thermal expansion valve developed by Bendapudi [118].
4.2.3.1 Orifice Plate
Assuming onedimensional flow, for the control volume drawn in Figure 4.12, the
mass balance of an orifice plate is given by
v,in v,out
0 m m + = (4.31)
Figure 4.12: Control volume of an orifice plate
120
Further assume that the throttle process is isenthalpic, i.e.
v,in v,out
h h = (4.32)
For compressible flow, a detailed formula for characterizing the orifice flow is given in
[212]:
( )
out
v,in
2
out in
1 /
D
A
m C x
A A
c =
(4.33)
where ( )
v,in v,in v,out
2 x p p = . In Eq. (4.33), an expansion factor () is adopted to
characterize the degree of compressibility, e.g., when 1 c = , Eq. (4.33) becomes the
orifice flow equation for incompressible flow. As Parr states [213], the difficulty is to
determine the discharge coefficient (C
D
) and it is also very difficult to quantify the
expansion factor () since it is related to the specific heat ratio, the inlet and outlet cross
sectional areas, and the pressures. In TILs orifice valve model, these terms are lumped
into a single parameter called effective flow area (A
eff
). The mass flow rate through the
orifice valve is thus simplified to
v,in eff
m A x =
(4.34)
In Eq. (4.34), it can be seen that
v,in
/ dm dx as 0 x , this would cause Newtonlike
solvers to stop due to the infinite derivative. In order to avoid such difficulty, an
alternative implementation of the valve equation is considered. For simplicity in the
analysis followed, let us define
2
smooth
0
eff _smooth
m
x
A
 
=


\ .
(4.35)
121
In TILs implementation, the valve equation is approximated by a cubic polynomial
3
v,in eff 0
0 0
5 1
4 4
x x
m A x
x x
(
 
( =

( \ .
(4.36)
within the interval
 
0 0
, x x x e
.
As 0 x , it follows
v,in
0 m , which gives the same
solution as Eq. (4.34). The cubic polynomial
3
0 0
5 1
4 4
x x
x x
 

\ .
has finite derivatives over the
whole interval
 
0 0
, x x x e and the problem of division by an infinite derivative is thus
avoided.
4.2.3.2 Thermal Expansion Valve
Unlike the orifice plate, the opening of a thermal expansion valve (TXV
8
) is regulated
by the pressure differential based on the temperature feedback at the evaporator outlet. In
particular, the operation of the TXV is determined by the pressure differential among
three forces: 1) bulb pressure
bulb
p , 2) evaporator pressure
e
p , and 3) minimum opening
pressure (parameter)
min
dp . Figure 4.13 illustrates the forces acting on a TXV.
8
Some literature also abbreviate thermal expansion valve as TEV.
122
Figure 4.13: Illustrative diagram of forces acting on a TXV
The TIL.Chillers package adopts the TXV model developed by Bendapudi [118]. The
key model equations are summarized as follows. Like the office plate, The TXV model is
also modeled as a static device under the assumption of isenthalpic process. The heat
transfer between the sensing bulb and the evaporator outlet temperature is modeled based
on the lumped capacitance method with a time constant C
b
[118]:
e,out
( )
b
b b
dT
C T T
dt
= (4.37)
The valve lift is computed based on the pressure differential from the three forces in
Figure 4.13 [118]:
lift spring bulb min
( )
e
y k p p dp = (4.38)
with k
spring
being the spring constant. For more details, refer to Bendapudi [118].
In the model implementation, the valve lift is numerically bounded to avoid a
negative lift, and the corresponding Modelica codes are given below.
123
Code Segment 4.3: Numerical bounds for the thermal expansion valve model
Notice that
bulb sat e,out
P ( ) p T = , (4.39)
and
sat
P ( ) is a thermodynamic state function that computes the saturation pressure given
the temperature. The valve flow area is approximated as a quadratic equation based on
the valve lift [118].
2
0 lift 1 lift v
A f y f y = + (4.40)
Finally, the mass flow rate across the valve is determined as [118]
v,in d v
m C A x =
(4.41)
where a constant discharge coefficient C
d
is adopted. Again, similar to the
implementation for the orifice plate model, the same numerical treatment is applied to
avoid infinite derivative at x = 0. The parameters of the TXV model are summarized in
Appendix B.
equation
dp_push = max(1e5, (p_bulb  portB.p  dp_min)) Lower bound;
// Upper bound
lift = min(maxlift, k_spring*(dp_push/1e3)) Divide by 1e3 to convert to kPa;
124
4.3 Simulation Study
4.3.1 Comparison of VRLBased and FVBased Heat Exchanger Models
Simulation study was conducted to compare the heat exchanger models developed
with the FV method (detailed) and the VRL method (simplified) respectively. In the
comparison tests, same inlet boundary conditions were applied for the condenser
operation in chiller. At the refrigerant side, the inlet pressure and specific enthalpy were
set as 9.04 10
5
Pa and 4.34 10
5
J/kg, respectively. A ramp change of the inlet
refrigerant mass flow rate was applied and will be described later. At the water side, the
inlet pressure, temperature, and mass flow rate were 1.210
5
Pa, 303.15K (30 C), and
16.7 kg/s, respectively. The geometric parameters of the FVbased model follow the
work by Bendapudi [118] and are summarized in Appendix B. In the VRLbased model
(see Figure 4.11), the volume of the Ideal Separator was set to the same as the
refrigerantside volume computed in the FVbased model. To study the transient
responses, after all the timevarying variables reached their first steady states, a 10
second ramp change was applied to the inlet refrigerant mass flow rate at 100 seconds.
Figure 4.14 and Figure 4.15 show the testing configurations of the FVbased model
and the VRLbased model, respectively.
125
Figure 4.14: Test configuration of FVbased heat exchanger simulation
Figure 4.15: Test configuration of VRLbased heat exchanger simulation
Since the FVbased model and the VRLbased model have different geometric and
system configurations, a comprised (and indeed suboptimal) comparison was performed
by tuning the same steadystate results of either outlet refrigerant enthalpy (case 1) or
refrigerant mass accumulation (case 2) before and after a ramp change in refrigerant mass
flow rate.
126
 Case 1: Same SteadyState Outlet Refrigerant Enthalpy
Figure 4.16 compares the transient responses of the VRLbased model and the FV
based model. As can be observed in subplot (a), the two models have similar settling time,
but there appears a difference in damping characteristics. The VRLbased model behaves
like an overdamped system, while the FVbased model behaves like a lightly damped
system with a delay at the beginning (100  108 seconds). Subplot (b) compares the
transient responses of the refrigerant mass accumulation in the condenser, which
indicates an apparent difference. Such difference is not a surprise. Although the steady
state values of the outlet refrigerant enthalpies were quite matched, the VRLbased model
has lumped volumes and thus cannot predict the phase conditions of the refrigerant in
different locations along the flow direction. In comparison, the refrigerant mass
accumulation in the FVbased method is calculated in each control volume and then
summed up over all the control volumes, which takes into account the effect of phase
conditions in calculating the local mass accumulation. Thus, identical steadystate results
cannot be obtained for the outlet enthalpy and the mass accumulation simultaneously.
This dilemma can be again demonstrated in case 2 shown below, where reinforcing the
same steadystate refrigerant mass accumulation leads to an apparent difference in the
outlet refrigerant enthalpy.
127
(a) Outlet refrigerant specific enthalpy
(b) Refrigerant mass accumulation
Figure 4.16: Comparison between the VRLbased model and the FVbased model (case 1)
 Case 2: Same SteadyState Refrigerant Mass Accumulation
In this simulation case, the steadystate values of the refrigerant mass accumulation
were tuned to the same, as shown in Figure 4.17 (a). Figure 4.17 (b) shows the
comparison of outlet refrigerant enthalpy. The two models can both predict an increase in
the outlet refrigerant enthalpy and have similar settling time, but there is a difference in
the steadystate predictions, as shown in Figure 4.17 (b).
128
(a) Refrigerant mass accumulation
(b) Outlet refrigerant specific enthalpy
Figure 4.17: Comparison between the VRLbased model and the FVbased model (case 2)
Computation efficiency is a crucial performance metric to evaluate dynamic models,
especially for the needs of realtime simulation of Modelica models. As described earlier,
the FVbased model is considered as detailed and thus quite complicated compared to
the VRLbased model, which is built based on lumped formulation. It is thus necessary to
129
evaluate of the possible improvement in computation speed by using the VRLbased
model.
Before comparing the computation efficiency of the two models, the respective model
complexity was first evaluated. For simulation of Modelica models, Dymola would
symbolically translates the original models into two parts before timestep integration:
one for the initialization problem and the other for timestep integration. Table 4.2 and
Table 4.3 compare the model sizes for the initialization problem and the timestep
integration after symbolic translations in Dymola, respectively. Note that the original
model size before symbolic translations are usually much larger, and the order reduction
is significant. As a side note, unlike Dymola, many other simulation tools, such as
TRNSYS, are not capable of performing symbolic transformation and automatic
differentiations to reduce the model size in terms of the dimensionality of linear and/or
nonlinear system of equations [214].
Table 4.2: Comparison of model size of heat exchangers for initialization problem
Model Size
9
for
Initialization
Model Type
FVBased VRLBased
Size of nls
10
361 7
Size after Manipulation of nls 15 1
Numerical Jacobians 1 1
9
Refer to the model size after symbolic translations in Dymola
10
Refer to nonlinear system of equations
130
Table 4.3: Comparison of model size of heat exchangers for timestep integration
Model Size for Time
Step Integration
Model Type
FVBased VRLBased
State Variables 60 3
Time Varying
Variables
2414 256
Size of nls
{16, 16, 16, 16, 11, 11, 11, 11, 11,
11, 11, 11, 11, 11, 11, 11, 16, 16,
16, 16, 16, 16, 16, 16, 16, 16, 16,
11, 11, 11, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6}
{3, 4, 6, 6, 6, 6, 6, 6, 6}
Size after Manipulation
of nls
{1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1}
{1, 1, 1, 1, 1, 1, 1, 1, 1}
Numerical Jacobians 50 7
For case 1, the computation time of the two modeling schemes is then compared in
Figure 4.18 via a dot chart [215]. The integration algorithm selected is DASSL
(Differential Algebraic System Solver). Brenan et al. [216] states that the DASSL method
applies a variablestepsize, variableorder, fixedleadingcoefficient [217]
implementation of the backward differentiation formula (BDF) to compute the solution
trajectory. The computation time was calculated by running the dynamic models on a
Dell desktop computer equipped with Intel(R) Core(TM) 2 Duo CPU E6850 @ 3.00 GHz,
3.00 GB RAM, and running on a 32bit Windows 7. The total simulation time is 200
seconds. Figure 4.18 show the significant advantage of the VRLbased model in
computational efficiency, which is 152 times faster than the FVbased model.
Nevertheless, the computation speeds of both models are much faster than the realtime.
131
Figure 4.18: Comparison of computation time of VRLbased model with FVbased model
4.3.2 Compare Chiller Simulations
For simulation study of the whole chiller, a first attempt was made to duplicate the
same chiller model described in Bendapudi [118]. However, there was an issue when
testing the individual TXV model based on the same model equations and parameters
provided in Bendapudi [118]. The TXV model implemented in Modelica produces a
much different result in mass flow rate. Major issues in matching the results are
summarized as follows.
In Bendapudi [118], it was stated that the refrigerant in the sensing bulb is R500,
which is a refrigerant mixture of R12 and R152a. Unlike pure refrigerant, which has a
unique saturation pressure, refrigerant mixtures such as R500 have different saturation
pressures at the bubble point and the dew point in a ph diagram [212]. Selecting different
saturation pressures would produce different results in computing the bulb pressure in Eq.
(4.39) and hence the mass flow rate. However, such details are not included in [118]. In
addition, it was stated in [118] that it is possible that the sensing bulb was not responding
to the superheat temperature sensed at the compressor inlet because it is mounted in a
well that extends into the shell and is surrounded by a heatconducting paste. However,
there is no further justification on how to set the actual sensed superheat from the bulb.
132
The equations in [118] still use the superheat temperature at compressor inlet, which is
consistent with common HVAC literature. Note that the TXV model was tuned based on
trialanderror method in [118], but the procedure for tuning the valve parameters, along
with the aforementioned details about saturation pressure and superheat temperature, are
not described. The effort in this study was then turned to implement an orifice plate
together with the compressor and heat exchanger models in Bendapudis work [118]. The
resulting chiller model, named as the modified Bendapudi chiller
11
, is considered as a
relatively simpler model due to its performancemap based compressor model. The model
developed in this dissertation research is deemed as a relatively more complex model,
with a detailed physicsbased compressor model while keeping the other three
components the same as the MBC. Note that the complexlevel model has initialization
issue and a proposed initialization method will be presented in the next Chapter.
Simulation study was conducted to compare the modified Bendapudi chiller with the
detailed chiller. The main objective of this comparison is to evaluate the transient
behavior under the same ramp changes, the robustness for initialization problems and the
computation speed. Since the compressor model developed in Bendapudi [118] is semi
empirical, only limited information is available for the compressor geometry. The
compressor model in the detailed chiller is physicsbased and detailed compressor
geometry is needed for the simulation study. In addition, some of the geometry
parameters are proprietary which may not be available from the manufacturer. For these
reasons, the compressor parameters were adopted from Gravdahl and Egelands work
[194] (see Appendix B). Compared to the constantspeed compressor model in
11
For simplicity, the modified Bendapudi chiller refers the same chiller model from Bendapudis work except that the TXV
model is replaced by the orifice plate model.
133
Bendapudis work, the detailed compressor has wider range of operating conditions since
it has two degrees of freedom (speed and IGV) in tuning the compressor capacity. By
comparison, the capacity of the compressor in [118] is limited by its maximum capacity
map (see Eq. (4.3)). Thus, although the two compressors have different geometries, it
would be still possible to compare the two chiller models by tuning the operating range of
the detailed compressor close to that of the performancemap based compressor.
To overcome the aforementioned odds, a parametric study was first conducted to tune
the compressor mass flow rate of the detailed compressor in order to match the same
chilledwater temperature with the modified Bendapudi chiller model under a ramp
change in its IGV position. Note that the modified Bendapudi chiller has a constantspeed
compressor with IGV as the only tuning mechanism. To be consistent, the speed of the
detailed compressor is also set to a constant in the comparison. In addition, to be
consistent with the linear IGV tuning mechanism in the modified Bendapudi chiller, the
linear IGV model aforementioned was adopted for the detailed chiller. The parameter
Linear
R was set to 310
5
Pas/kg. Figure 4.19 and Figure 4.20 show the Modelica model of
the modified Bendapudi chiller and the detailed chiller, respectively. Both chillers
adopted R134a as the refrigerant. The Pressure State Element in Figure 4.19 and
Figure 4.20 is used to compute the time derivative of pressures (dp/dt) for each pressure
level, which is TILs implementation of the aforementioned dp/dt assumption. The
notations (1) and (2) represents pressure levels 1 and 2, respectively. The geometric
parameters of the condenser and evaporator model followed Bendapudis work [118] and
are summarized in Appendix B.
134
Figure 4.19: Modelica model of the modified Bendapudi chiller in TIL
Figure 4.20: Modelica model of the detailed chiller in TIL
135
Three comparison cases are described in the following. As mentioned earlier, the
purpose was to tune the two different chillers to the same steadystate conditions in terms
of one variable, and then compare the qualitative behaviors of transient responses of other
variables of interest.
 Case 1: Same SteadyState Compressor Mass Flow Rate
In this case, the transient responses of the two chiller models were evaluated by
applying ramp changes to the IGV positions, and the steadystate results of compressor
mass flow rate before and after the ramp changes in IGV positions were tuned to the
same, as shown in Figure 4.21. Subplot (a) shows the ramp changes applied to the IGV
positions in the modified Bendapudi chiller and the detailed chiller, respectively.
Subplots (b) through (k) compare the compressor mass flow rate, the valve mass flow
rate, the water temperature at the condenser outlet, the chilledwater temperature, the
compressor inlet pressure, the compressor outlet pressure, the refrigerantside heat
transfer rates at the condenser and evaporator, the refrigerant mass accumulation at the
condenser and evaporator, respectively.
136
(a) Inlet guide vane (IGV) position
(b) Refrigerant mass flow rate through compressor
300 350 400
80
84
88
92
96
100
270
I
n
l
e
t
G
u
i
d
e
V
a
n
e
(
I
G
V
)
P
o
s
i
t
i
o
n
(
%
)
Time (sec)
Modified Bendapudi Chiller
Detailed Chiller
300 350 400
2
.
15
2.20
2.25
2.30
2.35
2.40
2.45
2.50
270
Time (sec)
C
o
m
p
r
e
s
s
o
r
M
a
s
s
F
l
o
w
R
a
t
e
(
k
g
/
s
)
Modified Bendapudi Chiller
Detailed Chiller
137
(c) Refrigerant mass flow rate through expansion valve
(d) Water temperature at condenser outlet
300 350 400
2.20
2.24
2.28
2.32
2.36
2.40
2.44
2.48
Time (sec)
V
a
l
v
e
M
a
s
s
F
l
o
w
R
a
t
e
(
k
g
/
s
)
270
Modified Bendapudi Chiller
Detailed Chiller
300 350 400
35.0
35.5
36.0
36.5
37.0
37.5
38.0
38.5
C
o
n
d
e
n
s
e
r
W
a
t
e
r
O
u
t
l
e
t
T
e
m
p
e
r
a
t
u
r
e
(
C
)
Modified Bendapudi Chiller
Detailed Chiller
270
Time (sec)
138
(e) Chilledwater temperature (evaporator outlet)
(f) Pressure at compressor inlet
300 350 400
9.0
9.2
9.4
9.6
9.8
10.0
10.2
Modified Bendapudi Chiller
Detailed Chiller
270
Time (sec)
C
h
i
l
l
e
d

W
a
t
e
r
T
e
m
p
e
r
a
t
u
r
e
(
C
)
300 350 400
2.0E5
2.4E5
2.8E5
3.2E5
3.6E5
4.0E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
P
r
e
s
s
u
r
e
a
t
C
o
m
p
r
e
s
s
o
r
I
n
l
e
t
(
P
a
)
270
139
(g) Pressure at compressor outlet
(h) Refrigerantside heat transfer rate at condenser
300 350 400
8.88E5
8.92E5
8.96E5
9.00E5
9.04E5
9.08E5
9.12E5
9.16E5
P
r
e
s
s
u
r
e
a
t
C
o
m
p
r
e
s
s
o
r
O
u
t
l
e
t
(
P
a
)
Time (sec)
270
Modified Bendapudi Chiller
Detailed Chiller
300 350 400
5.8E5
5.6E5
5.4E5
5.2E5
5.0E5
4.8E5
4.6E5
4.4E5
4.2E5
4.0E5
3.8E5
3.6E5
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
H
e
a
t
T
r
a
n
s
f
e
r
R
a
t
e
a
t
C
o
n
d
e
n
s
e
r
(
W
)
Modified Bendapudi Chiller
Detailed Chiller
270
140
(i) Refrigerantside heat transfer rate at evaporator
(j) Refrigerant mass accumulation at condenser
300 350 400
3.1E5
3.2E5
3.3E5
3.4E5
3.5E5
3.6E5
3.7E5
3.8E5
3.9E5
4.0E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
H
e
a
t
T
r
a
n
s
f
e
r
R
a
t
e
a
t
E
v
a
p
o
r
a
t
o
r
(
W
)
270
300 350 400
13
14
15
16
17
18
19
20
21
22
23
Modified Bendapudi Chiller
Detailed Chiller
270
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
M
a
s
s
A
c
c
u
m
u
l
a
t
i
o
n
a
t
C
o
n
d
e
n
s
e
r
(
k
g
)
141
(k) Refrigerant mass accumulation at evaporator
Figure 4.21: Comparisons of modified Bendapudi chiller and the detailed chiller (case 1)
 Case 2: Same SteadyState Water Temperature at Condenser Outlet
Figure 4.22 shows the second comparison case. The ramp change of the IGV
positions in case 1 is repeated, while the steadystate water temperature at the condenser
outlet before and after the ramp change in IGV positions were tuned to the same. Subplot
(a) shows the ramp changes applied to the IGV positions in the modified Bendapudi
chiller and the detailed chiller, respectively. Subplots (b) through (k) show the
comparisons of the compressor mass flow rate, the valve mass flow rate, the water
temperature at the condenser outlet, the chilledwater temperature, the compressor inlet
pressure, the compressor outlet pressure, the refrigerantside heat transfer rate at the
condenser and evaporator, the refrigerant mass accumulation at the condenser and
evaporator, respectively.
300 350 400
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
6.0
6.5
7.0
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
M
a
s
s
A
c
c
u
m
u
l
a
t
i
o
n
a
t
E
v
a
p
o
r
a
t
o
r
(
k
g
)
270
Modified Bendapudi Chiller
Detailed Chiller
142
(a) Inlet guide vane (IGV) position
(b) Refrigerant mass flow rate through compressor
280 300 320 340 360 380 400
55
60
65
70
75
80
85
90
95
100
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
I
n
l
e
t
G
u
i
d
e
V
a
n
e
(
I
G
V
)
P
o
s
i
t
i
o
n
(
%
)
280 300 320 340 360 380 400
1.6
1.7
1.8
1.9
2.0
2.1
2.2
2.3
2.4
2.5
2.6
Time (sec)
C
o
m
p
r
e
s
s
o
r
M
a
s
s
F
l
o
w
R
a
t
e
(
k
g
/
s
)
Modified Bendapudi Chiller
Detailed Chiller
143
(c) Refrigerant mass flow rate through expansion valve
(d) Water temperature at condenser outlet
280 300 320 340 360 380 400
1.6
1.7
1.8
1.9
2.0
2.1
2.2
2.3
2.4
2.5
2.6
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
V
a
l
v
e
M
a
s
s
F
l
o
w
R
a
t
e
(
k
g
/
s
)
300 350 400
35.5
35.6
35.7
35.8
35.9
36.0
36.1
36.2
36.3
Modified Bendapudi Chiller
Detailed Chiller
C
o
n
d
e
n
s
e
r
W
a
t
e
r
O
u
t
l
e
t
T
e
m
p
e
r
a
t
u
r
e
(
C
)
Time (sec)
144
(e) Chilledwater temperature (evaporator outlet)
(f) Pressure at compressor inlet
280 300 320 340 360 380 400
9.2
9.6
10.0
10.4
10.8
11.2
C
h
i
l
l
e
d

