Article Abt Autumobile

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Proceedings of the ASME 2016 Dynamic Systems and Control Conference

DSCC2016
October 12-14, 2016, Minneapolis, Minnesota, USA

DSCC2016-9769

A MODEL PREDICTIVE CONTROL LAW


FOR A VAPOR COMPRESSION CYCLE SYSTEM∗

Anhtuan D. Ngo† Joshua R. Cory


Air Force Research Laboratory (AFRL) University of Dayton Research Institute
Wright-Patterson Air Force Base, Ohio, 45433 Dayton, Ohio, 45469
Email: anhtuan.ngo@us.af.mi Email: joshua.cory@udri.udayton.edu

ABSTRACT ditional thermal management techniques of alleviating unwanted


During the development of next generation tactical aircraft, heat at the subsystem level will be inadequate due to the in-
thermal management is given significant consideration due to creased usage of power electronics, composite skin with low
higher transient cooling demands, with stricter temperature thermal conductivity, and reduced ram air inlet size and num-
limits along with the smaller size and weight in the cooling ber, coupled with the thin thermal margins due to weight and
system hardware. Traditional control approaches, such as size constraints. For better performance and reliability, a more
proportional-integral-derivative (PID), are sufficient to achieve energy-efficient, more-electric future aircraft is being developed
the desired steady-state error performance for a thermal system with stricter constraints on the available size and weight on the
with no constraints on the control inputs. The traditional subsystems [1]. For the modern aircraft, the electrical loads de-
control techniques may not be well-suited for thermal systems manded by the on-board weapons system can be unpredictable,
with constrained inputs. In this paper, we apply the Model intermittent, short in duration, and high in frequency. Thermal
Predictive Control (MPC) technique on an input-constrained management of the future aircraft is therefore a critical design
thermal system and examine the system performance under a consideration and, if not addressed successfully, can become a
large transient thermal load and control input limits through limiting factor in the vehicle’s performance envelope [2]. Active
the anticipation of the known thermal load. The results include cooling using the Vapor Compression Cycle (VCS) is one way
design and implementation of an MPC controller for a high- to enable the future vehicle to fully utilize its design capabili-
fidelity, nonlinear vapor compression cycle model, as well as ties. For such a system, an integrated method that involves man-
comparison of the MPC results to those of a finely tuned PID agement and control of the movement and disposal of unwanted
controller. thermal energy at the subsystem and system level is necessary in
enabling the vehicle to operate in its full capability envelope. Ac-
tive cooling using the Vapor Compression Cycle System (VCS)
is one way to enable the future vehicle to fully utilize its design
INTRODUCTION capabilities.
Motivation
Until recently, the traditional, passive cooling of aircraft sub-
systems, along with large built-in thermal margins at individual Problem Definition
thermal subsystems, has proven to be sufficient in managing the Thermal management for air vehicles has traditionally been
vehicle’s internally generated heat. For future aircraft, such tra- performed at the individual subsystem level since the pathways
for dissipation of locally generated heat are readily available. As
greater capabilities in terms of size, weight and power are envi-
∗ PUBLIC
sioned by the designers for future air vehicles, the electronic sub-
RELEASE CASE NUMBER: 88ABW-2016-1460
† CorrespondingAuthor