W
a
t
e
r
T
e
m
p
e
r
a
t
u
r
e
(
C
)
Time (sec)
Modified Bendapudi Chiller
Detailed Chiller
280 300 320 340 360 380 400
3.7E5
3.8E5
3.9E5
4.0E5
4.1E5
4.2E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
P
r
e
s
s
u
r
e
a
t
C
o
m
p
r
e
s
s
o
r
I
n
l
e
t
(
P
a
)
145
(g) Pressure at compressor outlet
(h) Refrigerantside heat transfer rate at condenser
280 300 320 340 360 380 400
8.65E5
8.70E5
8.75E5
8.80E5
8.85E5
8.90E5
8.95E5
9.00E5
9.05E5
9.10E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
P
r
e
s
s
u
r
e
a
t
C
o
m
p
r
e
s
s
o
r
O
u
t
l
e
t
(
P
a
)
280 300 320 340 360 380 400
4.5E5
4.4E5
4.3E5
4.2E5
4.1E5
4.0E5
3.9E5
3.8E5
3.7E5
3.6E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
H
e
a
t
T
r
a
n
s
f
e
r
R
a
t
e
a
t
C
o
n
d
e
n
s
e
r
(
W
)
146
(i) Refrigerantside heat transfer rate at evaporator
(j) Refrigerant mass accumulation at condenser
280 300 320 340 360 380 400
2.6E5
2.8E5
3.0E5
3.2E5
3.4E5
3.6E5
3.8E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
H
e
a
t
T
r
a
n
s
f
e
r
R
a
t
e
a
t
E
v
a
p
o
r
a
t
o
r
(
W
)
280 300 320 340 360 380 400
14.0
14.5
15.0
15.5
16.0
16.5
17.0
17.5
18.0
18.5
19.0
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
M
a
s
s
A
c
c
u
m
u
l
a
t
i
o
n
a
t
C
o
n
d
e
n
s
e
r
(
k
g
)
147
(k) Refrigerant mass accumulation at evaporator
Figure 4.22: Comparisons of modified Bendapudi chiller and the detailed chiller (case 2)
 Case 3: Same SteadyState ChilledWater Temperature
In case 3, the steadystate results of chilledwater temperature before and after the
ramp changes in IGV positions were tuned to the same, as shown in Figure 4.23. Subplot
(a) shows the ramp changes applied to the IGV positions in the modified Bendapudi
chiller and the detailed chiller, respectively. Subplots (b) through (k) show the
comparisons of the compressor mass flow rate, the valve mass flow rate, the water
temperature at the condenser outlet, the chilledwater temperature, the compressor inlet
pressure, the compressor outlet pressure, the refrigerantside heat transfer rate at the
condenser and evaporator, the refrigerant mass accumulation at the condenser and
evaporator, respectively.
280 300 320 340 360 380 400
5.0
5.2
5.4
5.6
5.8
6.0
6.2
6.4
6.6
6.8
7.0
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
R
e
f
r
i
g
e
r
a
n
t
M
a
s
s
A
c
c
u
m
u
l
a
t
i
o
n
a
t
E
v
a
p
o
r
a
t
o
r
(
k
g
)
148
(a) Inlet guide vane (IGV) position
(b) Refrigerant mass flow rate through compressor
280 300 320 340 360 380 400
65
70
75
80
85
90
95
100
Modified Bendapudi Chiller
Detailed Chiller
I
n
l
e
t
G
u
i
d
e
V
a
n
e
(
I
G
V
)
P
o
s
i
t
i
o
n
(
%
)
Time (sec)
280 300 320 340 360 380 400
1.9
2.0
2.1
2.2
2.3
2.4
2.5
2.6
C
o
m
p
r
e
s
s
o
r
M
a
s
s
F
l
o
w
R
a
t
e
(
k
g
/
s
)
Time (sec)
Modified Bendapudi Chiller
Detailed Chiller
149
(c) Refrigerant mass flow rate through expansion valve
(d) Water temperature at condenser outlet
V
a
l
v
e
M
a
s
s
F
l
o
w
R
a
t
e
(
k
g
/
s
)
280 300 320 340 360 380 400
1.9
2.0
2.1
2.2
2.3
2.4
2.5
2.6
Time (sec)
Modified Bendapudi Chiller
Detailed Chiller
280 300 320 340 360 380 400
35.0
35.5
36.0
36.5
37.0
37.5
38.0
C
o
n
d
e
n
s
e
r
W
a
t
e
r
O
u
t
l
e
t
T
e
m
p
e
r
a
t
u
r
e
(
C
)
Time (sec)
Modified Bendapudi Chiller
Detailed Chiller
150
(e) Chilledwater temperature (evaporator outlet)
(f) Pressure at compressor inlet
280 300 320 340 360 380 400
9.1
9.2
9.3
9.4
9.5
9.6
9.7
9.8
9.9
10.0
10.1
10.2
Modified Bendapudi
Chiller
Detailed Chiller
Time (sec)
C
h
i
l
l
e
d

W
a
t
e
r
T
e
m
p
e
r
a
t
u
r
e
(
C
)
280 300 320 340 360 380 400
2.0E5
2.4E5
2.8E5
3.2E5
3.6E5
4.0E5
4.4E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
P
r
e
s
s
u
r
e
a
t
C
o
m
p
r
e
s
s
o
r
I
n
l
e
t
(
P
a
)
151
(g) Pressure at compressor outlet
(h) Refrigerantside heat transfer rate at condenser
280 300 320 340 360 380 400
8.84E5
8.88E5
8.92E5
8.96E5
9.00E5
9.04E5
9.08E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
P
r
e
s
s
u
r
e
a
t
C
o
m
p
r
e
s
s
o
r
O
u
t
l
e
t
(
P
a
)
R
e
f
r
i
g
e
r
a
n
t
H
e
a
t
T
r
a
n
s
f
e
r
R
a
t
e
a
t
C
o
n
d
e
n
s
e
r
(
W
)
280 300 320 340 360 380
5.6E5
5.4E5
5.2E5
5.0E5
4.8E5
4.6E5
4.4E5
4.2E5
4.0E5
3.8E5
3.6E5
3.4E5
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
152
(i) Refrigerantside heat transfer rate at evaporator
(j) Refrigerant mass accumulation at condenser
300 325 350
2.9E5
3.0E5
3.1E5
3.2E5
3.3E5
3.4E5
3.5E5
3.6E5
3.7E5
3.8E5
3.9E5
4.0E5
R
e
f
r
i
g
e
r
a
n
t
H
e
a
t
T
r
a
n
s
f
e
r
R
a
t
e
a
t
E
v
a
p
o
r
a
t
o
r
(
W
)
Modified Bendapudi Chiller
Detailed Chiller
Time (sec)
280 300 320 340 360 380 400
13
14
15
16
17
18
19
20
21
22
Modified Bendapudi Chiller
Detailed Chiller
R
e
f
r
i
g
e
r
a
n
t
M
a
s
s
A
c
c
u
m
u
l
a
t
i
o
n
a
t
C
o
n
d
e
n
s
e
r
(
k
g
)
Time (sec)
153
(k) Refrigerant mass accumulation at evaporator
Figure 4.23: Comparisons of modified Bendapudi chiller and the detailed chiller (case 3)
Table 4.4 and Table 4.5 show the comparisons of model size for the initialization
problem and model size for the timestep integration after symbolic translations in
Dymola, respectively.
Table 4.4: Comparison of model size of chillers for initialization problem
Model Size for
Initialization
Model Type
Modified Bendapudi Chiller Detailed Chiller
Size of nls {376, 361} {1, 361, 376, 66}
Size after Manipulation of nls {30, 15} {1, 15, 30, 4}
No. of Numerical Jacobians 2 3
280 300 320 340 360 380 400
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
6.0
6.5
7.0
R
e
f
r
i
g
e
r
a
n
t
M
a
s
s
A
c
c
u
m
u
l
a
t
i
o
n
a
t
E
v
a
p
o
r
a
t
o
r
(
k
g
)
Time (sec)
Modified Bendapudi Chiller
Detailed Chiller
154
Table 4.5: Comparison of model size of chillers for timestep integration
Model Size for
TimeStep
Integration
Model Type
Modified Bendapudi Chiller Detailed Chiller
State Variables 137 138
Time Varying
Variables
5372 5412
Size of nls
{15, 10, 224, 10, 239, 12, 12, 12,
12, 12, 12, 12, 12, 12, 12, 12, 12,
12, 12, 12, 11, 11, 11, 11, 11, 11,
11, 11, 11, 11, 11, 11, 11, 11, 11, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6}
{15, 57, 239, 11, 11, 11, 11, 11, 11,
11, 11, 11, 11, 11, 11, 11, 224, 12,
12, 12, 12, 12, 12, 12, 12, 12, 12,
12,12, 12, 12, 12, 11, 11, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 1}
Size after
Manipulation of nls
{2, 1, 45, 1, 15, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1}
{2, 3, 15, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 45, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1}
No. of Numerical
Jacobians
103 102
Similar to the comparison made in Figure 4.18, Figure 4.24 shows the comparison of
commutation time between the modified Bendapudi chiller and detailed chiller in case 3.
The total simulation time is 1000 seconds. Due to the adoption of the performancemap
based compressor, the modified Bendapudi chiller has slight speed advantage over the
detailed chiller. It was also found that the modified Bendapudi chiller is more robust in
terms of initialization. However, the computation time of the detailed chiller is only
slightly behind and it is about 20 times faster than the realtime.
155
Figure 4.24: Comparison of computation time between the modified Bendapudi chiller
and detailed chiller
4.4 Summary
In this chapter, an addon Modelica package TIL.Chillers is developed based on TIL.
The package TIL.Chillers includes simpletocomplex level component models to
facilitate model reuse in system level modeling and simulation. For centrifugal
compressor models, both a performancemap based and a detailed physicsbased model
have been implemented for comparison purpose. The heat exchanger models are
implemented based on finite volume method for the detailed model and variable
refrigerant level method for the simplified model. Finally, both orifice plate and thermal
expansion valve models are implemented for the throttling process.
Two sets of simulation studies have been conducted to compare: 1) VRLbased heat
exchanger model with the FVbased model, and 2) the modified Bendapudi chiller model
with the detailed chiller model. In each comparison, the complexities of the dynamic
models are evaluated based on the sizes of nonlinear system of equations for the
initialization problem and the timeintegration problem, respectively. In addition, the
transient responses of the dynamic models are evaluated and compared based on realistic
scenarios. Finally, the computation time of each model is also evaluated and compared.
The VRLbased model has great advantage over FVbased model in terms of the
156
computation efficiency. However, by lumping the control volumes, local thermal
physical properties such as heat transfer and mass storage cannot be predicted accurately
through the flow path of the heat exchanger. For the whole chiller simulation, the
modified chiller model from Bendapudis work is compared with the detailed chiller with
three simulation cases. The detailed chiller was found to be more flexible to adapt to
different simulation scenarios once the relevant physical and geometric parameters are
available. The modified Bendapudi chiller has slight speed advantage and it does not
have initialization problems; while the detailed chiller model cannot pass the
initialization stage without special treatment. To overcome this problem, this dissertation
research developed an initialization scheme which will be presented in the next chapter.
157
Chapter 5 Consistent Initialization for Dynamic Simulation of
Centrifugal Chillers
The previous chapter presents the modeling details of each individual component in a
centrifugal chiller. Challenge in initialization arises for dynamic simulation of the whole
system when the individual components are connected. This chapter presents a method of
obtaining consistent initial conditions to deal with such difficulty. Section 5.1 describes
in more detail the difficulty of consistent initialization for the dynamic simulation of the
whole system of centrifugal chiller, followed by a comprehensive literature review of the
subject of consistent initialization in Section 5.2. Section 5.3 presents the direct method
for consistent initialization of the chiller model developed in Chapter 4. To adapt the
proposed direct initialization method to the specific case of centrifugal chiller, Section
5.4 presents a threestep preprocessing scheme to obtain suitable initial guesses for the
chiller model. Finally, Section 5.5 shows the simulation results of the proposed
initialization method.
5.1 Introduction
For numerical simulation of thermofluid systems which often include sets of
differential algebraic equations (DAEs), consistent initialization is often difficult to
obtain [1]. The procedure for solving DAE systems typically includes two steps. The first
step is to search and obtain a consistent set of initial conditions, and the second step is to
compute the solution trajectory through integration. However, the algebraic constraints in
the DAEs require initial conditions to be consistent with the differential equations, which
may be quite difficult to be satisfied in the initialization phase. In addition, for dynamic
158
simulations of large and complex systems, computation time is usually dominated by the
duration of the initialization phase [2], which may significantly limit the computational
efficiency of real time simulations. In this dissertation study, the consistent initialization
problem is investigated for the detailed centrifugal chiller model developed in Chapter 4.
Such chiller system is considered quite complicated as it forms via concatenation of
several components, i.e., the detailed centrifugal compressor, condenser, expansion valve,
and evaporator.
As mentioned in earlier chapters, the recent work by Bendapudi [118] on a dynamic
model for constantspeed watercooled centrifugal chiller showed the simulation results
agreeing with the steadystate data very well. However, there appeared some problems
with initialization. According to Bendapudi, the model failed to converge for certain
combinations of initial conditions during the earlier part (<150s) of the startup process.
This problem was overcome by adopting a linear ramp function for the condenser inlet
water temperature to start from an initial value to a desired value during the first 150
seconds. In addition, the time consumed by the initialization process was significant. For
a simulation case as shown in Figure 5.1, the initialization process (i.e. the numerical
transient) takes up about 300 seconds of the total simulation time of 1000 seconds.
Therefore, improving the initialization process can significantly enhance the
computational efficiency of the chiller simulation. In addition, for the detailed centrifugal
chiller model developed by this dissertation study, which adopts a variable speed
dynamic compressor with geometrybased characteristics, the ramp function method
failed in the initialization test. Thus, a more reliable approach is needed for the
initialization of centrifugal chiller system.
159
Figure 5.1 Total refrigerant mass charge at chiller startup in faultfree condition,
simulation of example 1 in Bendapudi [118].
Recently, technical challenges of modeling and simulating Modelica based DAE
systems have been reviewed by Wetter [19]. Wetter states that the most difficult issue is
to find consistent initial conditions for solving the hybrid differential algebraic equation
systems. Wetter said, Finding proper settings can in some cases require trial and error
even for experienced users. Therefore, developing reliable procedure for generating
good initial guesses has become a critical need for further promoting the acceptance of
equationbased, componentoriented modeling, especially for thermofluid systems.
5.2 Brief Review on DAE Consistent Initialization
Many HVAC systems, like many other thermofluid systems, can be naturally
modeled as Differential Algebraic Equations (DAE). A general form of DAE can be
represented as [218]:
160
( , , , ) 0 f x x y t =
(5.1)
where , ,
n m
t R x R y R e e e , and :
n n m n m
f D R R R R R
+
_
is a vector field. x and
y are typically classified as differential variables and algebraic variables, respectively.
The solution to the DAE system in Eq. (5.1) is composed of two steps. The first step is to
search for consistent initial values and the second step is the computation of a trajectory.
A set of initial conditions
0 0 0
( , , ) x x y is said to be consistent if the following DAE is
fulfilled at initial time
0
t [218].
0 0 0 0
( , , , ) 0 f x x y t = . (5.2)
Note that for the initialization problem in Eq. (5.2), ( ) f has n + m equations but there
are 2 n m + unknowns in
0 0 0
( , , ) x x y . If the Jacobian denoted by
0
( ) J t , i.e.
0
( )
f f
J t
x y
( c c
=
(
c c
(5.3)
is nonsingular at initial time
0
t , then there are exactly n additional unknowns that need to
be determined initially.
The subject of consistent initialization of DAEs has been widely studied in the area of
chemical engineering and applied mathematics. For applications, such problems are
generally case dependent. Hence a single generic solving procedure cannot be obtained
for various cases. To the authors best knowledge, this topic has not been well addressed
in the open literature of HVAC simulation and control. This section presents a brief
review on the subject.
161
In general, there are two main kinds of methods for consistent initialization, i.e., the
rigorous initialization techniques and the direct initialization methods [219]. The rigorous
initialization techniques typically require knowledge of the specific structure of the DAE
system. In addition, these methods generally need the solutions to a set of nonlinear
algebraic equations.
Several rigorous initialization techniques have been studied. Probably the simplest
method is the socalled initial Euler step method. According to Sincovec et al. [220], a
linear DAE system with constant coefficients can be decomposed into subsystems in
canonical form. The structure of the linear DAE system is characterized by a parameter m
known as the nilpotency of the system. Then, the canonical subsystem can be integrated
with an implicit Euler method initially with arbitrary initial conditions for at least m steps.
Finally, consistent initial conditions can be computed by using a backtracing function to
integrate backwards. However, the stability property of such method was not reported. In
addition, the authors proposed to only control the errors from the state components of the
solutions and remove the error control for the nonstate components. This approach was
not recommended to use by Petzold [221] and Krner et al. [222] since some parts of the
solution may produce large errors with the change of step size. Albet et al. [223]
proposed a modified Euler step method to calculate the state variables after sudden
operational changes in the distillation columns. Two Euler steps were used to calculate
algebraic variables and their derivatives separately without any error control. For the
disadvantage of Euler step methods, Vieira and Biscaia [224] said, The major drawback
of the Euler step method is that it makes no distinction between variables that are already
initialized and variables that must be determined. Considerable jumps will probably be
162
observed in some of the state variables, and the user cannot control which variables are to
be held constant.
The consistency equations methods are another class of rigorous initialization
methods. Such methods typically require differentiation of all or some equations of the
original DAE system to obtain hidden constraints, together with some userspecified
information about the initial conditions, to formulate an extended set of DAEs which are
called consistency equations. These equations would impose consistent requirements
on the initial values and the solution derivatives, which leads to a system of nonlinear
equations involving derivatives of the original problem. The objective is then to find
solution to the consistency equations in the initialization phase. The major limitation of
this class of methods lies in that sometimes the numerical solution to the consistency
equations may not converge, or even do not exist.
Leimkuhler et al. [225] proposed an approximation method for the solution to the
consistency equations. In this method, the derivatives in the consistency equations are
approximated via a singlesided finite difference scheme. The approximated consistency
equations can be solved by the method of leastsquares.
Instead of striving to find a solution to the consistency equations, alternative
approaches have been proposed based on optimization methods. Such approaches consist
of constructing a suitable objective function based on the extended set of the original
DAEs, i.e., the consistency equations. The goal is to search for a set of initial conditions
that is close enough to the consistent one. Suitable optimization methods are then applied
to minimize the deviation of the equations and constraints when substituting the
163
unknowns from the consistency equations. For the initialization problem, Gopal and
Biegler [226] proposed to use successive linear programming (SLP) method to solve the
DAE derivative array equations. By analysing the structure of the DAEs, the author
proposed a criterion to determine which variables would be still continuous under the
discontinuity caused by the input functions. Then, the SLP approach is used to find a
consistent set of initial conditions for the solution trajectory. Biscaia and Vieira [227]
proposed their DAE solver for consistent initialization problem based on a hybrid
optimization approach. The authors state that derivativebased optimization approaches
are likely to be trapped at local minima, and a hybrid optimization approach can be a
remedy. In their approach, a genetic algorithm is first applied to search for the
neighbourhood of the solution. Then a deterministic method is performed to further
locate the optimal solution. Gerdts and Bskens [228] proposed an optimization based
method to calculate consistent initial conditions for a special class of nonlinear semi
explicit DAEs up to index 2. The dependency of initial values on other parameters was
studied based on a sensitivity study of the consistent initialization method. Gerdts [229]
later proposed a direct multiple shooting method to numerically solve the optimal control
problems subject to highindex DAEs and state constrains. The method transforms the
optimal control problem to a finitedimension nonlinear programming problem (NLP).
Two methods were proposed to compute consistent initial values at the multiple shooting
nodes during the iterations. The projection method was shown to yield consistent initial
conditions at each iteration step. By comparison, the relaxation approach can only
achieve consistent initial conditions at the optimal solution.
164
The second class of methods work directly on the DAEs, thus known as the direct
initialization methods. Such methods are simple and efficient with less dependence on the
knowledge from the numerical solvers. Based on the simulation problem in staged batch
distillation unit, Cuill and Reklaitis [230] proposed a relaxation procedure to determine
the consistent initial conditions. The relaxation procedure was formulated based on
physical interpretation of the system. The discontinuity in the state variables at the initial
time was approximated by a cubic polynomial transition function. The authors state that
the method of setting the time derivatives of some variables to zero is an ad hoc approach
and it cannot guarantee to work well for all cases. The authors also mention that the
initial state variables are generally nonzero in some physical processes. Krner et al.
[219] presented a detailed analysis of the feasibility and applications of the direct
initialization method. The authors limited the claim from other authors that steady state is
always the consistent condition to semiexplicit index1 DAE systems. Vieira and Biscaia
[224] proposed a direct initialization method by assuming the dynamic response of a
physical system can be approximately treated as discontinuous perturbation to a similar
steadystate system. For implementation, a particular form of the regularization function
was proposed to smooth out the discontinuous perturbation.
5.3 Direct method for consistent initialization of dynamic centrifugal
compressor
In this section, the detailed chiller model described in Chapter 4 will be used in the
study of the proposed consistent initialization scheme. The individual component models
of the detailed chiller have been tested separately with assumed boundary conditions and
each simulation test worked very well. Figure 5.2 shows the simulation results for testing
165
each individual component by applying a 10second ramp change in compressor rotation
speed at 500 seconds.
(a) Response of the mass flow rate through the compressor
(b) Response of the condenser subcooling temperature
166
(c) Response of the mass flow rate through the expansion valve
(d) Response of evaporator superheat temperature
Figure 5.2: Simulation results for testing individual chiller components under a 10
seconds ramp change in compressor rotation speed at 500 seconds
167
5.3.1 Problem Statement
Although the individual chiller component test was successful, when the four
components were connected together, simulation failed to start due to initialization
problem. Figure 5.3 shows the failure message in Dymolas simulation log. The major
problem appears that the mass flow rates for the components should be solved implicitly
based on the equations from each individual component. For such a DAE system,
consistent initial mass flow rate distribution should be sought in order for the numerical
solver to compute its first integration step. To solve the problem, a first effort was made
to adopt the method proposed by Bendapudi [118], i.e., linearly ramping the inlet
condenser water temperature from the beginning to 150 seconds. But the solver still
failed to start.
Figure 5.3: Dymola simulation log of a failed initialization when the four chiller
components were connected.
168
In this dissertation study, the following remedies are proposed with the aim of solving
this problem in a systematic manner. For the initialization of the centrifugal compressor,
a direct initialization method is proposed with three major steps. First, a reasonable
compressor startup speed command is computed from the knowledge of the compressor
characteristic equation. Then, a pseudo physical analogy is established to approximate the
initial mass flow rate solution from the centrifugal compressor model. Second, a suitable
perturbation function is applied to approximate the step function in order to perturb the
compressor mass flow rate to its analytical nonsurge solution. Finally, a model switching
technique is proposed to switch the model equation back to the nonlinear compressor
characteristic equation that includes surge (unstable) solution. The details of each step of
the direct initialization method are given later.
Besides the direct initialization method applied to the compressor mass flow rate
solution, the initial conditions for the thermal variables at various states in the chiller
cycle remain to be determined. To achieve this, a preprocessing scheme with another
three steps is proposed. First, the initial guesses for the compressor discharge states are
obtained based on the chiller design specifications. Second, the initial conditions for the
condenser are determined based on the inequality constraints established from the
analysis of transient mass balance. Finally, the initial conditions for the evaporator are
determined based on the inequality constraints established from the analysis of its
minimum required initial mass inventory to compensate for the transient change of mass
storage in the condenser. The details of each step of the preprocessing scheme are given
in the case study section (see Section 5.4).
169
5.3.2 Computation of Reasonable Compressor Startup Speed Command
For detailed compressor modeling, it is important to note that the models described in
Gravdahl and Egeland [194] and Gravdahl et al. [197] were intended for typical air
compression systems, which are generally operated in openloop with dry air as the
working medium. However, for the chiller system, the working medium is refrigerant that
works in closeloop with multiphase conditions, which complicates the modeling
framework and the associated problem of consistent initialization.
In Chapter 4, the compressor characteristic equation is formulated based on Gravdahl
et al. [197]:
( )
2
1
2 2 2 2 1
2
c,out
c,in in
( )
2
, 1
f
c
p
r
r m k m
p
m
p c T
k
k
e e
e
(
(
= = +
(
(
(
(5.4)
where
1
c,in 1 1
cot
b
Ar

= (5.5)
o = (5.6)
For analysis of the consistent initialization problem, Eq. (5.4) can be analytically
rearranged into
2 2 2 2 2 2 2 2
1 1 1 2
1 1
( ) 0
2 2
f o
k r m r m r r e e e + + + = (5.7)
170
where
( )
1
in
, 1
o c p
m c T
k
k
e
(
=
(
(5.8)
In Eq. (5.7), a nonlinear equation should be solved for the mass flow rate. Fortunately,
this equation is a quadratic equation which can be solved analytically. To get real
solutions, the discriminant () of the quadratic equation should be greater than or equal to
zero. In Eq. (5.7), is given by,
2 2 2 2 2 2 2 2 2
1 1 1 2 0 1
1 1 1
( ) 4 4
2 2 2
f f
r k r r r k r e
(     
A = + +
  
(
\ .\ . \ .
(5.9)
Reinforcing 0 leads to the following constraint for the rotor speed
1
2
2 2
0 1
2 2 2 2 2 2
1 1 1 2
1
4
2
1 1
( ) 4
2 2
f
f
k r
r k r r r
e
(  
+

(
\ .
> (
  
(
+
 
(
\ .\ .
(5.10)
The corresponding solution of mass flow rate at the compressor outlet is then
2 2 2 2 2
1 1 1
2 2
1
1
( ) 4
2
1
2
2
f
f
r r k r
m
k r
e e 
 
+

\ .
=
 
+

\ .
(5.11)
where
2 2 2
1 2 0
1
2
r r  e
(  
= +

(
\ .
(5.12)
171
It is interesting to note that the solution of the mass flow rate is determined by the
sign of . If  < 0, there are one positive and one negative flow rate solutions. It can be
easily seen that the negative solution should be rejected. On the other hand, if  > 0, then
there are two positive solutions with a bigger one and a smaller one. For  = 0, one
solution of the mass flow rate is zero and the other is positive. For a given speed , the
surge mass flow rate could be explicitly calculated based on Eq. (5.4). By equating the
first derivative of
c
with respective to m to zero, the surge points could be determined.
By comparing the two possible solutions with the surge point, the unstable solution could
be detected and rejected. A closer look at Eq. (5.11) reveals that the sign of m is also
related to the rotation speed .
 Mass flow rate m has one positive and one negative solutions for
1
1
2
2
2 2
0 1
0
2 2 2 2 2 2 2 2
1 2 1 1 1 2
1
4
2
1 1 1
( ) 4
2 2 2
f
f
k r
r r r k r r r
e
(  
 
+

(

\ .
< s (

  
(
 +
 
( \ .
\ .\ .
(5.13)
 Mass flow rate m has two positive solutions for
1
1
2
2
2 2
0 1
0
2 2 2 2 2 2 2 2
1 2 1 1 1 2
1
4
2
1 1 1
( ) 4
2 2 2
f
f
k r
r r r k r r r
e
(  
 
+

(

\ .
< s (

  
(
 +
 
( \ .
\ .\ .
(5.14)
172
 Mass flow rate m has two identical positive solutions for
1
2
2 2
0 1
2 2 2 2 2 2
1 1 1 2
1
4
2
1 1
( ) 4
2 2
f
f
k r
r k r r r
e
(  
+

(
\ .
= (
  
(
+
 
(
\ .\ .
(5.15)
The above constraints for provide a useful guideline to set the initial speed
command. In addition, it provides a reasonable physical bound for the iteration range of
, if the initial value of is determined by solving for other initial equations.
5.3.3 PseudoPhysical Analogy for Centrifugal Compressor
One of the difficulties in centrifugal chiller initialization is to find consistent solutions
of the valve and compressor mass flow rates by simultaneously solving the nonlinear
equations. It is then desirable to change the equation structure at the initialization phase.
After the initialization, a suitable smoothing and switching method can be applied to
switch back to the original DAE system. Compared to some rigorous initialization
methods, the above scheme has more benefits in the following sense. In practice, a large
jump reference can be avoided with an appropriate startup controller. Whereas, for
simulation purpose, it is desirable to emulate such extreme behavior as close as possible.
Mathematically speaking, a jump startup command corresponds to a step forcing
function applied to the system. Physically, the change of mass flow rates and pressures
from zero to finite values require a transient period. However, in practice, the system
behaviors during this transient period are very difficult to measure. Therefore, it is very
hard to predict the initial conditions, e.g., the initial derivatives of some state variables.
Also, there are no general equations that are strictly accurate within that short transient
173
period. A remedy is to approximate the model behavior in that period based on a suitable
pseudophysical analogy
12
. It is important to note that such method is also very useful for
other system operations, e.g. a sudden open/close of a valve, or a step change of the
controller command. The simplest pseudophysical analogy can be a constant value, i.e.,
analogy
m C = (5.16)
Such method has been successfully applied to the calculation of heat transfer and
pressure drop correlations in TIL [231]. For a robust initialization, the correlations for
heat transfer and pressure drop can switch from constant values to their actual
correlations at different time instances during the simulation. For chiller simulation, an
alternative form of the pseudophysical analogy is proposed since the pressure rise ratio
and mass flow rate are coupled for the initialization problem. Fortunately, we could use
the compressor characteristic map to obtain a good startup relation in the following form:
analogy initial 0
m k = (5.17)
It is important to note that such a simple form is crucial for the success of the
initialization problem. It is much easier for numerical solvers to handle this linear
relation than the original nonlinear equation. Reducing the dimension of nonlinear system
of equations could greatly facilitate the computation or approximation of the inverse of
the Jacobian of the DAE system by Newtonlike solvers.
12
That is why it is termed as pseudophysical analogy.
174
5.3.4 Perturbation Method
After the pseudophysical analogy is established, the solution of mass flow rate can
be perturbed with the following perturbation function
 
analogy analytical
( ) 1 ( , ) ( , )
c c
m t t t t m t t t m o o = A + A (5.18)
where 0 ( , ) 1
c
t t t o s A s , t is the current simulation time, t
c
is the transition time to
apply the step change, and t is the time to complete the perturbation. The selection of an
appropriate perturbation function ( , )
c
t t t o A is given in a later section with details.
analogy
m is the mass flow rate obtained from the pseudophysical analogy shown in Eq.
(5.17). The selection of
analytical
m could be facilitated by examining the compressor
characteristic map. Depending on the rotation speede, the solutions from Eq. (5.7) could
be one negative and one positive, two positive solutions, or two identical positive
solutions. However, in either case, as shown in Figure 5.4, the solution on the righthand
side of the surge point can be uniquely determined by a single expression, i.e.,
2 2 2 2 2
1 1 1
analytical
2 2
1
1
( ) 4
2
1
2
2
f
f
r r k r
m
k r
e e 
 
+ +

\ .
=
 
+

\ .
(5.19)
175
Figure 5.4 Selection of
analytical
m based on compressor characteristic map
5.3.5 Variable Structure Modeling and Reinitialization
The perturbation method described in the previous section tunes the mass flow rate
solution to a stable one. However, for centrifugal chillers, it is also desirable to simulate
unstable (but realistic) behavior like surge and stall for the study of dynamics and control.
For such purpose, it is necessary to switch back to the nonlinear equations in the
centrifugal compressor model. It is important to note that such a switch involves two
crucial issues related to numerical simulation, i.e., variable structure modeling and
consistent reinitialization after the discontinuity jump introduced by switching the models.
Figure 5.5 presents a motivating example of an unsuccessful switch due to
inconsistent reinitialization. The pseudophysical analogy in Eq. (5.17) was used to find
the consistent initial conditions at very beginning. A transition to the original compressor
176
characteristic equation was applied at 10 seconds. However, the numerical solver seemed
to get stuck at that time instant. Consistent initial conditions could not be obtained for a
reinitialization. The simulation thus failed to proceed.
Figure 5.5: Illustrating example of inconsistent conditions for the reinitialization after
switching to the original compressor equation
Model switching during simulation dynamically changes the structure of the DAE
system. At the moment of switching, the numerical solver has to stop and a new set of
consistent initial conditions should be obtained in order for the solver to continue. In
Dymola, switching from model f
1
() to model f
2
() during simulation is realized by
( )
1 analytical 2 nonlinear
0 1 ( ) ( ) ( ) ( ) t f m t f m = + (5.20)
177
where
2 2 2 2 2
1 1 1
1 analytical
2 2
1
1
( ) 4
2
1
2
2
f
f
r r k r
f m
k r
e e 
 
+ +

\ .
=
 
+

\ .
, (5.21)
2 2 2 2 2 2 2 2
2 1 nonlinear 1 nonlinear 1 2
1 1
( )
2 2
f o
f k r m r m r r e e e = + + + , (5.22)
and
0
( )
1
s
s
t t
t
t t
<
=
>
. (5.23)
5.3.6 Choice of Suitable Perturbation Functions
A variety of functions can be used to approximate the step function o (,) in Eq.
(5.18). This subsection first reviews some candidate functions available in the literature,
and then proposes a simple yet effective function for the chiller problem. To solve the
startup problem in a constantspeed centrifugal chiller with heat exchangers formulated
in the FV methods, Bendapudi [118] adopted a ramp function linearly ramping the
condenser inlet water temperature. For chemical engineering applications such as
condenser model and finite batch simulation, Vieira and Biscaia [224] proposed a class of
smoothing function as
1
0
1
( , , ) 1 exp
!
j
n
c c
c
j
t t t t
t t n n n
j
o t
t t
=
 
 
=



\ .
\ .
(5.24)
178
For function ( , , )
c
t t n o t , the smoothing profile can be regulated via parameters n
and . To evaluate the smoothing quality, a smoothening error is defined as
,
( ) 1 ( , )
s
s
t
E t t n dt o o t
( =
}
(5.25)
Analytical integration showed that this smoothening error is equal to . Parameter n
controls the order of smoothness, while does not affect the smoothening error.
In this study, the following perturbation function is proposed
1
( , ) tanh tan 1
2
c
c
t t
t t t
t
o
   
A = +
 
A
\ . \ .
(5.26)
This function is a specialized hyperbolic tangent function, called SpliceFunction,
available from Modelica Standard Library. The SpliceFunction is typically used to handle
the smooth transition between different correlation functions in order to calculate a single
variable that has different operating ranges. To the authors best knowledge, this function
has not been reported being used as a perturbation function in the initialization problem.
Figure 5.6 compares the unit step function with the proposed function with t of 0.5, 1,
and 1.5 seconds.
179
Figure 5.6: Comparison of unit step function with the proposed perturbation function
with different time durations for transitions
5.3.7 Discussion on the Feasibility of SteadyState Initialization Method
As described in Chapter 4, the condenser and evaporator were modeled based on the
FVbased method with transient mass and energy balance equations. For the refrigerant
side, the state variable pair is pressureenthalpy. Special attention must be paid to the
choice of initialization method for the state variables. This is mainly because the mass
and energy equations are directly coupled to the state variables. Harsh initial transients of
the state variables would affect the initial solution of the mass balance, which would
make it difficult or even impossible for the solver to achieve a consistent set of mass flow
rate distributions.
180
Currently, Modelica offers two concepts of initialization [232]. The first concept is
the socalled steady state initialization, i.e., set the derivatives of the state variables to
zero initially. In many situations, this is the simplest treatment in that there are no initial
transients in system. Unfortunately, a typical airconditioning system is often composed
of different nonlinear models. When these models are connected, nonlinear algebraic
equations may be present and it might be difficult to seek a solution, as shown in the
previous example. One remedy is to assign guess values to the iteration variables.
However, for large and coupled system of equations, good initial guesses are needed in
order for Dymola to find a solution. Another issue is that it is not always possible to set
all the derivatives to zero for a given system. The second concept is to give explicit start
values to the states. This scheme offers the flexibility if the steadystate initialization fails
or the initial conditions could be readily obtained. For example, for simulation of a single
cooling coil, the temperatures (as state variables) can be directly assigned initial values
based on the simulation scenario or measurement. When solving particular problems, it is
also possible to resort to the combination of the two concepts. An example for initializing
thermal fluid system is given in Tummescheit and Eborn [233], the authors proposed a
pseudo steady state approach to initiate the variables with fast dynamics, such as
pressures, in steadystate, and give explicit start values to variables with slow dynamics,
such as temperature, enthalpy and composition.
In our chiller model, it has been found that steadystate initialization at the refrigerant
side does not work well. The reason is that the energy balance equation is simplified to an
algebraic equation relating the inlet mass flow rate, inlet and outlet specific enthalpies,
and the heat transfer rate. Such an equation is very sensitive to the change in heat transfer
181
rate computed. As consequence, the thermal variables like specific enthalpies will change
very quickly with the change in the computed heat transfer rate. This is not physically
correct and is most likely the reason for a failed simulation. A remedy is to assign a linear
distribution to the specific enthalpy or a constant start value of specific enthalpy from
certain offline calculations based on the refrigerant model and design conditions.
5.4 Case Study: Preprocessing Scheme for Dynamic Simulation of
Centrifugal Chillers
The above direct initialization method may well improve the numerical startup of
centrifugal compressor. However, for the overall chiller system initialization, proper
method is needed to obtain suitable initial guesses for the initialization of the dynamic
condenser and evaporator.
In general, the numerical convergence of the state variables in the refrigerantside
models is very sensitive to their initial guesses. As mentioned above, trialanderror has
been the typical practice; however, for a set of variables as in the chiller system, this
method could be extremely tedious and even without a guaranteed success. To overcome
such difficulties, a preprocessing scheme is proposed to determine suitable initial guesses
for all chiller components based on their physical connections.
To illustrate the proposed preprocessing scheme, Figure 5.7 shows a schematic of the
refrigerant cycle in a centrifugal chiller system, where notation ^ for the pressure and
enthalpy variables indicates that they are the initial guesses to be determined for the
chiller initialization. Equations (5.27) through (5.32) summarize the basic equations for
the chiller components used in the analysis of the initialization problem, where the
182
subscripts 1, 2, 3, 4 follow the graphical definition in Figure 5.7. To facilitate later
discussion, the sign convention is not applied to the flow variable itself, but it can be
recognized from the + or sign in the following equations.
Figure 5.7: Schematic of the four state points labeled on the centrifugal chiller system
2
1 initial
1
p
m k
p
= (5.27)
1 2
0 m m = (5.28)
cond
cond 2 3
d
V m m
dt
= (5.29)
183
3 eff 3 3 4
2 ( ) m A p p =
(5.30)
3 4
0 m m = (5.31)
evap
evap 4 1
d
V m m
dt
= (5.32)
A threestep procedure is proposed as follows.
 Step 1: Determine the Initial Guess for Compressor Discharge (State 2)
Since the pressures at the condenser and evaporator sides are normally measured in
practice and the nominal designed flow rate could be obtained from offline analysis, one
way to determine the parameter
initial
k is to solve the following equation:
2
initial
1 1,nom
p
k
p m
=
(5.33)
As for the proposed chiller model, the initial guesses of the pressure rise ratio
( )
2 1
/ p p is first determined based on the designed nominal conditions. Then, the initial
mass flow rate solution can be determined based on the selection of
initial
k as shown in Eq.
(5.33). To the authors experience, the selection of
initial
k is relatively flexible. But if
initial
k
is too large, then the solution of compressor mass flow rate would be too small and the
system trajectory would possibly enter the unstable region which may cause the
initialization to fail. On the other hand, if
initial
k is too small, then the solution of
compressor mass flow rate would be too large and may exceed the physical limit of the
compressor, e.g., the choke limit.
184
Next, the initial guess of specific enthalpy
2
( ) h h p = (5.34)
where
sup,vap
( ) h is the thermodynamic function to compute the specific enthalpy of the
superheat refrigerant vapor at a given pressure [234].
 Step 2: Determine the Initial Conditions for Condenser
As for the initial guesses, only approximate values are needed, thus the condenser
model is treated as a lumped formulation in the following analysis. From the mass
balance, the following inequality constraint should be satisfied during the simulation.
cond
cond 2 3 2
d
V m m m
dt
= < (5.35)
It is desirable to first derive a suitable constraint for the condenser outlet density and then
set up an appropriate outlet specific enthalpy in order to apply the linear enthalpy
distribution for initialization. Integrating Eq. (5.35) over [t
0
, t
1
] yields
1 1
0 0
cond
cond 2
t t
t t
d
V dt m dt
dt
<
} }
(5.36)
Next, it is necessary to estimate the time duration (t) for the refrigerant to flow from
the inlet to the outlet of the condenser, which can be determined based on the geometric
parameters of the condenser and the initial mass flow rate at condenser inlet, i.e.,
2 cond
1 0
2
c
A L
t t t
m
A = =
(5.37)
185
The assumption of lumped formulation implies
cond 0 2
( ) t = and
cond 1 3
( ) t = . Thus,
the integration in Eq. (5.36) yields
2 2 cond
cond 1 3
cond
( )
m t V
t
V
A +
= <
(5.38)
Note that
cond
(t
0
) could be adopted from Step 1. A suitable initial guess
3
h could thus be
obtained by a few trialanderrors based on the offline refrigerant property calculations
with the following guideline.
2 2 cond
2 3 3 3
cond
( , )
m t V
p h
V
A +
s <
(5.39)
 Step 3: Determine Initial Conditions for Evaporator
The initialization of evaporator is important for predicting the overall refrigerant mass
charge. During the very beginning of chiller startup, due to the transient mass storage at
the condenser, the mass flow rate at the inlet of the evaporator is generally smaller than
that of the evaporator outlet, i.e.,
4 1
m m < . For a reasonable startup, the evaporator should
have enough initial mass storage prior to the startup of the chiller. Thus, when the chiller
is turned on, as the mass flow rate command from the centrifugal compressor is greater
than the actual supply from the expansion valve, the initial mass storage in the evaporator
would serve as a backup source to compensate for the additional needs and maintain the
mass balance of the overall refrigerant cycle.
Therefore, for a physically reasonable initialization, an inequality constraint should be
set at the evaporator side to ensure enough initial mass storage prior to the start of the
simulation. To establish such a constraint, it is assumed that the mass flow rate entering
186
the evaporator will rise to a magnitude comparable to that leaving the evaporator by the
end of the period of pseudophysical analogy. Since the centrifugal compressor and
expansion valve are not modeled as mass storage devices, during the chiller startup, the
increase of mass storage at the condenser must equal to the decrease of mass storage at
the evaporator. Substituting Eq. (5.30) into Eq. (5.29), the variation of the transient mass
storage at the condenser can be obtained as
cond cond
cond 2 eff 3 3 4
2 ( )
dM d
V m A p p
dt dt
= = , (5.40)
Integrating Eq. (5.40) yields
( )
2 2
0 0
cond
cond cond 2 eff 3 3 4
2 ( )
t t
t t
d
M V dt m A p p dt
dt
A = =
} }
(5.41)
where
0
t is the start time of simulation, and
2
t refers to the end of the pseudophysical
analogy. As a conservative estimation, it is assumed that the transient mass storage is
minimum at the beginning of the chiller startup process, i.e.,
( ) ( ) ( ) ( ) ( )
( )
( )
cond 2 0 eff 3 0 3 0 4 0 2 0
2 M m t A t p t p t t t A > (5.42)
To ensure enough initial mass storage at the evaporator, the inequality constraint to be
satisfied is
evap cond
M M A > A (5.43)
Note that initially the evaporator should have enough refrigerant mass to compensate for
the change in
evap
M A caused by the compressor startup, so let
evap evap evap
M V A = , where
187
evap
is the initial refrigerant density at the evaporator, we have
cond
evap
evap
M
V
A
> (5.44)
Therefore,
( ) ( ) ( ) ( ) ( )
( )
( )
2 0 eff 3 0 3 0 4 0 2 0
evap
evap
2 m t A t p t p t t t
V
>
(5.45)
A good alternative assumption is to set
4 1
h h = , which corresponds to constant enthalpy
initialization. This assumption is advantageous in avoiding the computation of the mean
value of
evap
based on different enthalpy values to satisfy the above inequality constraint.
Given the initial guess of the refrigerant pressure at evaporator (
4
p ), the task is to find a
suitable initial guess of the refrigerant specific enthalpy
4
h such that
( )
( ) ( ) ( ) ( ) ( )
( )
( )
2 0 eff 3 0 3 0 4 0 2 0
evap 4 4
evap
2
,
m t A t p t p t t t
p h
V
>
(5.46)
Appendix C lists the initial guesses obtained based on the proposed threestep
preprocessing scheme used in the following chiller simulation study.
5.5 Simulation Study
5.5.1 Simulation Results of the Proposed Perturbation Function
Simulation study was conducted to evaluate the effectiveness of the proposed
consistent initialization method on the dynamic centrifugal model. The model adopted
188
R134a as refrigerant. The inlet water flow rates of the condenser and evaporator were
16.7 kg/s and 13.2 kg/s, respectively. The inlet water temperatures of the condenser and
evaporator were 30 C and 16 C, respectively. The dimensions of the compressor model
and the heat exchanger models follow the work by Gravdahl and Egeland [194] and
Bendapudi et al. [199], respectively. The details are summarized in Appendix B.
Figure 5.8 shows the Modelica layout of the detailed centrifugal chiller model
developed in TIL, which includes a detailed centrifugal compressor, a condenser, an
expansion valve, an evaporator, and two pressure state elements (dp/dt assumption).
Figure 5.8: Modelica model layout of the detailed centrifugal chiller model in TIL
189
5.5.2 Discussion on the PseudoPhysical Analogy
Unlike the air compression system described in Gravdahl and Egeland [194], the inlet
conditions for the compressor are not fixed in centrifugal chiller operation. According to
Bendapudi et al. [199], during the first few seconds of startup, it is numerically
(perhaps even physically) possible to have a twophase condition at compressor inlet and
also at the compressor outlet. Since the compressor characteristic equation is derived by
assuming the refrigerant is vapor, it is thus not suitable for the startup process.
As shown in Figure 5.9, when simulation starts, the quality [234] of the inlet
refrigerant, defined as the ratio of the refrigerant vapour mass over the total refrigerant
mixture mass, is about 0.2, which corresponds to twophase conditions. During the time
period of the proposed pseudophysical analogy, the refrigerant quality numerically
converged to the expected situation of refrigerant vapor with the unity quality. The
perturbation method was applied once the compressor characteristic equation became
accurate. Such a transition did not affect the solution profile of the refrigerant quality.
190
Figure 5.9: Refrigerant quality profile at the compressor inlet during chiller initialization
5.5.3 Simulation Results of the Direct Initialization Method on Compressor
Figure 5.10 shows the compressor mass flow rate during the initialization period (0 ~
30 seconds). As previously shown in Figure 5.5, the second transition (9 ~ 20 seconds) is
necessary for the Newton iteration solver to converge when switching back to the original
equation. The corresponding solution profile of the pressure rise at the centrifugal
compressor is given in Figure 5.11.
191
Figure 5.10: Mass flow rate at compressor outlet during chiller initialization
Figure 5.11: Compressor pressure rise ratio during chiller initialization
192
5.5.4 Validation of Mass and Energy Conservation
Due to perturbation method used in this study, it is important to check the mass and
energy conservation during the initialization phase and the steadystate. Figure 5.12
shows the mass accumulations at the condenser and evaporator, respectively, and Figure
5.13 shows the total mass imbalance of the chiller simulation. The total mass imbalance
was evaluated by subtracting the total steadystate mass accumulation from the total
transient mass accumulation during the first 30 seconds. As can be observed in Figure
5.13, the amount of mass imbalance is detectable only by the end of the perturbation
method, i.e., for 0 to 20 seconds. The mass imbalance is within about 0.0004 % of the
total refrigerant mass, which is negligibly small.
Figure 5.12: Refrigerant mass accumulation at condenser and evaporator sides during
chiller initialization
193
Figure 5.13: Total refrigerant mass imbalance during chiller initialization
The refrigerant energy balance is evaluated for the chiller cycle as well, which
requires at any simulation time, the work input to the centrifugal compressor plus the heat
transferred to the evaporator should equal to the heat removed from the condenser. As
shown in Figure 5.14, the refrigerant energy imbalance in the chiller cycle was bounded
by 2.510
9
W for the startup period. For longterm simulation, the energy imbalance
converged to less than 1.810
10
W from 150 seconds to 1000 seconds, which can firmly
be deemed negligible.
194
Figure 5.14: Total energy imbalance during chiller initialization
5.5.5 Computational Efficiency
Computation time is a crucial factor for realtime simulation of HVAC systems. The
proposed initialization method has been tested with two algorithms of numerical
integration, i.e. DASSL (Differential Algebraic System Solver) and Radau IIa. The
DASSL solver is introduced in the previous chapter. The Radau IIa solver is a fully
implicit RungeKutta algorithm. Both methods led to successful numerical convergence
at the initialization phase. Figure 5.15 presents the comparison between DASSL and
Radau IIa for the solution profiles of compressor mass flow rate at the initialization phase.
The two methods yielded similar results for the proposed initialization method. As shown
in the zoomed plot inside Figure 5.15, the solution with the DASSL method yielded
slightly better results in terms of the quality of smoothness.
195
Figure 5.15: Solution profiles of compressor mass flow rates based on the integration
algorithms of DASSL and Radau lla
The CPU time for integration was calculated by running the models on a Dell desktop
computer equipped with Intel(R) Core(TM) 2 Duo CPU E6850 @ 3.00 GHz, 3.00 GB
RAM, and running on a 32bit Windows 7. Figure 5.16 summarizes the computation time
of the proposed chiller model based on the DASSL and Radau IIa methods. The
computation time of the proposed chiller model with DASSL and Radau IIa yield close
result. Using the DASSL method was about 21 times faster than the real time, while the
Radau IIa method was about 26 times faster than the real time. In addition, it was found
that the computation time could be further reduced by loading the steadystate data as the
initial conditions for a new simulation. The new simulation could be started at any time in
the previous saved simulation time interval. For example, it is more beneficial to start the
simulation after the numerical convergence. For the same simulation scenarios as the
196
comparisons in Figure 5.16, the computation time of the proposed chiller model with
DASSL is reduced to about 3.01seconds by loading the initial values from previous
steadystate data, which is about 332 times faster than the real time.
Figure 5.16: Comparison of computation time for the proposed chiller model with the
integration algorithms of DASSL and Radau lla
5.6 Summary
In this chapter, a preprocessing scheme and a direct initialization method are
proposed to deal with the chiller initialization problem. Once the geometric parameters
and design conditions are available, reasonable initial guess values can be computed from
the preprocessing scheme. A direct initialization method can be applied to find consistent
initial conditions for the dynamic simulation of centrifugal chillers. The direct method
consists of three major steps: 1) establish a pseudophysical analogy, 2) apply the
perturbation method, and 3) switch the model structure. Simulation results have
demonstrated the effectiveness of proposed initialization method. The computation time
is about 21 times faster than the real time. If the initial values are loaded from steady
state data, the computation time is about 332 times faster than real time. Such quality and
efficient dynamic model could be later integrated to realtime simulation and test control
strategies.
197
Chapter 6 SelfOptimizing Operation of AirSide Economizer
Using Extremum Seeking Control (ESC)
Based on the 3state control strategy proposed for efficient and reliable operation of
airside economizers in [81], an extremum seeking control (ESC) scheme is designed and
evaluated in this chapter, which can achieve a selfoptimizing tuning of the AHU
dampers for minimizing the mechanical cooling needed. Section 6.1 gives a brief
introduction of the airside economizer. Then, the 3state control strategy [81] will be
briefly described in Section 6.2. Section 6.3 presents the details of the ESC control for
economizer operation, which includes the framework of the economizerbased ESC and
the associated controller design details. Furthermore, to avoid actuator saturation problem,
the antiwindup ESC proposed in [235] is adopted. Finally, in Section 6.4, simulation
study is carried out to validate the proposed ESC control with a highfidelity dynamic
model for the airside economizer developed based on Modelica (see Chapter 3). In the
simulation cases, different initial positions of the outdoorair damper are tested, and the
effectiveness of the antiwindup ESC scheme is also validated.
6.1 Introduction
The airside economizers have been developed as a class of energysaving HVAC
devices that may increase the energy efficiency by taking advantage of outdoor air during
cool or cold weather. However, many economizers do not operate in the expected manner
and waste even more energy than before installation, mostly due to the unreliable sensors
and actuators in practice. Better control strategy is needed for optimal and robust
operation. In this dissertation study, an ESC based selfoptimizing control strategy is
198
developed to minimize the energy consumption, with the feedback of chilled water
supply command rather than the measurements of air temperature and humidity. More
specifically, the mechanical cooling load is minimized by seeking the optimal opening of
the outdoorair damper in real time. This scheme does not need temperature or humidity
sensors, and depends much less on the knowledge of the economizer model. Simulation
was performed on the Modelica based transient model of a singleduct AHU developed in
Chapter 3. The simulation results have demonstrated the potential of using ESC to
achieve the minimal mechanical cooling load in a selfoptimizing manner. In addition,
the antiwindup ESC scheme [235] is implemented to handle the ESC windup issue due
to actuator (damper) saturation. With the antiwindup ESC scheme, optimal control of the
airside economizer can be achieved with smooth transitions among different regions of
operation on the psychrometric chart. The simulation results validated the effectiveness
of the antiwindup ESC scheme.
Figure 6.1 is a schematic of a typical singleduct AHU with controller. The AHU
includes a supply fan, three (outdoor air, relief air and mixed air) dampers for controlling
airflow between the AHU and the outdoors, heating and cooling coils for conditioning the
air, a filter for removing airborne particles, various sensors and actuators, and a controller
that receives sensor measurements (inputs) and computes and transmits new control
signals (outputs). The airside economizer moves the dampers to let in 100% outdoor air
when it is cool but not extremely cold outside. When it is hot outside, the dampers are
controlled to provide the minimum amount of outdoor air required for ventilation.
199
Figure 6.1: Schematic of a singleduct AHU (reproduced with permission from Figure 1
in [60])
6.2 Proposed Economizer Control Strategy
Modeling and optimal control of airhandling units and economizers have been
recently studied [9, 10]. Seem and House [11] described two modelbased strategies for
economizer control and use simulations to investigate the energy performance of the
strategies in comparison to traditional economizer strategies. While the modelbased
strategies achieved modest energy savings over the traditional strategies for perfect
sensors, the performance of all the strategies suffered when sensor errors were introduced.
In this study, an online selfoptimizing control approach is described for air economizers.
This approach is considered more robust than other modelbased approaches in part
because it does not require sensor measurement to achieve optimal control. This research
200
investigates the application of the extremum seeking control (ESC) [84, 86, 87, 236, 237]
to optimize the use of outdoor air so as to minimize the energy consumption. The input
and output of the proposed ESC framework are the damper opening and power
consumption (or equivalently, the chilled water flow rate), respectively. This approach
does not rely on the use of RH sensor and accurate model of the economizer for optimal
operation. Therefore, it provides a more reliable control strategy for economizer
operation. The proposed ESC scheme works as part of a 3state economizer control
strategy [81], as shown in the state transition diagram in Figure 6.2. State 1 uses heating
to maintain the supply air temperature. In state 2, the outdoor air is mixed with the return
air to maintain the supply air at a given setpoint. In state 3, the ESC is used to control the
dampers to minimize the mechanical cooling load. Also, the dampers must be controlled
to guarantee enough outdoor air inflow to satisfy the ventilation requirement for the
rooms. Figure 6.3 shows the control regions for different outside air conditions on a
psychrometric chart. The return air condition is 75 F and 50% RH, the cooling coil was
ideal, and the minimum fraction of outdoor air to supply air is 0.3. The heating region is
for state 1, the free cooling region is for state 2, and the three regions that need
mechanical cooling are combined into state 3 [81].
Compared to existing economizer control strategies, the proposed ESC has the
following advantages: 1) improving energy efficiency and 2) reducing the sensor related
costs. First, with the proposed ESC, optimal damper position can be determined
automatically and also adaptive to changes in the system such as external and internal
load changes. Thus, if implemented and operated correctly, ESC is expected to save
energy in a continuous manner. Second, with the proposed ESC framework, the
201
temperature and RH sensor at the return side and the RH sensor at the outdoor air side
can be removed, which eliminates the sensor related costs, including installation,
calibration and maintenance.
In order to design and simulate the proposed control strategy, a quality dynamic
model of economizer is needed. The dynamic AHU model developed in Chapter 3 is thus
adopted for the control study. In addition to application of the standard ESC to
economizer control, an enhancement on the ESC is proposed: an antiwindup ESC
scheme against damper (actuator) saturation. Due to the inherent integral action
incorporated in the ESC loop, the integral windup due to the damper saturation would
disable the ESC, as to be shown in the simulation section. The backcalculation scheme
[235] is applied to the ESC loop to achieve the antiwindup capability.
Figure 6.2: State transition diagram for the proposed control strategy (reproduced with
permission from Figure 4 in [81])
202
Figure 6.3: Illustration of the three control states with different outside air conditions for
an ideal coil with return conditions of 75F and 50% RH (reproduced with permission
from Figure 9 in [81])
6.3 Extremum Seeking Control (ESC)
6.3.1 Overview of ESC
The extremum seeking control deals with the online optimization problem of finding
an optimal input u
opt
(t) for the generally unknown and/or timevarying cost function l(t,
u), where u(t) e
m
is the input parameter vector [238], i.e.
opt
( ) argmin ( , )
m
u
u t l t u
e
=
(6.1)
Figure 6.4 shows the block diagram for a typical ESC system [238]. The
measurement of the cost function l(t, u), denoted by y(t), is corrupted by noise n(t). The
transfer function F
I
(s) represents the linear dynamics of the mechanism that command the
203
control or optimization tuning parameter vector u(t). F
O
(s)
represents the transfer function
of the sensor dynamics that measure the cost function, which is often a lowpass filter for
removing noise from the measurement. The basic components of the ESC loop are
defined as follows [239]. The dithering and demodulating signals are denoted by
 
2 1 1 1
( ) sin( sin( ) )
T
m m m d d
d t a t a t e o e o = + + and  
1 1
( ) sin( ) sin( )
T
m d d
d t t t e e = ,
respectively, where e
di
are the dither frequencies for each input parameter channel, and
o
i
are the phase angles introduced intentionally between the dithering and demodulating
signals. The signal vector
2
( )
T
d t includes the dither signals used to perturb and extract the
gradient of the cost function l(t, u). These signals work in conjunction with the highpass
filter F
HP
(s), the demodulating signal
1 1
( ) sin( ) sin( )
T
dm d
d t t t e e = (
and the lowpass
filter F
LP
(s), to generate a vectorvalued signal that is proportional to the gradient ( )
l
u
u
c
c
of the cost function at the input of the multivariable integrator, where u is the estimated
control input based on the gradient estimation. By integrating the gradient signal,
asymptotic stability of the closed loop system will make the gradient vanish, i.e.,
achieving the optimum. Adding compensator K(s) may enhance the transient
performance by compensating the input/output dynamics. For a detailed explanation of
ESC, refer to references [86, 87, 238].
Figure 6.4: Block diagram of ESC [50]
F
HP
(s)
F
I
(s) l(t, u)
d
1
d
2
n
y
u
F
LP
(s) K(s)
F
O
(s)
204
6.3.2 ESC for Energy Efficient Operation of Economizers
The proposed ESC based economizer control is illustrated in Figure 6.5. This control
strategy can be considered as a dualloop structure. The inner loop is the supply air
temperature control for the cooling coil, which has faster dynamics. The outer loop is the
damper opening tuning for minimizing the cooling coil demand, which is realized with an
ESC framework. The nonlinear performance mapping is from the outdoor air damper
opening to the cooling coil demand, and the input dynamics are effectively the closed
loop dynamics for supply air temperature control. In the previously described 3state
economizer operation scheme [81], ESC is applied to the state3 operation where
mechanical cooling is required.
(a) Detailed block diagram
(b) Simplified block diagram
Figure 6.5: ESC based airside economizer control [50]
F
HP
(s)
K
cc
(s)
d
1
d
2
n
C
Q
y
u
F
LP
(s)
K(s)
Cooling
Coil
Cooling coil demand
Damper
command
Setpoint Supply
Air Temperature
Supply Air
Temperature
+
+
+
+
+
Mixing
Damper
Mixed Air
F
HP
(s)
F
cc_cL
() ( , )
C
Q t u
d
1
n
y
u
F
LP
(s)
K(s)
205
6.3.3 ESC Design
Typical ESC design needs to determine the following parameters: the dither
amplitude a, the dither frequency
d
and the phase angle o, the high pass filter F
HP
(s),
the low pass filter F
LP
(s), and the dynamic compensator K(s). Based on averaging
analysis, the dither frequency should be relatively large with respective to the adaptation
gain, but should not be too large to trigger unmodeled dynamics [239] and make the
system more sensitive to measurement noise. Also, if the dither frequency is well out of
the bandwidth of the input dynamics, the rolloff in the magnitude response will slow
down the convergence [87]. Therefore, dither frequency
d
is typically chosen to be just a
moderate value smaller than the cutoff frequency of the input dynamics as long as it is
enough to separate the time scales of the dither signal and the inner loop dynamics.
Generally, the dynamic compensator should be designed based on the dither signal,
adaptation gain and the frequency responses of the input dynamics. Particularly, a proper
proportionalderivative (PD) action can increase the phase margin of the input dynamics
and thus make the inner loop more stable. However, extreme values of the adaptation
gain, especially the derivative gain, will make the system more affected by noise and thus
destabilize the system. Further design guidelines are summarized as follows [95, 238].
1) The dither frequency should be in the passband of the high pass filter and in the
stopband of the low pass filter, and it should be below the cutoff frequency of the
input dynamics of the respective channel.
2) The dither amplitude should be chosen to be sufficiently small to avoid large
oscillation of output, and meanwhile sufficiently large to overcome the noise effect.
206
3) The dither phase angle should be chosen such that
( ) ( )
2 2
, I HP F j F j
t t
u e e o = Z + Z + e
 

\ .
, and it is desirable to make u close to 0.
6.3.4 AntiWindup ESC
Actuator saturation is often encountered in control systems. For the ESC based
economizer control, the actuator saturation will happen when it is cool or hot outside.
Such implementation issue should be addressed for the 3state control strategy described
in Figure 6.2. For example, when the outdoor air is around 53 F, the outdoor air damper
will be positioned fully open to allow 100% outdoor air to enter the AHU (freecooling
state). When it is warmer than 100 F, the damper will be closed to a minimum opening
which only maintain the lowest ventilation for indoor air quality [240] (mechanical
cooling state). In other words, the optimal reference input is not inside the saturation limit,
but rather at either limit point. Transition between the ESC operation and the nonESC
operation is affected by the saturation issue. The averaging analysis of ESC [238] showed
that, at a large time scale, the ESC can be deemed as a linear system regulating the
gradient signal with a PI controller. When saturation presents in the ESC loop, integrator
windup is unavoidable, and in consequence, leads to the undesirable windup phenomena.
Later in the simulation section, results will show that, due to the windup issue, the ESC
action may be totally disabled even when the air condition changes to a point demanding
its reactivation. It is thus necessary to modify the standard ESC structure in order to
avoid integrator windup.
A lot of work has been reported in the field of antiwindup control (AWC) [241, 242].
In order to keep the simple nature of ESC, a backcalculation method was proposed in
207
[235] following the spirit of the references [242244]. The difference between the input
and output of the actuator is fed back to the input end of the integrator through some gain
factor. Figure 6.6 shows the block diagram of the antiwindup ESC with backcalculation
method. This dissertation study adopts the backcalculation method in [235], and the
simulation study in the next section will evaluate this method to prevent the integrator
windup in ESC based economizer control system.
Figure 6.6: Block diagram for the antiwindup ESC [50]
6.4 Simulation Study
6.4.1 ESC with Standard Design
As previously stated, the control objective in this study is to minimize the chilled
water flow rate of the cooling coil by tuning the OAD opening. The following second
order linear model was used to approximate the input dynamics:
2
2 2
( )
2
n
I
n n
F s
s s
e
,e e +
=
+
(6.2)
where e
n
is the undamped natural frequency and , is the damping ratio. The input
dynamics from the OAD opening to the chilled water flow rate was approximated based
+
F
HP
(s) Integrator
F
I
(s) l(t, u)
n
y
u
F
LP
(s)
F
O
(s)
+
+
+
+
+
+
Gain
d
1
d
2
208
on several inner loop simulations. Timedomain system identification was used based on
the step response based parameter estimation. A fast (3 seconds) ramp input was used to
approximate step input in order to remove the output jitter due to the inner loop PID
control. The e
n
was estimated from the 10%to90% rise time
1.8
r
n
T
e
= [245].
A suitable e
n
could be then obtained by further tuning the responses manually based
on the above approximation. A group of tests indicated that the average value of e
n
was
about 0.2 rad/sec. The damping ratio was estimated to be 1. Generally speaking, a higher
dither frequency will increase the convergence speed, but the maximum value is limited
to the cutoff frequency of the input dynamics. As a conservative design, to properly
separate the dither signal and plant dynamics, the dither frequency
d
was selected as
approximately onethird of the inner loop cutoff frequency. Next, a high pass filter F
HP
(s)
was designed to wash out the low frequency component and is given by:
2
2 2
( )
2 0.65 0.0125 0.0125
HP
s
F s
s s
=
+ +
(6.3)
which has unity gain at
d
. Ideally, after demodulation, the signal will contain a low
frequency component that is proportional to the gradient. A lowpass filter was designed
to keep this component and suppress high frequency components and is given by:
2
2 2
0.0167
( )
2 0.65 0.0167 0.0167
LP F s
s s
=
+ +
(6.4)
which has approximately 15 dB and 30 dB attenuation at
d
and 2
d
, respectively. The
dither amplitude was designed to have 5% opening variation. To compensate for the
209
phase lag and phase lead from the input dynamics F
I
(s) and the high pass filter ( )
HP
F s ,
the dither phase o was selected as 0.0071 radian, which led to
( ) ( ) 0
d d
FI j FHP j u e e o = Z +Z + ~ .
The designed ESC controller was first tested with a fixed operating condition. The
initial air temperature and RH were set to be 298.15 K (77 F) and 33%, respectively.
The supplyair temperature was controlled at 286 K (55 F) and the return air temperature
was maintained around 297 K (75 F) by providing a constant heat input to the indoor
space. A minimal OAD opening of 30% was assumed to ensure adequate indoor air
quality. In addition, the indoor humidity gain was assumed to be generated by people so
that the RH of the return air was maintained around 50%. The static mapping from the
OAD opening to the chilledwater flow rate was approximated with the steadystate
simulation data in Figure 6.7. The optimal OAD opening and chilledwater flow rate
were around 53% and 2.365 kg/s, respectively.
210
Figure 6.7: Static map from OAD opening to chilled water flow rate
Figure 6.8 shows the time histories of the optimized chilled water flow rate and OAD
opening. The system started at the minimal OAD opening, i.e. 30%. The ESC controller
was turned on at t = 2000 seconds, and the system output was brought to the optimum
with the settling time of about 884 seconds. At steady state, the mean values of the OAD
opening and the chilledwater flow rate were 52.42% and 2.376 kg/s, respectively, which
are off from the optimum in Figure 6.7 by only 0.58% and 0.011 kg/s, respectively.
211
Figure 6.8: Tuning results of standard ESC design based on 30% initial OAD opening
A further study was conducted to test the ESC tunings with different initial OAD
openings. Figure 6.9 through Figure 6.14 show the tuning results of different initial
condition based on the same outdoor air condition as above. These initial OAD opening
test points are given by 20%, 25%, 40%, 65%, 75% and 85%, respectively. Note that in
the leftmost of the time axis in all the cases, at around time smaller than 1000 seconds,
there appears a decrease and then an increase in water flow rate. At that time period, the
ESC is not turned on. During the system startup without ESC, the damper is first
initialized in a certain position (depending on the test cases), based on which the initial
cooling water flow rate is computed. However, due to the supplyair temperature control
(implemented as a PID controller). The valve in the coolingwater supply side is adjusted
by the PID controller to achieve the control target. As a result, the water flow rate varies
with the change of the valve positions until the supply temperature setpoint is achieved.
Then, the system will reach steadystate. Comparison of Figure 6.9 and Figure 6.13
212
reveals that the settling time is similar for initial OAD opening of 20% or 75%. It seems
that the settling time increases with the distance between the initial condition and the
optimum point.
Figure 6.9: Tuning results of standard ESC design based on 20% initial OAD opening
Figure 6.10: Tuning results of standard ESC design based on 25% initial OAD opening
213
Figure 6.11: Tuning results of standard ESC design based on 40% initial OAD opening
Figure 6.12: Tuning results of standard ESC design based on 65% initial OAD opening
214
Figure 6.13: Tuning results of standard ESC design based on 75% initial OAD opening
Figure 6.14: Tuning results of standard ESC design based on 85% initial OAD opening
215
6.4.2 AntiWindup ESC
Simulation study was also conducted to verify the effectiveness of the back
calculation based antiwindup ESC scheme [235]. Assume that a 30% damper opening is
still the minimum requirement for indoor air quality, and this was set as the lower
saturation limit. The upper saturation limit was 100%. The initial outdoor air damper
opening was again set at 30%, same as the lower saturation limit. The change of outdoor
air conditions is represented on the psychrometric chart shown in Figure 6.15. Figure
6.16 shows the ramp change of the outdoorair conditions at 5000 and 8000 seconds. The
initial air temperature and RH were again set to be 298.15 K and 33% (State 1),
respectively. Figure 6.17 shows the static map of chilled water flow rate to OAD opening
at State 2. Figure 6.18 shows the integrator windup phenomenon when the standard ESC
scheme was applied. Driven by the ESC, the damper opening first reached 52.42%,
which was supposed to be the first optimum. Then the outdoorair temperature was
suddenly (100 seconds ramp) increased to 305.15 K (89.6 F) at 5000 seconds (State 2).
As shown in Figure 6.18, the standard ESC scheme was able to detect such change and
the damper opening was decreased from 52.42% to 30% which was the corresponding
achievable optimal setting. However, when another 100 seconds ramp signal was applied
to bring the outdoor air temperature and RH (State 3) back to the initial settings at 8000
seconds, the new optimal opening was supposed to be switched back to the first optimum,
i.e. 52.42%. However, the results show that the standard ESC was unable to respond to
such change by increasing the damper opening. Rather the damper appeared stuck at
the previous position. In comparison, as shown in Figure 6.19, the problem was
effectively solved after the backcalculation based antiwindup ESC was applied from
216
2000 seconds. Therefore, the backcalculation based antiwindup ESC scheme is capable
of handling the saturation windup problem. Note that the OAD opening kept oscillating
in Figure 6.19. Such behavior is acceptable for durability concerns. In actual operation,
the ESC can be turned off once the optimum (convergence) is obtained. For building
HVAC operations, the optimal setpoint varies slowly with the changes from thermal
responses in the conditioned space and/or changes from the outdoor air conditions. In
addition, the period of the dither signal was set to be 150 seconds in this study, which
could hardly damage the dampers.
Figure 6.15: Illustration for the changes of the outdoorair conditions on the
psychrometric chart for antiwindup ESC simulation
217
Figure 6.16: Ramp change of outdoor air conditions at 5000 and 8000 seconds
Figure 6.17: Static map from OAD opening to chilled water flow rate (State 2)
218
Figure 6.18: Integral windup of standard ESC under damper saturation
Figure 6.19: Antiwindup ESC under damper saturation
219
6.5 Summary
As part of the 3state economizer control strategy [81], the extremum seeking control
has been applied to the selfoptimizing operation for the airside economizers, resulting
in higher efficiency and reliability. The mechanical cooling load can be minimized via
searching for the optimal outdoor air damper opening. Simulation study has been
conducted on a Modelica based transient model of a singleduct AHU developed in
Chapter 3. Simulation results have demonstrated that the optimal damper opening can be
obtained based on the feedback of the chilledwater flow command. Due to the inherent
integral element in ESC, the windup phenomenon has been observed through simulation,
which may disable ESC during practical operation. An antiwindup ESC strategy [235] is
adopted. Its effectiveness was validated with simulation. The proposed ESC strategy
indicates a perspective of sensorfree operation for economizers, which would greatly
enhance the efficiency and reliability for such devices, and also reduce the cost of system
operation and energy consumption.
The ESC design described in this study is applicable to the situation when the
nonlinear dynamics can be decoupled as a nonlinear static map cascaded with a linear
input dynamics. When the nonlinear static map is convex, then it is possible to use ESC
to search for the optimum. The choice of dither frequency and thus the design of F
HP
(s)
and F
LP
(s) are based on the knowledge of the input dynamics, as discussed earlier.
Dithering amplitude and loop gain K can be determined from the Hessian of the static
map near the optimum. A worstcase estimation can be used to achieve a safe (robust)
solution.
220
Chapter 7 Experimental Validation
This chapter presents experimental validation of the cooling coil modeling and an
important assumption generally made in vapor compression cycle modeling. The
remainder of this chapter is organized as follows. Section 7.1 presents the experimental
validation of the proposed dynamic model for the chilledwater cooling coil. Section 7.2
presents the computation efficiency of the proposed coil model. Finally, Section 7.3
presents the experimental validation of the dp/dt assumption in heat exchanger modeling
for dynamic simulation of vapor compression refrigeration system.
7.1 Benchmark with Expermental Data from Zhou (ASHRAE RP
1194) [18]
The dynamic model of a chilledwater cooling coil was developed in the previous
section. This section shows the benchmark results of the proposed model with
experimental data as well as Zhous model [18] (denoted as ASHRAE RP1194). The
proposed model adopted the same geometries as the cooling models developed by Zhou
[18]. On both sides of the coil wall, several parallel flow channels are lumped into one
uniform flow path. The cooling water path through the component is treated as one pipe
flow with circular cross section and one air element associated with each flow segment.
Each air element could be further discretized along its flow direction. In the following,
the 4row cooling coil is selected as an illustrating example to describe the air and water
flow patterns throughout the coil. Figure 7.1 schematically shows the detailed airflow
pattern and the waterflow circuiting throughout the example 4row coil. For the 8row
coil, the number of water flow passes is doubled. To be consistent with Zhou's model, the
221
total number of control volumes for the 4row and 8row coils were set to be 4 and 8,
respectively. The dimensions of the 4row and 8row coils are summarized in Table 7.1.
(a) General view of the 4row cooling coil.
(b) Detailed view of flowcircuiting in the 4row cooling coil.
Figure 7.1: Schematic diagrams for the 4row cooling coil
222
Table 7.1: Dimensions of the Test Coils [18]
Coil Physical Parameters 8Row Coil 4Row Coil
Fin geometry Wavy Louvered
Coil depth (m) 0.264 0.132
Number of fins per inch (/0.0254m) 8 12
Coil face width (m) 0.6096 0.6096
Coil face height (m) 0.6096 0.6096
Tube material Copper Copper
Tube outer diameter (m) 0.0127 0.0127
Tube thickness (m) 0.0004 0.0004
Tube longitudinal pitch (m) 0.033 0.033
Tube transverse pitch (m) 0.0381 0.0381
Fin material Aluminum Aluminum
Fin thickness (m) 0.0002 0.0002
The following shows the benchmark results of the proposed coil model against the
experimental data collected by Zhou [18]. For the purpose of comparison, the
corresponding simulation results obtained from Zhous dynamic coil model are also
presented. Both the 4row and 8row coils were employed in this study. The comparison
cases cover a wide range of practical HVAC operations for the chilled water cooling coils.
Table 7.2 summarizes the inlet variations of the coil boundary conditions in terms of flow
rate, temperature, humidity ratio and relative humidity. For cases 4 and 5 in Table 7.2, the
inlet variations of air and water flow rates caused by feedback controls are given in
Figure 7.2 (a) and Figure 7.2 (b), respectively. Finally, Figure 7.3 through Figure 7.9
223
present the benchmark results for all 7 comparison cases. To verify the underlying
simulation code, the energy balance in case 1 was presented via the corresponding energy
profile (see Figure 7.3 (d)). At any given time instance of the simulation , the sum of the
heat transfer rates from the airside, the wallside and the waterside is expected to be
zero.
Table 7.2: Coil Inlet Conditions for Sample Transient Comparisons [1]
Case Coil Operation
a
m (kg/s) T
a,in
(C) RH
air, in
(%) W
a,in
(g/kg)
w
m (kg/s) T
w,in
(C)
1 8row Fullywet 0.76 1.13 23.94 70.8 13.3 0.43 2.33
2 4row Partiallywet 0.761.25 25.39 54.04 11.08 0.25 1.19
3 4row Partiallywet 1.42 25.63 53.15 10.93 0.230.47 2.5
4 8row Partiallywet Fig. 7.2(a) 22.12 43.51 7.28 0.42 2.4
5 4row Partiallywet 1.26 26.13 53.04 11.36 Fig. 7.2(b) 3.55
6 8row Partiallywet 1.05 27.6423.28 35.0845.45 8.18 0.45 2.93
7 8row Partiallywet 1.04 26.67 40.4763.01 8.7214.08 0.44 2.75
(a) Feedback control of air flow rate
224
(b) Feedback control of water flow rate
Figure 7.2: Inlet flow rate variations of cases 4 and 5 in Table 7.2
(a) Air outlet temperature
225
(b) Air outlet humidity ratio
(c) Water outlet temperature
226
(d) Energy balance profile
Figure 7.3: Increase air inflow rate for an 8row fully wet coil (case 1)
For all the cases, the proposed coil model yielded better airside transient and steady
state predictions as compared to the experimental results. For some cases, the proposed
coil model did not yield better steadystate predictions of the waterside. This may be
associated with the energy balances reinforced in the simulations. For example, in case 2,
the proposed model has better airside steadystate predictions, but worse waterside
steadystate predictions. In steady state, the heat loss from the air side is assumed to be
completely transferred to the water side. To achieve better waterside steadystate
predictions, the waterside temperature is expected to increase to be closer to the
experiment results, the total heat transfer rate at the water side would then increase with
the adjustment of the water temperatures. Due to the energy balance constraint, the total
227
heat transfer rate at airside would also increase and hence the airside temperatures
would decrease, which in turn makes the airside predictions worse. Therefore, there
seems a tradeoff between the performances in the air and waterside steady state
predictions.
(a) Air outlet temperature
228
(b) Air outlet humidity ratio
(c) Water outlet temperature
Figure 7.4: Increase air inflow rate at low water flow rates for a 4row partially wet coil
(case 2)
229
(a) Air outlet temperature
(b) Air outlet humidity ratio
230
(c) Water outlet temperature
Figure 7.5: Increase water inflow rate for 4row partially wet coil (case 3)
(a) Air outlet temperature
231
(b) Air outlet humidity ratio
(c) Water outlet temperature
Figure 7.6: Feedback control of air flow rate for 8row partially wet coil (case 4)
232
(a) Air outlet temperature
(b) Air outlet humidity ratio
233
(c) Water outlet temperature
Figure 7.7: Feedback control of water flow rate for a 4row partially wet coil (case 5)
(a) Air outlet temperature
234
(b) Air outlet humidity ratio
(c) Water outlet temperature
Figure 7.8: Decrease air inlet temperature for an 8row partially wet coil (case 6)
235
(a) Air outlet temperature
(b) Air outlet humidity ratio
236
(c) Water outlet temperature
Figure 7.9: Increase air inlet humidity for an 8row partially wet coil (case 7)
7.2 Comparion of Compution Efficiency
Computation time is a crucial factor for evaluating dynamic simulation of HVAC
systems. This section will compare the CPUintegration time of the proposed coil model
developed in Dymola with Zhous dynamic forward model developed in C++. For the
proposed model, two integration methods were evaluated. The first integration method
was DASSL (Differential Algebraic System Solver) [217] and the second integration
method was the 5
th
order Radau IIa, which is a fully implicit RungeKutta method [246].
The CPUintegration time was calculated by running the models on a Dell desktop
computer equipped with Intel Core 2 CPU 6300 with 1.86GHz and 3GB RAM. Table 7.3
summarizes the computation time of the proposed coil model based on the DASSL and
Radau IIa integration algorithms as well as Zhous model developed in C++ and run from
Matlab
2007b. The evaluation was based on the same 7 cases as shown in the above coil
237
benchmark section. Radau IIa was the fastest among all the cases. The computation time
of the proposed model appears more sensitive to the number of control volumes. This
could be better illustrated from the comparisons between the DASSL solver and Zhous
method. For the 8row cases, the computation speed of Zhous method was nearly 2.5
times faster than the DASSL solver. However, when the number of control volumes were
halved, i.e., for the 4row cases, the DASSL solver became more efficient and its
computation speed was about 1.5 times faster than that of Zhous method.
The solution profiles of the DASSL, Radau IIa and Zhous methods are compared in
Figure 7.10. For illustration purpose, the air outlet temperature profile from case 2 is
adopted. Compared with the Radau IIa solver, the DASSL solver yields slightly better
results in terms of the quality of smoothness. However, the Radau IIa solver has greater
advantage in computational speed. Zhous model adopted explicit integration scheme
with a fixed timestep of 0.1 second [18]. For stability and accuracy concerns, explicit
methods need to adopt smaller timestep than the implicit methods. This may explain the
reason why the solution profile looks a little more corrugated for some time intervals. As
can be better illustrated in the zoomed plot inside Figure 7.10, at certain time spans such
as from 65 to 70 seconds, the slope of the air temperature did not change sharply;
however, the explicit solver still marched forward with a 0.1 second timestep, which
may lose some computational efficiency.
238
Figure 7.10: Comparison of the simulation profiles of DASSL, Rauda IIa and Zhous
method (ASHRAE RP1194) [18]
Table 7.3: Comparison of Computation Time (CPUIntegration Time)
Case
Proposed Model Results (sec) Zhous
(sec)
DASSL Radau lla
1 14.70 4.80 6.78
2 4.20 1.83 6.16
3 4.30 1.89 6.18
4 30.00 9.17 13.82
5 8.33 3.53 13.58
6 15.20 4.50 6.27
7 14.90 4.55 6.74
239
7.3 Experimental Validation for the dp/dt Assumption of Heat
Exchangers in Vapor Compression Refrigeration Cycles
7.3.1 Introduction
For dynamic modeling and simulation of vaporcompression cycles, the time
derivative of pressure (dp/dt) is generally calculated either in each discretized volume for
finite volume methods or in each phase zone for moving boundary methods. Such
treatment usually result in inferior computational efficiency and sometimes the solver
may fail to converge for large and complex thermalfluid systems due to the sensitivity of
pressure changes. Recently, a new concept of modeling dp/dt in heat exchangers was
proposed by Lemke [200] and implemented in a Modelica Library TIL [40] for dynamic
simulation of thermofluid system. As Lemke proposed, the time derivative of pressure at
the inlet of the heat exchanger can be treated equal to that at the outlet, i.e.,
in out
dp dt dp dt ~ . As a result, the overall computational efficiency is claimed to have
improved greatly [40]. This assumption is also very useful to deal with the common
difficulty in finding consistent initial conditions for large and coupled set of differential
algebraic equations, which are very common in thermalfluid modeling.
To the authors best knowledge, the experimental validation of this key assumption
(
in out
dp dt dp dt ~ ) has only been conducted by Lemke with a gas cooler in R744 (CO2)
based refrigeration cycle [200]. However, the operating phase and physical properties of
the refrigerants for typical vapor compression cycle system, such as R22, R134a and
R410A, are quite different from those of R744 based systems. For R744 based
refrigeration cycle, the refrigerant in the gas cooler is in the supercritical state only.
240
While for the typical vaporcompression systems, such as chillers, it is very possible that
the refrigerant will undergo a phase transition from the superheated, twophase, to
subcooled regimes throughout the condenser. Therefore, the observations and
conclusions in Lemkes work cannot be simply extended to these systems without further
experimental investigation.
Thus, the main focus of this study is to experimentally evaluate the change of time
derivatives of pressures (dp/dt) at the inlet and outlet of the condenser and evaporator on
a typical vapor compression cycle. The experiments were conducted on a York water
cooled screw chiller operated with R134a as the working medium. In addition, compared
to Lemkes work, both the condenser and the evaporator sides are considered. Finally, as
temporal differentiation of pressure signals lead to significant numerical noises, a
particular filtering technique, the SavitzkyGolay smoothing filter [247], is applied for
denoising the differentiated experimental data.
7.3.2 Experiment Setup
The experiments were conducted on a York (model number: YR) screw chiller testing
facility at Johnson Controls China in WuXi, China. The YR chiller uses electric drive
and the working fluid is refrigerant R134a. Figure 7.11 shows the photograph of the
experimental setup.
241
Figure 7.11: Photograph of the chiller test bench
7.3.2.1 Selection and Installation of Pressure Sensors
There were two major concerns when selecting pressure sensors, i.e., pressure range
and bandwidth. The upper pressure bound of the test chiller is 1034.2 kPa. As the
objective of this experimental study is to capture the transient, it is hard to define the
bandwidth directly regarding to the frequency contents of the measured signals, as
Fourier transform is applicable only to (nearly) steadystate signals. As a remedy, we
consider the choice of sampling period from the digital control perspective, i.e. the
sampling period is chosen for a computer controlled chiller system. As suggested by
strm and Wittenmark [241], it is preferred that, within the rise time, there should be 4
to 10 samples in order to have small distortion in the discretetime frequency response
from the continuoustime system.
242
With the Modelica based dynamic chiller developed in Chapter 4, the harshest
possible transient change scenarios have been simulated. The rise time of the condenser
and evaporator pressures were found to be around 2 to 3 seconds. By considering 10
samples within the rise time, the sampling frequency is at least 5 Hz. The AST4000AV
0250P4E1000 pressure sensor is thus chosen, which has bandwidth of 0 250 Hz. In
addition, considering the electromagnetic interference (EMI) issue of high power AC
motor, the 420 mA current loop was selected for better quality over higher current range.
Figure 7.12 schematically shows the sensor allocation on the screw chiller for the test.
The instrumentation of the test is summarized in Table 7.4.
Figure 7.12: Schematic of sensor allocation and data acquisition system for dp/dt test
243
Table 7.4: Instrumentation for the dp/dt Test
Instrumentation
Location Number of
Probes
Accuracy
Platinum resistance
thermometers
Condenser water inlet 1
0.05 C
Condenser water outlet 1
Chilled water inlet 1
Chilled water outlet 1
Stainless steel media
isolated pressure
sensor
Condenser refrigerant inlet 1
1.5 kPa
Condenser refrigerant outlet 1
Evaporator refrigerant inlet 1
Evaporator refrigerant outlet 1
7.3.2.2 Compressor and Expansion Valve [248]
The rotary screw compressor is the twinscrew type and the compressor housing is
made of cast iron. The rotors are made of forged steel and have asymmetric profiles. The
maximum speed of the rotors is either 3570 RPM (~60 Hz) or 2975 RPM (~50 Hz). An
internal oil reservoir is used to supply oil to the bearings. A check valve is included to
avoid the problem of rotor backspin during shutdown. The squirrelcage induction motor
is with 2pole and has continuousduty operation. A slide valve is used to control the
compressor by modulating its capacity from 20 % to 100 % of the fullload. The slide
valve is driven by system differential pressure, which can be manipulated by external
solenoid valves via the OptiView control center. The compressor can be operated with
offdesign cooling tower water during partload operation in line with ARI Standard
550/590. Table 7.5 shows the dimensions of the screw compressor.
244
Table 7.5: Dimensions of the Screw Compressor Tested
The expansion device is a butterfly valve with electromagnetic actuator. Figure 7.13
shows the photographs of the valve.
(a) Close view (b) Distant view
Figure 7.13: Photographs of the butterfly valve used in York YR chiller
7.3.2.3 Condenser and Evaporator
Both the condenser and the evaporator are horizontal shellandtube heat exchangers.
The condenser is of flooded type, while the evaporator is a combination of a falling film
section and a flooded section. For both condenser and evaporator, water flows on the tube
side and refrigerant flows on the shell side. For condenser, refrigerant enters at the top
and leaves from the bottom. While for the evaporator, the refrigerant enters and leaves at
Compressor Physical Parameters Values
Male Rotor Diameter (mm) 226
Suction Diameter (mm) 254
Discharge Diameter (mm) 203
Displacement (m
3
/rev) 0.0106
Max. Condenser Temp. (C) 49.7
245
the top. The inner surface of the tube is enhanced by ribs, and the outer surface of the
tube is enhanced by integral fins. Figure 7.14 is a schematic drawing of the cross
sectional view of the water tubes in the condenser and evaporator. Table 7.6 through
Table 7.8 summarize the dimensions and configurations of water tubes in the condenser
and evaporator of the test chiller.
Figure 7.14: Cross sectional view of water tubes in the condenser and evaporator
Table 7.6: Dimensions of the Condenser and Evaporator
Tube Shell
Inner Diameter Outer Diameter Length Shell Diameter
Evaporator 16.5 mm 19.1 mm 3658 mm 635 mm
Condenser 15.8 mm 19.1 mm 3658 mm 641 mm
Table 7.7: Arrangement of Water Tubes at Condenser and Evaporator
Tube Row
Number of Tubes
Evaporator Condenser
1 10 21
2 9 22
246
3 10 23
4 9 24
5 10 25
6 9 26
7 10 27
8 9 26
9 10 27
10 9 27
11 10 26
12 9 27
13 10 26
14 9 27
15 10 26
16 9 25
17 10 18
18 5 12
19 11 12
20 12 N/A
21 11 N/A
22 10 N/A
23 9 N/A
24 8 N/A
25 7 N/A
26 6 N/A
27 5 N/A
Table 7.8: Total Number of Water Tubes at Condenser and Evaporator
Heat Exchanger Type Total Number of Tubes
Evaporator 246
Condenser 447
247
7.3.3 Resutls and Discussion
The tests were conducted by the following procedure and scenarios: 1) the evaporator
water flow rate increases incrementally from 40% to 130% of the nominal flow rate, 2)
the compressor capacity incrementally decreases from 100% to 50%, and 3) evaporator
water outlet temperature increases incrementally from 5 C to 10 C. Each test was
performed after the system reached steadystate for a long enough time.
In the data analysis, the dp/dt is estimated based on taking the finite difference of two
successive pressure measurements and dividing by the sample time (1 second). The
uncertainty of the calculated dp/dt was further evaluated based on [249].
( ) ( )
2 2
/
2
dp dt p p p
o o o o = + = (7.1)
The uncertainty of the individual pressure measurement (
p
o ) is 1.5 kPa (see Table 7.4)
and thus the uncertainty of the dp/dt, i.e. 2
p
o , is 2.1 kPa. The test results were
collected and summarized as the 8 test cases tabulated in Table 7.9. For numerical
differentiation, the general finite difference formula tends to produce noisy results and
thus gives misleading information. To overcome this issue, the SavitzkyGolay
smoothing filter [247] was adopted to smooth out the noisy signal due to the numerical
differentiation. The basic idea of the SavitzkyGolay smoothing filter is to fit local data
points using polynomial with the leastsquares method in order to determine the
smoothed data points. For post processing of the test data, the SavitzkyGolay filter
function from MATLAB

\ .
,
(7.2)
and is evaluated at each time instant to check the sensitivity of dp/dt differences to the
operating pressures.
Table 7.9: Operating Ranges for the dp/dt Tests
Test Cases
Pressure Range at
Condenser (kPa)
Pressure Range at
Evaporator (kPa)
1 [758.0, 966.4] [315.1, 633.6]
2 [801.4, 1253.5] [298.3, 684.9]
3 [881.0, 1157.1] [327.8, 732.3]
4 [804.6, 1079.3] [329.7, 673.1]
5 [830.7, 1183.0] [333.5, 600.5]
6 [824.1, 926.4] [345.2, 578.5]
7 [1023.7, 1125.0] [333.8, 532.6]
8 [649.3, 815.4] [358.8, 615.9]
Table 7.10 and Table 7.11 summarize the maximum values of the dp/dt differences
and the normalized dp/dt differences of all the test cases, for the condenser and
evaporator sides, respectively. Figure 7.15 through Figure 7.22 show the experiment
results corresponding to cases 1 through 8, respectively, in which the dp/dt profiles were
249
compared for the inlet and outlet of the condenser and evaporator, respectively, as shown
in plots (a) and (b) in each figure. The result of the normalized dp/dt difference of
condenser and evaporator are shown in plots (c) and (d), respectively.
Table 7.10: Summary of dp/dt Difference at Condenser Inlet and Outlet
Test Cases
Max. dp/dt Difference
(kPa/s)
Max. Normalized dp/dt
Difference (/s)
1 1.610
1
1.910
4
2 7.510
2
0.810
4
3 8.310
2
0.810
4
4 1.210
1
1.210
4
5 1.010
1
1.210
4
6 5.410
2
0.610
4
7 7.210
2
0.710
4
8 1.410
1
1.910
4
Table 7.11: Summary of dp/dt Difference at Evaporator Inlet and Outlet
Test Cases
Max. dp/dt Difference
(kPa/s)
Max. Normalized dp/dt
Difference (/s)
1 5.210
1
9.710
4
2 2.810
1
5.110
4
3 1.810
1
3.110
4
4 6.410
1
1.110
3
250
5 5.610
1
1.410
3
6 1.810
1
3.210
4
7 4.210
1
9.310
4
8 2.610
1
1.210
3
(a)
251
(b)
(c)
252
(d)
Figure 7.15: Experimental results of case 1 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
253
(b)
(c)
254
(d)
Figure 7.16: Experimental results of case 2 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
255
(b)
(c)
256
(d)
Figure 7.17: Experimental results of case 3 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
257
(b)
(c)
258
(d)
Figure 7.18: Experimental results of case 4 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
259
(b)
(c)
260
(d)
Figure 7.19: Experimental results of case 5 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
261
(b)
(c)
262
(d)
Figure 7.20: Experimental results of case 6 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
263
(b)
(c)
264
(d)
Figure 7.21: Experimental results of case 7 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
(a)
265
(b)
(c)
266
(d)
Figure 7.22: Experimental results of case 8 where (a) and (b) are the pressures at inlet and
outlet of condenser and evaporator respectively and comparison of dp/dt at their
respective inlet and outlet, (c) and (d) are the normalized dp/dt difference at condenser
and evaporator, respectively.
In summary, for the condenser side, the maximum dp/dt difference was 1.610
1
kPa/s and the maximum normalized dp/dt difference was 1.910
4
/s. For the evaporator,
the maximum dp/dt difference was 6.410
1
kPa/s and the maximum normalized dp/dt
difference was 1.410
3
/s. The experimental results of all the test cases show that the
change of time derivatives of pressures is quite small both in the sense of magnitude and
its relative change with respective to the operating pressure. Thus, the assumption of
equal time derivative of pressures (
in out
dp dt dp dt ~ ) is valid for vapor compression
refrigeration system that undergoes refrigerant phase transitions on both condenser and
evaporator sides. After validation, this key assumption can be readily applied to the
267
dynamic modeling of vapor compression refrigeration system. The dynamic model can be
further applied for testing new control methods to optimize the thermal efficiency such as
COP of the system, or can be applied for testing fault detection and diagnostics methods
to ensure better controller performance.
7.4 Summary
This chapter first evaluates the dynamic model of chilledwater cooling coil proposed
in Chapter 3 by use of the experimental results in an earlier ASHRAE project [18]. The
proposed model agrees with the experimental results very well both in transient and
steadystate for a wide range of operating conditions. In addition, an experimental study
on a commercial chiller system is presented for validating the dp/dt assumption for
transient simulation of vapor compression refrigeration systems. The experimental results
show that the time derivative of pressures at the inlet of the heat exchangers are
extremely close to that at the respective outlet, and can be thus treated as equal and
applied to the transient simulation of vapor compression system to improve its numerical
robustness and efficiency.
268
Chapter 8 Conclusions and Future Work
This chapter summarizes the contributions of this dissertation research, and then
points out the possible future work that may be followed.
8.1 Summary of Research Contributions
Dynamic modeling and simulation of HVAC equipment is an important topic for
control system design, validation, and FDI. The multiphysical nature of such simulation
challenge determines that equation based simulation platforms like Modelica have great
potential in this area. This dissertation research has developed dynamic simulation
models for airhandling unit and centrifugal chiller, and meanwhile investigating some
relevant research issues. In parallel, a selfoptimizing control strategy is applied to
maximize the efficiency of the airside economizer.
In summary, this dissertation study covers the following aspects of research relevant
to the dynamic modeling and control for chilledwater cooling systems:
1) Development of Modelica based dynamic simulation model for AHU;
2) Implementation and simulation of ESC based control strategy for efficient AHU
operation;
3) Development of Modelica based dynamic simulation model for centrifugal chiller
system;
4) Development of a consistent initialization strategy for the dynamic model of
centrifugal chiller;
269
5) Experimental validation of dp/dt assumption for heat exchangers in chiller systems.
The major research contributions of this dissertation research are described in the
following subsections.
8.1.1 Dynamic Modeling of Building ChilledWater Systems
One major contribution of this dissertation study is the development of Modelica
based dynamic simulation models for chilledwater cooling coil and centrifugal chiller in
central HVAC systems in buildings.
For the modeling of chilledwater cooling coil, a simple yet accurate water medium is
developed that is suitable for the use of building HVAC applications. To improve
numerical efficiency, the dynamic mass and energy balance equations are transformed
into the state variable pairs, pressuretemperature (p, T) and pressurespecificenthalpy (p,
h), respectively. The friction factor calculation for smooth pipes is important in
calculating the heat transfer characteristics for transition and turbulent flow regions. A
new explicit equation for the friction factor calculation for smooth pipes is developed and
compared with the best existing correlation. For the airside, a comprehensive literature is
conducted to find the plausible method of evaluating the latent heat transfer caused by
condensation. Instead of simply following the apparently wellreceived practice of
assuming the Lewis number to be 1, the author has conducted a comprehensive study on
the past literatures and clarified the definition and applicability of Lewis number and
Lewis relation. Based on that analysis, the Lewis relation is applied to calculate the
condensation flow rate for the cooling and dehumidifying process.
270
The transient behavior and steadystate predictions of the dynamic coil model have
been benchmarked with experimental data and also the dynamic model presented in
ASHRAE RP1194. The developed model demonstrated improved performance in both
transient and steadystate performance.
Watercooled chillers are generally regarded as the heart of the overall integrated
chilledwater cooling system. The most energy consumption device in a centrifugal
chiller refrigeration cycle is the centrifugal compressor. This dissertation study considers
a detailed geometrybased centrifugal compressor model as opposed to the mapbased or
datadriven compressor modeming in most existing works. Detailed losses at the impeller
and the diffuser are considered based on the compressor geometry parameters. The chiller
capacity is controlled by both variable inlet guide vane and variable speed drive. The
dynamic responses of the speed adjustment mechanisms are determined by the
momentum equation relating the drive torque and the compressor torque. The shelland
tube heat exchangers are discretized based on the finite volume method. When the control
volumes are discretized fine enough, the finite volume method is considered to yield
more accurate results and has more robust initialization over the moving boundary
method. The mass and energy balances are formulated in each divided control volume in
a numerically efficient way. The heat transfer coefficients are determined continuously
over the transition between different phase regions in the condenser and evaporator. An
orifice plate is used to regulate the refrigerant flow rate. The effective flow area can be
adjusted by an external input.
When the relevant chiller components are connected into a centrifugal chiller system
and formulated a nonlinear system of DAEs to be solved, the issue of consistent
271
initialization for the chiller cycle model is detected and is discussed in details in Chapter
5. This study proposes a direct initialization method, which contains threesteps that are
easy to apply in practice. For the particular case of centrifugal chiller initialization, a
threestep preprocessing scheme is proposed to obtain suitable initial guesses to guide the
nonlinear solvers to start at physically reasonable values. Simulation results have shown
the effectiveness of the proposed modeling framework and the initialization method.
8.1.2 SelfOptimizing Operation using Extremum Seeking Control
Another major contribution of this dissertation study is to apply the ESC to the self
optimizing operation for the airside economizers, resulting in higher efficiency and
reliability. Existing control strategies are usually ineffective mostly due to the large
sensor errors in relative humidity reading. On the contrary, the proposed ESC framework
does not rely on this sensor reading. The mechanical cooling load of the chilledwater
cooling coil can be minimized via searching for the optimal outdoor air damper opening.
Simulation study has been conducted on the detailed AHU model developed in Chapter 3.
Simulation results have demonstrated that the optimal damper opening can be obtained
based on the feedback of chilled water flow command. The ESC can also search for the
optimal damper opening at different initial positions. The integrator windup problem in
ESC based AHU control has been identified and handled using a backcalculation scheme
of antiwindup ESC. Its effectiveness has been validated with simulation.
The ESC strategy demonstrated in this study indicates a perspective of sensorfree
selfoptimization, which would greatly enhance the efficiency and reliability for
operation of economizers, thus lowering the operation and maintenance costs.
272
8.2 Suggested Future Work
The study of dynamic modeling and control of building HVAC systems is a broad
and fastgrowing field. Anchored on chilled water cooling systems for commercial
building HVAC systems, this dissertation study has developed solid ground for the future
development of dynamic simulation models for many other types of building HVAC
systems. Also, the AHU control work in this dissertation can be continued and further
improved in the future. By defining different setpoints and control objectives, the
proposed control method can also be applied to optimize the performance and efficiency
of chiller system and integrated chilledwater systems. Some suggestions for future
research opportunities are listed as follows.
8.2.1 Future Research for ChilledWater System
The effectiveness of the ESCbased economizer control is tested and validated
through simulation studies. In the future, it is desirable to evaluate and possibly improve
the controller design with experimental studies on a realistic AHU facility. It would be
desirable to conduct tests under various ambient conditions for different seasons, and also
for different load conditions. Foreseeable challenges include the mismatch between the
actual system behavior and the developed model, and also the possible nonlinearities and
defects in damper actuators.
Extensive research and development has been observed for surge characterization/
dynamics and active surge control for dynamic air compression systems, in particular by
aerospace industry, which has led to higher efficiency, better dynamic performance and
improved robustness. However, such effort is meager in the HVAC sector, e.g. for the
centrifugal chiller system as studied in this dissertation. In particular, the centrifugal
273
chiller does not present a less challenging scenario. As a matter of fact, such system
features a closedloop system by design, with distributedparameter, highly nonlinear,
and potentially unstable subsystems. Advanced dynamic control strategies that can locate
the actual maximum efficiency, rely on minimal surge margin and maintain good
robustness would significantly improve the operational efficiency for such systems.
The Modelica based dynamic model for centrifugal chiller follows the physics based
strategy, rather than datadriven. Therefore, validation of such model requires detailed
design and calibration parameters. However, such information is often proprietary for
chiller manufacturer. It is unfortunate that by the end of this dissertation research, it is
difficult to obtain the necessary parameters and test results to validate the developed
model, due to the unavailability of the relevant proprietary information. Nevertheless,
such detailed, physics based model, is highly beneficial for simulation, design
optimization, control design and fault detection among other purposes. Experimental
validation of the proposed model would still great benefit the HVAC equipment industry.
The extremum seeking control can be extended to study the selfoptimizing operation
of integrated chilledwater system by minimizing the total energy consumption from the
fans in cooling tower, the centrifugal compressor in chiller, and the cooling coil in AHU.
For the centrifugal chiller in particular, there could be interesting combination of some
outer loop ESC searching for the maximum achievable efficiency, and some robust
stabilizing controller that achieve robust stability near the surge boundary.
This dissertation concerns with the operation of single equipment, i.e. single AHU
and single chiller only. Consistent initialization has been investigated for singly unit of
274
chiller as well. In practice, a chiller plant may include multiple chillers and multiple
cooling towers. There may be a number of AHUs within a building attached to the chiller
plant. Multiple chillers may be connected in parallel, in series or a combination of both.
Dynamic simulation of such system of DAE systems will bring additional challenge for
consistent initialization. Further effort is expected on this subject.
8.2.2 Modelica Models for RealTime Applications
For large and complex Modelica models, it is desirable to further study the hybrid
simulation of continuoustime systems with digital control systems for realtime
simulation applications such as hardwareintheloop (HIL). Unlike the situations with
continuoustime models alone, discretetime simulation requires a fixed sampling time is
used (or multiples of a base sampling period for multirate control systems), and the
maximum timestep of the solver is limited to the sampling time interval. Such problems
can be handled via cosimulation. For example, Dymola and Matlab can be linked
together
13
to simulate digital control of HVAC systems. For some situations, it would be
more desirable to have the capabilities to study such problems solely in Dymola without
issues, especially for large and complex hybrid systems.
8.2.3 Integration with Smart Grid and Renewable Energy
As the technology of integrating smart grid and renewable energy with building
systems become more amenable, it is noteworthy that future demand response control
may require the HVAC equipment to interact more and more with local distributed
generation units and battery charging systems via electric grid. Therefore, dynamic
simulation of HVAC equipment would have to be compatible with the dynamic co
13
Synchronize data between each other by calling userdefined external C functions
275
simulation of electrical systems, grid, motor drives and power electronics. Under such
circumstances, a multiphysical platform like Modelica would be a very suitable tool. For
instance, it could be a possible scenario for a microgrid master controller might have to
coordinate with the transient controls for building HAVC equipment, wind and solar
power generators, and electric vehicle chargers. Such interesting scenarios would
naturally further complicate the simulation task, e.g. initial conditions for a number of
heterogeneous subsystems. More research would then be needed to achieve robust
simulation for the modeling tasks and optimal operations for the control tasks.
276
References
[1] DOE. (2010). Buildings Energy Data Book. Available:
http://buildingsdatabook.eren.doe.gov/
[2] P. Haves, T. Salsbury, D. Claridge, and M. Liu, "Use of Whole Building
Simulation OnLine Performance Assessment: Modeling and Implementation
Issues," presented at the Seventh International IBPSA Conference, Rio de Janeiro,
Brazil, 2001.
[3] P. Jacobs and H. Henderson, "StateoftheArt Review: Whole Building, Building
Envelope, and HVAC Component and System Simulation and Design Tools,"
2002.
[4] K. W. Roth, P. Llana, D. Westphalen, and J. Brodrick, "Automated Whole
Building Diagnostics," ASHRAE Journal, vol. 47, pp. 8284, 2005.
[5] M. Wetter, W. Zuo, and T. S. Nouidui, "Recent Developments of the Modelica
Buildings Library for Building Heating, Ventilation and AirConditioning
Systems," presented at the 8th International Modelica Conference., Dresden,
Germany, 2011.
[6] S. Katipamula and M. R. Brambley, "Methods for Fault Detection, Diagnostics
and Prognostics for Building Systems  A Review Part II," HVAC & R Research,
vol. 11(2), pp. 169187, 2005.
[7] H. Tummescheit, "Design and Implementation of ObjectOriented Model
Libraries using Modelica," Ph.D., Department of Automatic control, Lund
Institute of Technology, pp. 4449, 99100, 2002.
[8] F. Tang, "HVAC System Modeling and Optimization: A Datamining Approach,"
M.S., Industrial Engineering, The University of Iowa, 2010.
[9] DOE, "Energy Consumption Characteristics of Commercial Building HVAC
Systems Volume I: Chillers, Refrigerant Compressors, and Heating Systems,"
DEAC0196CE23798, 2001.
[10] J. E. Seem, Personal Communications, March, 2011.
[11] Modelica. (2011). Modelica and the Modelica Association. Available:
http://www.modelica.org/
277
[12] K. J. strm and R. M. Murray, Feedback Systems: An Introduction for Scientists
and Engineers Princeton University Press, p. 24, 2008.
[13] M. R. Brambley, P. Haves, S.C. McDonald, P. Torcellini, D.G. Hansen, D.
Holmberg, and K. Roth, "Advanced Sensors and Controls for Building
Applications: Market Assessment and Potential R&D Pathways," Pacific
Northwest National Laboratory, Richland, WA p. 6.18, 2005.
[14] D. E. Seborg, T. F. Edgar, and D. A. Mellichamp, Process Dynamics and Control,
2nd ed.: Wiley, pp. 510561, 2003.
[15] R. Huang, "Nonlinear Model Predictive Control and Dynamic Real Time
Optimization for Largescale Processes," Ph.D., Chemical Engineering, Carnegie
Mellon University, p. 4, 2010.
[16] S. Wang and Z. Ma, "Supervisory and Optimal Control of Building HVAC
Systems: A Review," HVAC&R Research, vol. 14, no. 1, pp. 332, 2008/01/01
2008.
[17] A. I. Dounis and C. Caraiscos, "Advanced Control Systems Engineering for
Energy and Comfort Management in a Building EnvironmentA Review,"
Renewable and Sustainable Energy Reviews, vol. 13, no. 67, pp. 12461261.
[18] X. Zhou and J. E. Braun, "Dynamic Modeling of Chilled Water Cooling Coils.
ASHRAE RP1194 Final Report," American Society of Heating, Refrigerating
and AirConditioning Engineers, Inc, Atlanta 2005.
[19] M. Wetter, "Modelicabased Modeling and Simulation to Support Research and
Development in Building Energy and Control Systems," Journal of Building
Performance Simulation, vol. 2, no. 2, pp. 143161, 2009.
[20] Mathworks. (2011). Simulink  Simulation and ModelBased Design,
http://www.mathworks.com/products/simulink/
[21] Dynasim. (2011). Available: http://www.3ds.com/products/catia/portfolio/dymola
[22] F. C. Winkelmann, B. E. Birsdall, W. F. Buhl, K. L. Ellington, A. E. Erdem, J. J.
Hirsch, and S. Gates., "DOE2 Supplement, Version 2.1E.," Lawrence Berkeley
National Laboratory, Berkeley 1993.
[23] L. K. L. Drury B. Crawley, Frederick C. Winkelmann, Walter F. Buhl, Y. Joe, C.
O. P. Huang, Richard K. Strand, Richard J. Liesen, Daniel E. Fisher,, and a. J. G.
278
Michael J. Witte, "EnergyPlus: Creating a NewGeneration Building Energy
Simulation Program," Energy and Buildings, vol. 33, no. 4, pp. 443457, 2001.
[24] K. J. strm, H. Elmqvist, and S. E. Mattsson, "Evolution of ContinuousTime
Modeling and Simulation," presented at the the 12th European Simulation
Multiconference, Manchester, UK, 1998.
[25] D. C. Augustin, M. S. Fineberg, B. B. Johnson, R. N. Linebarger, F. J. Sansom,
and J. C. Strauss, "The SCi continuous system simulation language (CSSL),"
Simulation, vol. 9, pp. 281303, 1967.
[26] D. R. Clark, "Building Systems and Equipment Simulation Program HVACSIM+
User Manual," National Bureau of Standards and Technology, Washington DC
1985.
[27] TRNSYS, A Transient Simulation Program. Reference Manual. Madison, WI:
Solar Energy Laboratory, University of Wisconsin, 1983.
[28] W. Borutzky, Development and Analysis of Multidisciplinary Dynamic System
Models, 1st ed.: Springer, 2010.
[29] D. C. Karnopp, "A Bond Graph Modeling Philosophy for Thermofluid Systems,"
ASME Journal of Dynamic Systems, Measurement, and Control vol. 100, no. 1,
70 (6 pages), 1978.
[30] D. Karnopp, "State Variables and Pseudo Bond Graphs for Compressible
Thermofluid Systems," Journal of Dynamic Systems, Measurement, and Control,
vol. 101, no. 3, pp. 201204, 1979.
[31] S. Azarbaijani and D. C. Karnopp, "Bond Graph Structural Models for Heat
Exchangers," International Journal of Modelling and Simulation, vol. 2, no. 3, pp.
138141, 1982.
[32] R. Shoureshi, "Dynamic Analysis and Failure Detection of HVAC Systems Using
TemperatureEntropy Bond Graph," presented at the ASHRAE Workshop on
HVAC Control, Modeling, and Simulation, Atlanta, Georgia, 1984.
[33] DOE. (2011). Building Energy Software Tools Directory. Available:
http://www.eere.energy.gov/buildings/tools_directory/
[34] Thermosys. (2007). Thermosys Description. Available:
http://www1.mengr.tamu.edu/TFCL/thermosys.htm
279
[35] J. Eborn, "On Model Libraries for Thermohydraulic Applications," Ph.D.,
Department of Automatic Control, Lund Institute of Technology, pp. 2745, 2001.
[36] Modelon. (2007). Available: http://www.modelon.se/
[37] F. Casella, M. Otter, K. Proelss, C. C. Richter, and H. Tummescheit, "The
Modelica Fluid and Media Library for Modeling of Incompressible and
Compressible ThermoFluid Pipe Networks," in Modelica Conference Vienna,
Austria, 2006.
[38] J. I. Videla and B. Lie, "A New Energy Building Simulation Library," in
Proceedings of Modelica 2006, pp. 685693, 2006.
[39] M. Ljubijankic, C. NytschGeusen, and S. Unger, "Modelling of Complex
Thermal Energy Supply Systems Based on the ModelicaLibrary Fluidflow," in
Proceedings of the 7th Modelica Conference, Como, Italy, Sep. 2022, 2009.
[40] C. C. Richter, "Proposal of New ObjectOriented Model Libraries for
Thermodynamic Systems," Ph.D., TU Braunschweig, pp. 57, 70, 8182, 2008.
[41] M. Grber, K. Kosowski, C. Richter, and W. Tegethoff, "Modelling of Heat
Pumps With an ObjectOriented Model Library for Thermodynamic Systems,"
Mathematical and Computer Modelling of Dynamical Systems: Methods, Tools
and Applications in Engineering and Related Sciences, vol. 16, no. 3, pp. 195 
209, 2010.
[42] TLKThermo. (2011). TLKThermo GmbH Engineering Services and Software for
Thermal Systems. Available: http://www.tlkthermo.com/
[43] S. Bendapudi and J. E. Braun, "A Review of Literature on Dynamic Models of
Vapor Compression Equipment," ASHRAE Report #40365, 2002.
[44] J. E. Seem, C. Park, and J. M. House, "New Sequencing Control Strategy for Air
Handling Units," HVAC&R Research, vol. 5, no. 1, pp. 3558, 1999.
[45] J. E. Seem, G. E. Kelly, J. M. House, and C. J. Klaassen, "Damper Control
System for Preventing Reverse Airflow through the Exhaust Air Damper of
VariableAirVolume AirHandling Units," HVAC&R Research, vol. 6, no. 2, pp.
135148, 2000.
280
[46] L. Song and M. Liu, "Optimal Outside Airflow Control of an Integrated Air
Handling Unit System for Large Office Buildings," Journal of Solar Energy
Engineering, Transactions of the ASME, vol. 126, no. 1, pp. 614619, 2004.
[47] C. Guo, Q. Song, and W. Cai, "A Neural Network Assisted Cascade Control
System for Air Handling Unit," IEEE Transactions on Industrial Electronics, vol.
54, no. 1, pp. 620628, 2007.
[48] T. I. Salsbury, "ModelBased Diagnostics for Air Handling Units," in
Proceedings of Diagnostics for Commercial Buildings: from Research to Practice,
June 1617, 1999.
[49] A. S. Glass, P. Gruber, M. Roos, and J. Todtli, "Qualitative ModelBased Fault
Detection in AirHandling Units," IEEE Control Systems Magazine, vol. 15, no. 4,
pp. 1122, 1995.
[50] P. Li, Y. Li, and J. E. Seem, "Efficient Operation of AirSide Economizer Using
Extremum Seeking Control," ASME Journal of Dynamic Systems, Measurement
and Control, vol. 132, no. 3, p. 031009 (10 pages), 2010.
[51] S. E. Mattsson, "On Modeling of Heat Exchangers in Modelica," in Proceedings
of the 9th European Simulation Symposium, Oct 1923, Passau, Germany, 1997.
[52] J. M. Jensen and H. Tummescheit, "Moving Boundary Models for Dynamic
Simulations of TwoPhase Flows," presented at the Proceedings of the 2nd
International Modelica Conference, Oberpfaffenhofen, Germany, March 1819,
2002.
[53] T. Pfafferott and G. Schmitz, "Modelling and Transient Simulation of CO2
Refrigeration Systems with Modelica," International Journal of Refrigeration
Volume 27, vol. 27, no. 1, pp. 4252, 2004.
[54] H. Tummescheit and J. Eborn, "Design of a ThermoHydraulic Model Library in
Modelica," presented at the Proceedings of the 12
th
European Simulation
Multiconference, ESM'98, Manchester, UK, 1998.
[55] T. Skoglund, K.E. rzn, and P. Dejmek, "Dynamic ObjectOriented Heat
Exchanger Models for Simulation of Fluid Property Transitions," International
Journal of Heat and Mass Transfer, vol. 49, no. 1314, pp. 22912303, 2006.
281
[56] D. Fu, G. Poncia, and Z. Lu, "Implementation of An ObjectOriented Dynamic
Modeling Library for Absorption Refrigeration Systems," Applied Thermal
Engineering, vol. 26, no. 23, pp. 217225, 2006.
[57] W.J. Zhang, C.L. Zhang, and G.L. Ding, "Transient Modeling of an AirCooled
Chiller with Economized Compressor. Part I: Model Development and
Validation," Applied Thermal Engineering, vol. 29, no. 1112, pp. 23962402,
2009.
[58] DOE, "Energy Consumption Characteristics of Commercial Building HVAC
Systems Volume II: Thermal Distribution, Auxiliary Equipment, and
Ventilation," DEAC0196CE23798, 1999.
[59] P. Haves, T. I. Salsbury, and J. A. Wright, "Condition Monitoring in HVAC
Subsystems Using First Principles Models," ASHRAE Transactions, vol. 102, no.
1, pp. 519527, 1996.
[60] J. E. Seem and J. M. House, "Development and Evaluation of OptimizationBased
Air Economizer Strategies," Applied Energy, vol. 87, no. 3, pp. 910924, 2010.
[61] J. Eborn, H. Tummescheit, and K. Prl, "AirConditioning  a Modelica Library
for Dynamic Simulation of AC Systems," in Proceedings of the 4th International
Modelica Conference, Hamburg, March 78, pp. 185192, 2005.
[62] W. Chen and S. Deng, "Development of a Dynamic Model for a DX VAV Air
Conditioning System," Energy Conversion and Management, vol. 47, no. 1819,
pp. 29002924, 2006.
[63] N. Nassif, S. Moujaes, and M. Zaheeruddin, "SelfTuning Dynamic Models of
HVAC System Components," Energy and Buildings, vol. 40, no. 9, pp. 1709
1720, 2008.
[64] S. Li and J. Wen, "Development and Validation of a Dynamic Air Handling Unit
Model, Part 1," in 2010 ASHRAE Winter Conference, January 23, 2010  January
27, 2010, Orlando, FL, United states, ASHRAE Transactions (116 PART 1), pp.
4556, 2010.
[65] S. Li, J. Wen, X. Zhou, and C. J. Klaassen, "Development and Validation of a
Dynamic Air Handling Unit Model, Part 2," in 2010 ASHRAE Winter Conference,
282
January 23, 2010  January 27, 2010, Orlando, FL, United states, ASHRAE
Transactions (116 PART 1), pp. 5773, 2010.
[66] H. Wang and S. Wang, "A Mechanistic Model of a Centrifugal Chiller to Study
HVAC Dynamics," Building Services Engineering Research and Technology, vol.
21, no. 2, pp. 7383, 2000.
[67] I. N. Grace and S. A. Tassou, "Dynamic Simulation of Liquid Chillers," presented
at the Proceedings of 2000 International Refrigeration Conference at Purdue,
2000.
[68] J. W. MacArthur and E. W. Grald, "Prediction of Cyclic Heat Pump Performance
with a Fully Distributed Model and a Comparison with Experimental Data,"
ASHRAE Transactions, vol. 93, no. 2, 1987.
[69] M. W. Browne and P. K. Bansal, "Transient Simulation of VapourCompression
Packaged Liquid Chillers," International Journal of Refrigeration, vol. 25, no. 5,
pp. 597610, 2002.
[70] K. O. Beatty and D. L. Katz, "Condensation of Vapors on Outside of Finned
Tubes," Chemical Engineering Progress, vol. 44, no. 1, pp. 5570, 1948.
[71] J. C. Chen, "Correlation for Boiling Heat Transfer to Saturated Fluids in
Convective Flow," Industrical Engineering and Chemical Process Design and
Development, vol. 5, no. 3, pp. 322329, 1966.
[72] Z. Lei and M. Zaheeruddin, "Dynamic Simulation and Analysis of a Water Chiller
Refrigeration System," Applied Thermal Engineering, vol. 25, no. 1415, pp.
22582271, 2005.
[73] S. Bendapudi, J. E. Braun, and E. A. Groll, "Dynamic Model of a Centrifugal
Chiller System  Model Development, Numerical Study, and Validation,"
ASHRAE Transactions, vol. 111, pp. 132148, 2005.
[74] S. Bendapudi, J. E. Braun, and E. A. Groll, "A Comparison of MovingBoundary
and FiniteVolume Formulation for Transients in Centrifugal Chillers,"
International Journal of Refrigeration, vol. 31, pp. 14371452, 2008.
[75] S. Wang, Intelligent Buildings and Building Automation, 1st ed.: Spon Press, p.
171, 2009.
283
[76] H. Tan and A. Dexter, "Estimating Airflow Rates in AirHandling Units from
Actuator Control Signals," Building and Environment, vol. 41, no. 10, pp. 1291
1298, 2006.
[77] ASHRAE, "Energy Standard for Buildings Except LowRise Residential
Buildings," American Society of Heating, Refrigerating and AirConditiioning
Engineers, Inc., 1791 Tullie Circle NE, Atlanta, GA 30329,
ANSI/ASHRAE/IESNA Standard 90.12004 2004.
[78] S. T. Taylor and C. H. Cheng, "Why Enthalpy Economizers Don't Work,"
ASHRAE Journal, November, 2010.
[79] M. Hydeman, S. Taylor, J. Stein, E. Kolderup, and T. Hong, "Advanced Variable
Air Volume System Design Guide," California Energy Commission,Octobor 2003.
[80] J. Zhou, G. Wei, W. D. Turner, and D. E. Claridge. Airside Economizer
Comparing Differing Control Strategies and Common Misconceptions. Available:
http://hdl.handle.net/1969.1/90822
[81] J. E. Seem, "Sensorfree Optimal Control of AirSide Economizer," US Patent
Application No. 11/699,859, Publication No. US 2008/0179408 A1, 2008.
[82] M. Leblanc, "Sur lelectrification des Chemins de fer au Moyen de Courants
Alternatifs de Frequence Elevee," Revue Generale de lElectricite, 1922.
[83] H. S. Tsien, Engineering Cybernetics. New York:McGrawHill, pp. xii, 289, 1954.
[84] J. Sternby, "Extremum Control Systems: An Area for Adaptive Control?" in
Preprints of the Joint American Control Conference, ed. San Francisco, CA, 1980.
[85] K. J. strm and B. Wittenmark, Adaptive Control, 2nd ed. Reading, Mass.:
AddisonWesley, pp. xvi, 574, 1995.
[86] M. Krsti and H.H. Wang, "Stability of Extremum Seeking Feedback for General
Nonlinear Dynamic Systems," Automatica, vol. 36, no. 4, pp. 595601, 2000.
[87] M. Krsti, "Performance Improvement and Limitations in Extremum Seeking
Control," Systems and Control Letters, vol. 39, no. 5, pp. 313326, 2000.
[88] W. HsinHsiung, S. Yeung, and M. Krstic, "Experimental Application of
Extremum Seeking on an AxialFlow Compressor," Control Systems Technology,
IEEE Transactions on, vol. 8, no. 2, pp. 300309, 2000.
284
[89] J. Y. Choi, M. Krstic, K. B. Ariyur, and J. S. Lee, "Extremum Seeking Control for
DiscreteTime Systems," IEEE Transactions on Automatic Control, vol. 47, no. 2,
pp. 318323, 2002.
[90] M. A. Rotea, "Analysis of Multivariable Extremum Seeking Algorithms,"
Chicago, IL, USA, Proceedings of the American Control Conference (1), pp. 433
437, 2000.
[91] H.H. Wang, M. Krsti, and G. Bastin, "Optimizing Bioreactors by Extremum
Seeking," International Journal of Adaptive Control and Signal Processing, vol.
13, pp. 651669, 1999.
[92] A. Banaszuk, S. Narayanan, and Y. Zhang, "Adaptive Control of Turbulent Flow
Separation in a Planar Diffuser," in 54th Annual Meeting of the Division of Fluid
Dynamics, American Physical Society, San Diego, CA, November 18  20, 2001.
[93] P. Binetti, K. B. Ariyur, M. Krsti, and F. Bernelli, "Formation Flight
Optimization Using Extremum Seeking Feedback," Journal of Guidance, Control,
and Dynamics, vol. 26, no. 1, 2003.
[94] K. S. Peterson and A. G. Stefanopoulou, "Extremum Seeking Control for Soft
Landing of an Electromechanical Valve Actuator," Automatica, vol. 40, no. 6, pp.
10631069, 2004.
[95] Y. Li, M. A. Rotea, G. T.C. Chiu, L. G. Mongeau, and I.S. Paek, "Extremum
Seeking Control of a Tunable Thermoacoustic Cooler," IEEE Transactions on
Control Systems Technology, vol. 13, no. 4, pp. 527536, 2005.
[96] C. Zhang, A. Siranosian, and M. Krsti, "Extremum Seeking for Moderately
Unstable Systems and for Autonomous Vehicle Target Tracking without Position
Measurements," Automatica, vol. 43, no. 10, pp. 18321839, 2007.
[97] J. Creaby, Y. Li, and J. E. Seem, "Maximizing Wind Turbine Energy Capture
Using Multivariable Extremum Seeking Control," Wind Engineering, vol. 33, no.
4, pp. 361387, June 2009.
[98] P. Lei, "Modeling and Global MPPT Control for Photovoltaic Array with
Variable Shading," IEEE Transactions on Sustainable Energy, vol. 2, no. 3, pp.
348358, 2011.
285
[99] N. J. Killingsworth and M. Krstic, "PID Tuning Using Extremum Seeking: Online,
Model FreePerformance Optimization," IEEE Control Systems Magazine, vol. 26,
no. 1, pp. 7079, 2006.
[100] P. Cougnon, D. Dochain, M. Guay, and M. Perrier, "RealTime Optimization of a
Tubular Reactor with Distributed Feed," AIChE Journal, vol. 52, no. 6, pp. 2120
2128, 2006.
[101] M. Guay, D. Dochain, and M. Perrier, "Adaptive Extremum Seeking Control of
Continuous Stirred Tank Bioreactors with Unknown Growth Kinetics,"
Automatica, vol. 40, no. 5, pp. 881888, 2004.
[102] M. Guay, D. Dochain, M. Perrier, and N. Hudon, "FlatnessBased Extremum
Seeking Control Over Periodic Orbits," Automatic Control, IEEE Transactions on,
vol. 52, no. 10, pp. 20052012, 2007.
[103] D. Dochain, M. Perrier, and M. Guay, "Extremum Seeking Control and its
Application to Process and Reaction Systems: A survey," Mathematics and
Computers in Simulation,vol. 82, no. 3, pp. 369380, 2011.
[104] V. Tyagi, H. Sane, and S. Darbha, "An Extremum Seeking Algorithm for
Determining the Set Point Temperature for Condensed Water in a Cooling
Tower," in Proceedings of the American Control Conference, pp. 11271131,
2006.
[105] H. S. Sane, C. Haugstetter, and S. A. Bortoff, "Building HVAC Control Systems 
Role of Controls and Optimization," in Proceedings of the American Control
Conference, pp. 11211126, 2006.
[106] J. Lebrun. (2004). Variable Speed Fan. Available:
http://cbs.lbl.gov/diagnostics/model_library
[107] W. Wagner, J. R. Cooper, A. Dittmann, J. Kijima, H. J. Kretzschmar, A. Kruse, R.
Mares, K. Oguchi, H. Sato, I. Stocker, O. Sifner, Y. Takaishi, I. Tanishita, J.
Trubenbach, and T. Willkommen, "The IAPWS Industrial Formulation 1997 for
the Thermodynamic Properties of Water and Steam," Journal of Engineering for
Gas Turbines and Power, vol. 122, no. 1, pp. 150180, 2000.
286
[108] P. Li, Y. Li, and J. E. Seem, "Dynamic Modeling and Control of an AirSide
Economizer Using Modelica," in Modelica 2008 Conference, University of
Applied Sciences Bielefeld, Germany, pp. 447464, 2008.
[109] J.P. Bourdouxhe, M. Grodent, and J. Lebrun, Reference Guide for Dynamic
Models of HVAC Equipment: American Society of Heating, Refrigerating & Air
Conditioning Engineers, 1998.
[110] Modelon. (2007). AirConditioning Library Users Manual Version 1.5.
[111] S. Patankar, Numerical Heat Transfer And Fluid Flow, 1
st
ed.: Taylor & Francis,
1980.
[112] D. Limperich, M. Braun, G. Schmitz, and K. Prl, "System Simulation of
Automotive Refrigeration Cycles," in Proceedings of the 4th International
Modelica Conference, Hamburg, March 78, 2005.
[113] R. K. Shah, A. D. Kraus, and D. Metzger, Compact Heat Exchangers. New York:
Hemisphere Pub. Corp, 1990.
[114] M. Fowler, UML Distilled: A Brief Guide to the Standard Object Modeling
Language, 3
rd
ed.: AddisonWesley Professional, 2003.
[115] W. Wagner and A. Pru, "The IAPWS Formulation 1995 for the Thermodynamic
Properties of Ordinary Water Substance for General and Scientific Use," J. Phys.
Chem. Ref. Data vol. Volume 31, no. Issue 2, pp. 387535, June 2002.
[116] FLUIDCAL. (2011). Software FLUIDCAL for the Calculation of Thermodynamic
Properties for a Great Number of Substances. Available: http://www.ruhruni
bochum.de/thermo/Software/Seiten/Fluidcaleng.htm
[117] H. Elmqvist, H. Tummescheit, and M. Otter, "ObjectOriented Modeling of
ThermoFluid Systems," in Proceedings of the 3rd International Modelica
Conference, Linkping, November 34, 2003.
[118] S. Bendapudi, "Development and Evaluation of Modeling Approaches for
Transients in Centrifugal Chillers," Ph.D., Purdue University, pp. 2861, 2004.
[119] T. Pfafferott and G. Schmitz, "Implementation of a Modelica Library for
Simulation of Refrigeration Systems," in Proceedings of the 3rd International
Modelica Conference, Linkping, November 34, pp. 197206, 2003.
[120] Modelon. (2009). Available: http://www.modelon.se/
287
[121] F. Casella and A. Leva, "Modelica Open Library for Power Plant Simulation:
Design and Experimental Validation," presented at the Proceedings of the 3rd
International Modelica Conference, Linkping, 2003.
[122] O. Bauer, "Modeling of TwoPhase Flows with Modelica," M.S., Department of
Automatic Control, Lund Institue of Technology, pp. 2426, 1999.
[123] P. Li, J. E. Seem, and Y. Li, "A New Explicit Equation for Accurate Friction
Factor Calculation of Smooth Pipes," International Journal of Refrigeration, vol.
34, no. 6, pp. 15351541, 2011.
[124] L. Prandtl, Fhrer durch die Strmungslehre, 3
rd
ed.: F. Vieweg & S,
Braunschweig, 1949.
[125] S. Kakac, R. K. Shah, and W. Aung, "Handbook of SinglePhase Convective Heat
Transfer," pp. 16108, 1987.
[126] W. M. Rohsenow, J. P. Hartnett, and Y. I. Cho, Handbook of Heat Transfer, 3 ed.:
McGrawHill Professional, p. 5.20, 1998.
[127] P. K. Swamee and A. K. Jain, "Explicit Equations for PipeFlow Problems,"
Journal of the Hydraulics Division, vol. 102, no. 5, pp. 657664, 1976.
[128] S. E. Haaland, "Simple and Explicit Formulas for the Friction Factor in Turbulent
Pipe Flow," Journal of Fluids Engineering, Transactions of the ASME, vol. 105,
no. 1, pp. 8990, 1983.
[129] T. K. Serghides, "Estimate Friction Factor Accurately," Chemical Engineering
(New York), vol. 91, no. 5, pp. 6364, 1984.
[130] C. F. Colebrook, "Turbulent Flow in Pipes with Particular Reference to the
Transition Region between the Smooth and Rough Pipes Laws," J. Inst. Civ. Eng.,
vol. 11, pp. 133156, 1939.
[131] O. Reynolds, "On the Extent and Action of the Heating Surface of Steam
Boilers," in Lit. Phil. Soc. Manchester, (14), pp. 712, 1874.
[132] A. P. Colburn, "A Method of Correlating Forced Convection HeatTransfer Data
and a Comparison with Fluid Friction," Trans. Am. Inst. Chem. Engs., vol. 29, pp.
174210, 1933.
[133] T. H. Chilton and A. P. Colburn, "Mass Transfer (Absorption) Coefficients," Ind.
Eng. Chem, vol. 26, pp. 11831186, 1934.
288
[134] L. Prandtl, "Eine Beziehung zwischen Wrmeaustausch und
Strmungswiderstand der Flssigkeiten," ZPhysik, vol. 11, pp. 10721078, 1910.
[135] G. I. Taylor, "Conditions at the Surface of a Hot Body Exposed to the Wind,"
British Advisory Committee for Aeronautics, vol. 2, R&M No. 272, pp. 423429,
1916.
[136] T. von Krmn, "The Analogy between Fluid Friction and Heat Transfer," Trans.
Amer. Soc. Mech. Engrs., vol. 61, pp. 705710, 1939.
[137] W. L. Friend and A. B. Metzner, "Turbulent Heat Transfer inside Tubes and the
Analogy among Heat, Mass and Momentum Transfer," J. AIChE, vol. 4, no. 4, pp.
393402, 1958.
[138] Petukhov and Kirillov, "The Problem of Heat exchange in the Turbulent Flow of
Liquids in Tubes," Teploenergetika, vol. 4, no. 4, pp. 6368, 1958.
[139] R. L. Webb, "A Critical Evaluation of Analytical Solutions and Reynolds
Analogy Equations for Turbulent Heat and Mass Transfer in Smooth Tubes,"
Heat and Mass Transfer/Waerme und Stoffuebertragung, vol. 4, no. 4, pp. 197
204, 1971.
[140] S. W. Churchill, "Comprehensive Correlating Equations for Heat, Mass and
Momentum Transfer in Fully Developed Flow in Smooth Tubes," Industrial &
Engineering Chemistry, Fundamentals, vol. 16, no. 1, pp. 109116, 1977.
[141] O. C. Sandall, O. T. Hanna, and P. R. Mazet, "A New Theoretical Formula for
Turbulent Heat and Mass Transfer with Gases or Liquids in Tube Flow," The
Canadian Journal of Chemical Engineering, vol. 58, no. 4, pp. 443447, 1980.
[142] W. M. Kays, M. E. Crawford, and B. Weigand, Convective Heat and Mass
Transfer, 4
th
ed., 2005.
[143] V. Gnielinski, "New Equations for Heat and Mass Transfer in Turbulent Pipe and
Channel Flow," International Chemical Engineering, vol. 16, no. 2, pp. 359368,
1976.
[144] T. von Krmn, "Turbulence and Skin Friction," J. Aerospace Sci, vol. 1, no. 1,
pp. 120, 1934.
289
[145] J. Nikuradse, "Gesetzmssigkeit der turbulenten strmung in glatten Rohren,"
Forsch. Arb. Ing.Wes., vol. No.356; English Translation., NASA TT F_10, 359,
1966, 1932.
[146] H. Blasius, "Das hnlichkeitsgesetz bei Reibungsvorgngen in Flssigkeiten,"
ForschungsArbeit des IngenieurWesens, vol. 131, 1913.
[147] R. Techo, R. R. Tickner, and R. E. James, "An Accurate Equation for the
Computation of the Friction Factor for Smooth Pipes from the Reynolds
Number," Journal of Applied Mechanics, vol. 32, p. 443, June 1965.
[148] J. J. J. Chen, "Systematic Explicit Solutions of the Prandtl and ColebrookWhite
Equations for Pipe Flows," Proceedings of the Institution of Civil Engineers
(London), vol. 79, no. pt 2, pp. 383389, 1985.
[149] C. T. Goudar and J. R. Sonnad, "Explicit Friction Factor Correlation for Turbulent
Flow in Smooth Pipes," Industrial and Engineering Chemistry Research, vol. 42,
no. 12, pp. 28782880, 2003.
[150] A. K. Jain, "Accurate Explicit Equations for Pipe Flow Analysis," J. Hydraul.
Div., vol. 102, pp. 674677, 1976.
[151] G. I. Barenblatt, A. J. Chorin, and V. M. Prostokishin, "Scaling Laws for Fully
Developed Turbulent Flow in Pipes: Discussion of Experimental Data,"
Proceedings of the National Academy of Sciences of the United States of America,
vol. 94, no. 3, p. 773, 1997.
[152] A. Avci and I. Karagoz, "A Novel Explicit Equation for Friction Factor in Smooth
and Rough Pipes," Journal of Fluids Engineering, Transactions of the ASME, vol.
131, no. 6, 2009.
[153] W. M. Rohsenow and J. P. Hartnett, "Forced Convection, Internal Flow in Ducts
" in Handbook of Heat Transfer: McGrawHill, Inc, 1973.
[154] A. Bejan and A. D. Kraus, Heat Transfer Handbook: John Wiley & Sons, Inc.,
2003.
[155] EES. (2008). Engineering Equation Solver. Available:
http://www.fchart.com/ees/ees.shtml
290
[156] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical
Recipes: the Art of Scientific Computing, 3rd ed.: Cambridge University Press,
2007.
[157] J. A. Nelder and R. Mead, "A Simplex Method for Function Minimization,"
Computer Journal, vol. 7, pp. 308313, 1965.
[158] T. B. Drew, E. C. Koo, and W. H. McAdams, "The Friction Factor for Clean
Round Pipes," Trans. AIChE, vol. 28, pp. 5672, 1932.
[159] W. H. McAdams, Heat Transmission, 3rd ed.: McGrawHill, 1954.
[160] G. K. Filonenko, "Hydraulic Resistance in Pipes (in Russian)," Teploenergetika,
vol. 1, no. 4, pp. 4044, 1954.
[161] M. S. Bhatti and R. K. Shah, "Turbulent and Transition Flow Convective Heat
Transfer in Ducts," in Handbook of SinglePhase Convective Heat Transfer, ed
New York: WileyInterscience, 1987.
[162] D. R. Mirth, S. Ramadhyani, and D. C. Hittle, "Thermal Performance of Chilled
Water Cooling Coils Operating at Low Water Velocities," Chicago, IL, USA,
ASHRAE Transactions (99), pp. 4353, 1993.
[163] F. P. Incropera, D. P. DeWitt, T. L. Bergman, and A. S. Lavine, Fundamentals of
Heat and Mass Transfer: Wiley, pp. 71, 371, 2006.
[164] L. Prandtl, Fhrer durch die Strmungslehre, 3 ed. Braunschweig: F. Vieweg, p.
155, 1949.
[165] J. Nikuradse, "Gesetymssigkeit der turbulenten strmung in glatten rhren," Ver.
Dtsch. Ing. Forschungsh, p. 356, 1932.
[166] W. M. Kays and H. C. Perkins, "Forced Convection, Internal Flow in Ducts," in
Handbook of Heat Transfer, W. M. Rohsenow and J. P. Hartnett, Eds., New York:
McGrawHill, pp. 71, 1973.
[167] I. E. Idelchik, Handbook of Hydraulic Resistance: Begell House., 2001.
[168] C. C. Wang, W. L. Fu, and C. T. Chang, "Heat Transfer and Friction
Characteristics of Typical Wavy FinandTube Heat Exchangers," Experimental
Thermal and Fluid Science, vol. 14, no. 2, pp. 174186, 1997.
[169] S. G. Kandlikar, "Thermal Design Theory for Compact Evaporators," in Compact
Heat Exchangers, ed, pp. 245286, 1990.
291
[170] F. C. McQuiston, "Heat Mass and Momentum Transfer Data for Five PlateFin
Tube Transfer Surface," ASHRAE Transactions, Part 1, vol. 84, pp. 266293,
1978.
[171] F. C. McQuiston, "Correlation of Heat, Mass, and Momentum Transport
Coefficients for PlateFinTube Heat Transfer Surfaces With Staggered Tubes,"
ASHRAE Transactions, Part 1, vol. 84, pp. 294309, 1978.
[172] C.C. Wang, Y.C. Hsieh, and Y.T. Lin, "Performance of Plate Finned Tube Heat
Exchangers under Dehumidifying Conditions," Heat Transfer, vol. 119, February
1997.
[173] W. Pirompugd, S. Wongwises, and C.C. Wang, "A TubebyTube Reduction
Method for Simultaneous Heat and Mass Transfer Characteristics for Plain Fin
AndTube Heat Exchangers in Dehumidifying Conditions," Heat and Mass
Transfer/Waerme und Stoffuebertragung, vol. 41, no. 8, pp. 756765, 2005.
[174] A. Klinkenberg and H. H. Mooy, "Dimensionless Groups in Fluid Friction, Heat,
and Material Transfer," Chemical Engineering Progress, vol. 44, no. 1736, 1948.
[175] F. C. McQuiston, J. D. Parker, and J. D. Spitler, Heating Ventilating, and Air
Conditioning: Analysis and Design, 6 ed.: John Wiley & Sons, Inc.
[176] E. R. G. Eckert, Heat and Mass Transfer, with Part A, Heat Conduction and
Appendix of Property Values, by Robert M. Drake, Jr., 2 ed. New York: McGraw
Hill, 1959.
[177] W. M. Rohsenow and H. Y. Choi, Heat, Mass, and Momentum Transfer New
Jersey: PrenticeHall, 1961.
[178] A. Bejan, Convection Heat Transfer 2
nd
ed. New York John Wiley & Sons, Inc.,
1995.
[179] W. K. Lewis, "The Evaporation of a Liquid into a Gas," ASME Transactions, vol.
44, pp. 325340, 1922.
[180] T. Kusuda, "Calculation of the Temperature of a FlatPlate Wet Surface under
Adiabatic Conditions with Respect to the Lewis Relation," vol. 1, pp. 1632, 1965.
[181] L. C. Burmeister, Convective Heat Transfer: John Wiley & Sons, Inc., 1983.
[182] A. T. Prata, "Analogy between Heat and Mass Transfer for Diffusion Driven Non
Isothermal Evaporation," Revista Brasileira Mec, vol. 10, no. 1, pp. 8797, 1988.
292
[183] ASHRAE, ASHRAE Handbook  HVAC Systems and Equipment  IP Edition, pp.
37.1, 37.28, 2008.
[184] W. F. Stoecker and J. W. Jones, Refrigeration and Air Conditioning, 2nd ed., p.
205, 1983.
[185] J. E. Braun, J. W. Mitchell, S. A. Klein, and W. A. Beckman, "Models for
VariableSpeed Centrifugal Chillers," vol. 93(pt1), pp. 17941813, 1987.
[186] A. Beyene, H. Guven, Z. Jawdat, and P. Lowrey, "Conventional Chiller
Performances Simulation and Field Data," International Journal of Energy
Research, vol. 18, pp. 391399, 1994.
[187] J. M. Gordon, K. C. Ng, and H. T. Chua, "Centrifugal Chillers: Thermodynamic
Modelling and a Diagnostic Case Study," International Journal of Refrigeration,
vol. 18, no. 4, pp. 253257, 1995.
[188] M. W. Browne and P. K. Bansal, "SteadyState Model of Centrifugal Liquid
Chillers," International Journal of Refrigeration, vol. 21, no. 5, pp. 343358, 1998.
[189] D. J. Swider, "A Comparison of Empirically Based SteadyState Models for
VaporCompression Liquid Chillers," Applied Thermal Engineering, vol. 23, no.
5, pp. 539556, 2003.
[190] J. Saththasivam and K. C. Ng, "Predictive and Diagnostic Methods for
Centrifugal Chillers," ASHRAE Trans, vol. 114 no. 1, pp. 282287, 2008.
[191] E. Greitzer, "Surge and Rotating Stall in Axial Flow Compressors, Part I:
Theoretical Compression System Model," Journal of Engineering for Power, vol.
98, pp. 190198, 1976.
[192] K. E. Hansen, P. Jrgensen, and P. S. Larsen, "Experimental and Theoretical
Study of Surge in a Small Centrifugal Compressor," Journal of Fluids
Engineering, vol. 103, pp. 391394, 1981.
[193] D. A. Fink, N. A. Cumpsty, and E. Greitzer, "Surge Dynamics in a FreeSpool
Centrifugal Compressor System.," Journal of Turbomachinery, vol. 114, pp. 321
332, 1992.
[194] J. T. Gravdahl and O. Egeland, Compressor Surge and Rotating Stall: Modeling
and Control. London: Springer Verlag, pp. 143160, 1999.
293
[195] T. B. Ferguson, The Centrifugal Compressor Stage. London: Butterworths, pp.
4964, 1963.
[196] N. Watson and M. S. Janota, Turbocharging the Internal Combustion Engine:
MacMillan, pp. 86109, 1982.
[197] J. T. Gravdahl, O. Egeland and S.O. Vatland, "Drive Torque Actuation in Active
Surge Control of Centrifugal Compressors," Automatica, vol. 38, no. 11, pp.
18811893, 2002.
[198] N. A. Cumpsty, Compressor Aerodynamics: Longman, pp. 3944, 1989.
[199] S. Bendapudi, J. E. Braun, and E. A. Groll, "A Comparison of MovingBoundary
and FiniteVolume Formulations for Transients in Centrifugal Chillers,"
International Journal of Refrigeration, vol. 31, pp. 14371452, 2008.
[200] N. C. Lemke, "Untersuchung zweistufiger Flssigkeitskhler mit dem Kltemittel
CO2," Forschungsberichte des Deutschen Klte und Klimatechnischen Vereins
Nr. 73, DKV, Stuttgart, pp. 4243, 2005.
[201] P. J. Marto, "Heat Transfer in Condensation," in Boilers Evaporators &
Condensers, S. Kakac, Ed., ed: WileyInterscience, pp. 525570, 1991.
[202] V. K. Dhir and J. H. Lienhard, "Laminar Film Condensation on Plane and
Axisymmetric Bodies in NonUniform Gravity," International Journal of Heat
Transfer, vol. 93, no. 7, pp. 97100, 1971.
[203] W. Nusselt, "The Condensation of Steam on Cooled Surfaces," Zeitschirift des
Vereines Deutscher Ingenieure, vol. 60, pp. 541546, 1916.
[204] M. Belghazi, A. Bontemps, J. C. Signe, and C. Marvillet, "Condensation Heat
Transfer of a Pure Fluid and Binary Mixture Outside a Bundle of Smooth
Horizontal Tubes. Comparison of Experimental Results and a Classical Model,"
International Journal of Refrigeration, vol. 24, no. 8, pp. 841855, 2001.
[205] M. M. Chen, "An Analytical Study of Laminar Film Condensation: Part 2  Single
and Multiple Horizontal Tubes," Trans. of ASME, Journal of Heat Transfer, vol.
83, pp. 5560, 1961.
[206] D. Q. Kern, "Mathematical Development of Loading in Horizontal Condensers,"
AIChE Journal, vol. 4, pp. 157160, 1958.
294
[207] M. Belghazi, A. Bontemps, and C. Marvillet, "Experimental Study and Modelling
of Heat Transfer during Condensation of Pure Fluid and Binary Mixture on a
Bundle of Horizontal Finned Tubes," International Journal of Refrigeration, vol.
26, no. 2, pp. 214223, 2003.
[208] V. K. Dhir and J. H. Lienhard, "Laminar Film Condensation on Plane and
Axisymmetric Bodies in NonUniform Gravity," International Journal of Heat
Transfer vol. 93, no. 7, pp. 97100, 1971.
[209] W. M. Rohsenow and M. Cambridge, "Heat Transfer and Temperature
Distribution in LaminarFilm Condensation," Transations of the ASME, vol. 78,
pp. 16451648, 1956.
[210] W. Tegethoff, Personal Communications, 2011.
[211] S. K. Wang, Handbook of Air Conditioning and Refrigeration: McGrawHill, p.
13.5, 2001.
[212] R. C. Arora, Refrigeration and Air Conditioning: PrenticeHall of India Pvt.Ltd,
pp. 490492, 2010.
[213] E. Parr, Industrial Control Handbook, 3rd ed.: Industrial Press, Inc., pp. 104106,
1999.
[214] M. Wetter and C. Haugstetter, "Modelica versus TRNSYS  A Comparison
between an EquationBased and a Procedural Modeling Language for Building
Energy Simulation," Proc. of the 2nd SimBuild Conference, August 2006.
[215] W. S. Cleveland, The Elements of Graphing Data. Monterey, California:
Wadsworth Advanced Books and Software, p. 5, 1985.
[216] K. E. Brenan, S. L. Campbell, and L. R. Petzold, Numerical Solution of Initial
Value Problems in DifferentialAlgebraic Equations. Philadelphia, PA: Society
for Industrial and Applied Mathematics (SIAM), p. 118, 1996.
[217] K. R. Jackson and R. SacksDavis, "An Alternative Implementation of Variable
StepSize Multistep Formulas for Stiff ODEs," ACM Transactions on
Mathematical Software (TOMS), vol. 6, no. 3, pp. 295318, 1980.
[218] C. Pantelides, "The Consistent Initialization of DifferentialAlgebraic Systems,"
SIAM Journal of Scientific and Statistical Computing, pp. 213231, 1988.
295
[219] A. Krner, W. Marquardt, and E. D. Gilles, "Getting around Consistent
Initialization of DAE Systems?," Computers & Chemical Engineering, vol. 21, no.
2, pp. 145158, 1997.
[220] R. F. Sincovec, A. M. Erisman, E. L. Yip, and M. A. Epton, "Analysis of
Descriptor Systems Using Numerical Algorithms," IEEE Transactions on
Automatic Control, vol. 26, pp. 139146, 1981.
[221] L. R. Petzold, "Differential/Algebraic Equations Are Not ODEs," SIAM J. Sci.
Stat. Comput., vol. 3, pp. 367384, 1982.
[222] A. Krner, W. Marquardt, and E. D. Gilles, "Computing Consistent Initial
Conditions for DifferentialAlgebraic Equations," Computers & Chemical
Engineering, vol. 16, no. Supplement 1, pp. S131S138, 1992.
[223] J. Albet, J. M. Le Lann, X. Joulia, and B. Koehret, "Operational Policies for the
StartUp of Batch Reactive Distillation Column," in Escape 4 (Symposium Series
133)  IChemE (Symposium Series 133), Dublin, Ireland, March 2830 pp. 6370,
1994.
[224] J. R. C. Vieira and E. C. Biscaia, "Direct Methods for Consistent Initialization of
DAE Systems," Computers & Chemical Engineering, vol. 25, no. 910, pp. 1299
1311, 2001.
[225] B. Leimkuhler, L. R. Petzold, and C. W. Gear, "Approximation Methods for the
Consistent Initialization of Differential Algebraic Equations," SIAM Journal on
Numerical Analysis, vol. 28, no. 1, pp. 205226, 1991.
[226] V. Gopal and L. T. Biegler, "A Successive Linear Programming Approach for
Initialization and Reinitialization after Discontinuities of DifferentialAlgebraic
Equations," SIAM Journal on Scientific Computing, vol. 20, no. 2, pp. 447467,
1998.
[227] E. C. Biscaia and R. C. Vieira, "Heuristic Optimization for Consistent
Initialization Of DAEs," Computers & Chemical Engineering, vol. 24, no. 910,
pp. 21832191, 2000.
[228] M. Gerdts and C. Bskens, "Consistent Initialization of Sensitivity Matrices for a
Class of Parametric DAE Systems," BIT Numerical Mathematics, vol. 42, no. 4,
pp. 796813, 2002.
296
[229] M. Gerdts, "Direct Shooting Method for the Numerical Solution of HigherIndex
DAE Optimal Control Problems," Journal of Optimization Theory and
Applications, vol. 117, no. 2, pp. 267294, 2003.
[230] P. E. Cuille and G. V. Reklaitis, "Dynamic Simulation of Multicomponent Batch
Rectification with Chemical Reactions," Computers & Chemical Engineering, vol.
10, no. 4, pp. 389398, 1986.
[231] R. Kossel, W. Tegethoff, and N. C. Strupp, "Effects of Tool Coupling on
Transient Simulation of a Mobile AirConditioning Cycle," in Proceedings 7th
Modelica Conference, Sep. 2022, Como, Italy, Sep. 2022, pp. 318325, 2009.
[232] S. E. Mattsson, H. Elmqvist, M. Otter, and H. Olsson, "Initialization of Hybrid
DifferentialAlgebraic Equations in Modelica 2.0," in Proceedings of 2nd
International Modelica Conference, Oberpfaffenhofen, Germany, March 1819,
pp. 915, 2002.
[233] H. Tummescheit and J. Eborn, "Chemical Reaction Modeling with
ThermoFluid/MF and MultiFlash," in Proceedings of 2nd International Modelica
Conference, Oberpfaffenhofen, Germany, March 1819, pp. 3139, 2002.
[234] M. J. Moran and H. N. Shapiro, Fundamentals of Engineering Thermodynamics,
5
th
ed.: 5th ed., Wiley, pp. 81, 777, 2003.
[235] Y. Li and J. E. Seem, "Extremum Seeking Control with Actuator Saturation
Control," US Patent Application No. 12/683,883, Publication No. US
2010/0106331 A1, 2010.
[236] P. F. Blackman, "ExtremumSeeking Regulators," in An Exposition of Adaptive
Control, ed: Pergamon Press, 1962.
[237] K. J. strm and B. Wittenmark, Adaptive Control, 2
nd
ed. Reading, Mass.:
AddisonWesley, pp. xvi, 574 p., 1995.
[238] M. A. Rotea, "Analysis of Multivariable Extremum Seeking Algorithms,"
Proceedings of the American Control Conference, vol. 1, pp. 433437, 2000.
[239] Y. Li, M. A. Rotea, G. T. C. Chiu, L. G. Mongeau, and I.S. Paek, "Extremum
Seeking Control of a Tunable Thermoacoustic Cooler," IEEE Transactions on
Control Systems Technology, vol. 13, no. 4, pp. 527536, 2005.
297
[240] ASHRAE, ASHRAE Standard: Ventilation for Acceptable Indoor Air Quality.
Atlanta, GA: American Society of Heating, Refrigerating and AirConditioning
Engineers, Inc., p. 34, 2001.
[241] K. J. strm and B. Wittenmark, ComputerControlled Systems: Theory and
Design, 3
rd
ed. Upper Saddle River, N.J.: Prentice Hall, p. 194, 1997.
[242] K. J. strm and L. Rundqwist, "Integrator Windup and How to Avoid It," in
Proceedings of 1989 American Control Conference, pp. 16931698, 1989.
[243] H. A. Fertik and C. W. Ross, "Direct Digital Control Algorithms with Anti
Windup Feature," ISA Transactions, pp. 317328, 1967.
[244] K. J. strm, "Advanced Control Methods Survey and Assessment of
Possibilities," in Advanced control in computer integrated manufacturing.
Proceedings of the thirteenth annual Advanced Control Conference, H. M. Morris,
et al., Eds., 1987.
[245] G. Franklin, Feedback Control of Dynamic Systems, 5 ed., 2005.
[246] F. E. Cellier and E. Kofman, Continuous System Simulation, 1st ed.: Springer, p.
332, 2006.
[247] A. Savitzky and M. J. E. Golay, "Smoothing and Differentiation of Data by
Simplified Least Squares Procedures," Analytical Chemistry, vol. 36, no. 8, pp.
16271639, 1964.
[248] Johnson Controls Inc., Model YR TwinScrew Compressor Liquid Chillers Design
Level C, 2008.
[249] J. R. Taylor, An Introduction to Error Analysis: the Study of Uncertainties in
Physical Measurements, 2
nd
ed.: University Science Books, p. 58, 1997.
[250] MathWorks. (2010). Documentation of MATLAB Signal Processing Toolbox
(R2010b). Available: http://www.mathworks.com/
[251] J. T. Gravdahl, "Modeling and Control of Surge and Rotating Stall in
Compressors," Ph.D., Dept. of Engineering Cybernetics, NTNU, pp. 8194, 1998.
[252] R. C. Pampreen, "Small Turbomachinery Compressor and Fan Aerodynamics,"
Journal of Engineering for Power, vol. 95, pp. 251256, 1973.
298
Appendices
A. Review of Centrifugal Compressor Modeling in Gravdahl
and Egeland [194]
As described in Chapter 4, the development of the detailed compressor model in this
work follows the work by Gravdahl and Egeland [194]. Difference is noteworthy
between an air compression system and a chiller system such as in this study. For chiller
operation, the fluid properties are in general not constant at the inlet of compressor,
whereas the condition of the inlet air for air compressor is usually constant or at least
varies much slower. Due to its natural closedloop structure, the fluid properties at the
inlet of chillers compressor are timevarying and subject to transient operations such as
startup, shutdown, load and setpoint changes. The book by Gravdahl and Egeland [194]
is largely dedicated to Gravdahls thesis work [251]. The following reviews the major
model equations to complete the chiller modeling in this work. For more details, refer to
Gravdahl and Egeland [194].
Figure A.1 shows the schematic of the velocity profiles at the impeller of a
centrifugal compressor. Plot (a) shows the fluid angle at the inducer. The absolute
velocity of the fluid entering the inducer, denoted by
1
C , can be calculated as
1
c,in 1
m
C
A
=
(A.1)
where m is the compressor mass flow rate,
c,in
is the refrigerant density at the
compressor inlet, A
1
is the cross section area of the impeller eye. The inducer diameter
299
(
1
D ) is determined by averaging the diameters at the inducer tip (
1 t
D ) and the hub casing
(
1 h
D ), respectively.
( )
2 2 2
1 1 1
1
2
t h
D D D = + (A.2)
The tangential velocity of the inducer (
1
U ) is determined based on
1
D
1
1
2
D
U e = (A.3)
(a) (b)
Figure A.1: Fluid velocity angles at the impeller (Reproduced with permission from
Figure 5.2 and Figure 5.3 in Gravdahl and Egeland [194]), where (a) is the fluid velocity
angle at the inducer (inlet). (b) is the fluid velocity angle at the impeller tip (outlet).
Plot (b) illustrates the fluid angle at the impeller tip, from which the compressor
torque can be determined as
2 2 c
m r U t o = (A.4)
where
2
r is the radius at the impeller tip and o is the slip factor.
2 2 2
2 U r Nr e t = = (A.5)
1b
U
1
U
1
1
, C m
U
2
D
2
2b
300
The spool dynamics is modeled as a firstorder differential equation, i.e.
d c
d
J
dt
e
t t = (A.6)
where
d
t and
c
t are the drive torque of the motor and the compressor torque, respectively.
The isentropic efficiency can be determined based on the ideal energy transfer and
various losses at different locations of the compressor. Based on the original definition by
Cumpsty [198], Gravdahl and Egeland [194] extended the calculation of isentropic
efficiency q
is
by considering more loss terms described in Watson and Janota [196] and
Cumpsty [198]:
ideal
is 1 bf
ideal loss
( , )
c v d
h
mU
h h
q q q q q
A
= A A A A
A + A
(A.7)
where Aq
bf
, Aq
c
, Aq
v
and
Aq
d
denote the efficiency degradation due to backflow loss,
impeller tip clearance loss, volute loss, and energy conversion in diffuser, respectively.
Note in Figure A.1(b), the outlet fluid angle
2
usually changes to
2b
at the impeller tip
due to the fluid slip. By assuming the backsweep angle
2b
=90, the ideal energy transfer
Ah
ideal
can be simplified to
2
ideal 2
h U o A = (A.8)
and the main enthalpy loss is defined as
loss ii id fi fd
h h h h h A = A +A +A +A , where
2
1
ii 1
c,in 1
1 cot
2
b
m
h U
A

 
A =


\ .
(A.9a)
2
2 1 2
id
1 c,in 1
1 cot
2
b
D U m
h
D A
o u
 
A =


\ .
(A.9b)
301
2
fi 2 2 2
c,in 1 1
2 sin
i b
l
h Ch m
D A 
A = (A.9c)
2
fd fd
h k m A =
(A.9d)
As shown in Figure A.1(b), the fluid leaves the impeller tip at the angle of
2,
but when it
enters the diffuser, the incidence angle will change to the fixed inlet angle
2b.
Other loss terms in Eq. (A.7) are described as follows. According to Watson and
Janota [196], it is very difficult to find an accurate theoretical model to describe the back
flow loss (
bf
q A ), and a constant value can be utilized, i.e.,
bf
0.03 q A = . The impeller tip
clearance losses (
c
q A ) was adopted from Pampreens work [252].
cl
0.3
c
l
b
q A =
(A.10)
Usually a volute is placed after the diffuser, and is designed to guide the fluid to flow out
of the compressor in an efficient manner. The efficiency degradation due to the volute
loss was estimated by Cumpsty [198] as
0.02 0.05
v
q s A s
(A.11)
Finally, the efficiency degradation of the energy conversion in diffuser (
d
q A ) is
dependent on the actual design. Detailed formula was studied in Watson and Janota [196]
and Cumpsty [198] by considering the pressure recovery coefficient. For simplicity,
Gravdahl and Egeland [194] adopted a constant value.
302
B. Geometric Parameters for Centrifugal Chiller Modeling
Table B.1: Geometric Parameters of the Centrifugal Compressor [194]
Part Description Values
Impeller
Diameter at the inducer tip, D
t1
0.074 m
Diameter at impeller hub casing, D
h1
0.032 m
Diameter at impeller tip, D
2
0.128 m
Inlet relative blade tip angle,
1b
35
Backsweep angle,
2b
90
Mean hydraulic channel diameter, D
i
0.02 m
Impeller tip width, b 0.07 m
Mean channel length of the impeller, l 0.12 m
Momentum of inertia of the compressor, J 0.001 kgm
2
Vaned
Diffuser
Vane inlet angle,
2b
19
Table B.2: Geometric Parameters of the Condenser and Evaporator [118]
Tube Shell
Inlet Diameter Outlet Diameter Length Diameter
Evaporator 0.01606 m 0.01960 m 2.4384 m 0.39288 m
Condenser 0.01554 m 0.01905 m 2.4384 m 0.37760 m
303
Table B.3: Geometric Parameters of the TXV [118]
Parameter Value
dp
min
80 kPa
k
spring
42105 kPa/m
f
0
0.020400
f
1
0.416180
C
d
100 kW/C
304
C. Initial Guess for Chiller Initialization Study
Table C.1: Initial Guesses for Dynamic Chiller Simulation Obtained from the ThreeStep
Preprocessing Scheme
Component Initial Guesses Values
Compressor
(Step 1)
Inlet pressure
1
p 410
5
Pa
Outlet pressure
2
p 810
5
Pa
Coefficient of pseudophysical analogy
initial
k 1
Outlet specific enthalpy
2
h 4.1510
5
J/kg
Condenser
(Step 2)
Outlet pressure
3
p 810
5
Pa
Outlet specific enthalpy
3
h 3.310
5
J/kg
Evaporator
(Step 3)
Inlet pressure
4
p 410
5
Pa
Inlet specific enthalpy
4
h 2.510
5
J/kg
Outlet specific enthalpy
1
h 2.510
5
J/kg
305
D. Simulation Layout of ESC Based Economizer Control
Figure D.1 shows the simulation model layout of the ESC based economizer control
described in Chapter 6. Figure D.2 shows the detailed view of the ESC block in Figure
D.1. The detailed view of the AHU block is shown in Chapter 3. In Chapter 6, the
transition among the three states in the antiwindup ESC simulation was handled by the
TimeTable marked with the green dash line in Figure D.1.
Figure D.1 Simulation model layout of ESC based economizer control
306
Figure D.2: Modelica model layout of the ESC block with antiwindup logic
307
E. Model Development Based on TIL
Figure E.1 shows the library structure of the customized TIL. A Chillers package
was created that includes most of the major components for the centrifugal chiller
simulations in Chapter 4. The Chillers package can be added to newer version of TIL
with small modifications because only the most basic models and interfaces are used
from the root package of TIL (first tree diagram in Figure E.1).
The four major chiller components are all included in the Chillers package. The
Compressors package contains the detailed centrifugal compressor model with the
proposed direct initialization method and the performancebased centrifugal compressor
described in Bendapudis work [118]. The HeatExchangers package contains the finite
volume method based models for the condenser and evaporator that were described in
Chapter 4. The Valves package contains the orifice valve and the thermal expansion
valve models described in Chapter 4. The Cells package contains the objects of
CondRefrigerantCell and EvapRefrigerantCell, which are modified based on the
RefrigerantCell to include condensation or evaporation heat transfer calculation for the
condenser and evaporator, respectively. The RefrigerantTubes package contains the
objects of CondensationTube and the EvaporationTube, where are the basic
component used in the heat exchanger models. The Separators package contains the
VariableRefLevelSeparator, which is the variable refrigerant level (VRL) based heat
exchanger model described in Chapter 4. The Summaries package contains the storage
of some common variables of centrifugal chiller simulations to facilitate further analysis.
308
Finally, the package Testers contains some example and test cases of chiller
simulations
Figure E.1: Library structure of the modified TIL (the customized modeling efforts are
marked with the shaded blocks)
Chillers
TIL
GasComponents
MositAirComponents
LiquidCompoents
RefrigerantComponents
SLEMediumComponents
HeatExchangers
OtherComponents
PressureStateElements
Cells
GeneralTransportPhenomena
Connectors
Utilities
Internals
Examples
Internals
Compressors
LiquidComponents
HeatExchangers
RefrigerantTubes
Separators
Testers
Valves
Summaries
Cells
CentrifugalCompressor
IGVLinear
IGVQuadratic
Condenser
Evaporator
OrificeValve
ThermalExpansionValve
CondRefrigerantCell
EvapRefrigerantCell
...
CondensationTube
EvaporatorionTube
VariableRefLevelSeparator
CentrifugalCompressor
CentrifugalChiller
....
....
....
....
....
....
TestCentrifugalChiller
....
BendapudiCompressor
309
CURRICULUM VITAE
Pengfei Li
Place of Birth: Beijing, China
Education
B. Sc., North China University of Technology, Beijing, July, 2005
Major: Automation
Dissertation Title: Dynamic Modeling and SelfOptimizing Control of Building HVAC
Systems
Journal Publications:
[1] Li, P., Seem, J.E. and Li, Y. Experimental Validation for dp/dt Assumption of Heat
Exchangers in Vapor Compression Refrigeration Cycles, ASME Journal of Heat
Transfer, vol. 134, 114502 (6 pages), November, 2012
[2] Li, P., Seem, J.E. and Li, Y. A New Explicit Equation for Accurate Friction Factor
Calculation of Smooth Pipes, International Journal of Refrigeration, vol. 34, no. 6,
pp. 15351541, September 2011
[3] Li, P., Li, Y. and Seem, J.E. Consistent Initialization of System of Differential
Algebraic Equations for Dynamic Simulation of Centrifugal Chillers, Journal of
Building Performance Simulation, DOI:10.1080/19401493.2010.547949, August,
2011
[4] Li, P., Li, Y. and Seem, J.E., 2010, Efficient Operation of AirSide Economizer
Using Extremum Seeking Control, ASME Journal of Dynamic Systems,
Measurement and Control, vol. 132, no. 3, 031009 (10 pages), May 2010
[5] Li, P., Li, Y. and Seem, J.E. Modelica Based Dynamic Modeling of a ChilledWater
Cooling Coil, HVAC&R Research, vol. 16, no. 1, pp. 3558, January 2010
Major Professor Date
Bien plus que des documents.
Découvrez tout ce que Scribd a à offrir, dont les livres et les livres audio des principaux éditeurs.
Annulez à tout moment.