1 Copyright © 2016 by ASME


This work was authored in part by a U.S. Government employee in the scope of his/her employment.
ASME disclaims all interest in the U.S. Government’s contribution.
Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo
systems for communication, control, actuation, life support, etc., tem during the design process in return for requiring upfront a
are becoming more powerful and smaller in size, while light- plant model and a centralized control law implementation. For a
weight, composite airframe structures and skin are less thermally VCS, the MPC technique is more suitable since the system can
conductive. These design factors along with a more stream- have multiple inputs and multiple outputs with a high degree of
lined configuration requirement make active cooling of subsys- couplings among themselves.
tems and integrated thermal management of the vehicle a pri-
mary consideration during the design process. The challenges in
active cooling and integrated thermal management can be seen Gap 2: System Constraints
in meeting the transient cooling demands while being energy ef- Linear models of a physical system do not reflect the ranges
ficient at the same time. One solution toward these challenges of operations of its inputs and outputs. The resultant control law
is to employ low-weight, high-capacity vapor compression cycle derived from such models can command control inputs that are
systems (VCS) to deliver on-demand, intermittent cooling for the outside of the operation ranges. This unwelcome feature of the
vehicle. The VCS in this discussion has a variable compressor, control law in commanding physically infeasible inputs is more
two controllable expansion valves, along with a condenser and likely to occur when the design objective is to optimize an per-
two evaporators. In this paper, we will apply the model predic- formance index. This is because an optimal control law tends
tive control (MPC) technique on a high-fidelity, nonlinear model to generate aggressive input signals that can be near or exceed
of the VCS to track the temperature commands under the control the operating limits of the physical plant inputs. Control input
input constraints. The MPC control law fits well with the above saturation can have a destabilizing effect on the closed loop sys-
design goal of providing the cooling demanded while minimizing tem when it is implemented in hardware. The lack of considera-
compressor work, i.e. energy, under the constraints of the expan- tion for system constraints, as a result, forces the designer to be
sion valve openings, compressor speed and coolant flow rate. In conservative during the design process, and to include additional
an application setting, the use of proportional-integral (PI) con- margins into the hardware during the implementation and testing
trol law is typically sufficient for a slow, weakly coupled VCS of the closed-loop system, creating the unnecessary weight and
dynamics, with additional on-line tuning for performance in the power. The MPC technique allows for the inclusion of the con-
presence of unmodeled system dynamics [3]. Sequential loop straints on the inputs and outputs during the design process, thus
closings for a multi-input, multi-output (MIMO) system with PI eliminating the design conservatism and its associated penalties.
controller to alleviate cross-couplings were used successfully to
control a thermal system [4]. This paper aims to shows the fea- Gap 3: Tool for Controller Tuning
sibility and advantages of the MPC technique in optimizing over An optimal controller for a high-dimensional, multi-input,
the current time steps in the control horizon while keeping the multi-output system can have many design parameters over
future time steps in the prediction horizon in account. which to minimize the performance cost index. The design
parameters are associated with design weightings on the plant
states, its outputs, and inputs. By varying these design weight-
Gap 1: Large Dimension and Dynamics Coupling
ings, the designer attempts to arrive at an optimal controller while
Classical control design for the VCS, such as PI design, re- meeting all the system constraints. The lack of a systematic ap-
lies on the designer’s practical insights into the system’s input- proach to vary the design weights makes it necessary for the de-
to-output dynamics so that appropriate input-output pairings can signer to perform the design weights tradeoff manually. Such
be leveraged. Controllers are sequentially developed to create a effort can be tedious and time-consuming. A more automatic
feedback control loop between each of these input-output pair- technique for controller design parameter selection is discussed
ings one at a time. The input-output pairings may be not be a in this paper to reduce the controller design time.
straightforward one-to-one mapping from the inputs to the out-
puts. Furthermore, in such decentralized methods, the cross-
coupling dynamics between the inputs and the outputs are only Contributions
considered by fine tuning of the control gains when the system In this paper, we seek to address the aforementioned limita-
is simulated with all the control loops. For a system with a tions of the traditional control design approach by using a linear
small number of input-output pairings, this tried-and-tested de- model predictive control approach for a nonlinear, high dimen-
sign approach is tractable. However, the fine tuning task can be sioned vapor compression cycle system enabling the system to
time-consuming and iterative when the system is of high dimen- track temperature commands under the constraints of inputs and
sion [4]. When an analytical model of the system is not readily outputs while minimizing the compressor usage. The design ap-
available, the approach can also be used to design the PI con- proach is model-based and the design tools for system model
troller online for a stable system [5–8]. The MPC technique acquisition, controller design and simulation are from commer-
consider simultaneously all the inputs and outputs of the sys- cially available software. The controller design process can be

2 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


readily applied on other VCS configurations. The multi-input,
multi-output (MIMO) VCS under study exhibits cross-couplings
between its inputs and outputs, with physical constraints on its
control inputs. The contributions can be seen in the suggested
methods to manage the large number of system state, handle
the MIMO cross-couplings, optimize the control objective of
temperature command tracking, and meet the system constraints
while using a systematic approach to select the design parame-
ters.

Paper Outline
Two main points are discussed in the paper: Model descrip-
tion and MPC synthesis and design. The design tools, techniques
employed and model simplification are discussed in the model FIGURE 1. ATTMO VCS Model
description. The design approach discussion includes the MPC
formulation, closed-loop control results for the nonlinear model,
and performance comparison.

MODEL DESCRIPTION
Vapor Compression Cycle High-Fidelity Model
At the Air Force Research Laboratory (AFRL), a Vapor Cy-
cle System Research Facility (VCSRF) was constructed to con-
duct experiments in control and thermal management techniques.
The VCS under study is a simplified version of the VCSRF and
is constructed from the MATLAB Simulink R TM
-based AFRL
Transient Thermal Modeling and Optimization (ATTMO) tool-
box [9–11]. The overall diagram of the system is shown in Fig-
ure 1. The operation and framework of the system can be de-
scribed as follows: heat is rejected to a mixture of 75% propylene
glycol and 25% water from the circulating refrigerant R134A at
the condenser (the red block near the top of Figure 1). Exiting
the condenser, the refrigerant passes through a flow distributer,
which diverts the refrigerant into 2 parallel paths, with each con-
FIGURE 2. Steady state P-h diagram of the VCSRF model
taining an identical electronic expansion valve (called EXV-4
and EXV-5 in [3]). The refrigerant then passes into evaporators
1 and 2 (called Panel 4 and Panel 5 Evaporators in [3]). In these
heat exchangers, heat is added to the refrigerant from the exter- valve openings are from 29.6% to 39.6%. Midpoint settings of
nal fluid polyalphaolefin (PAO), which is set to a constant inlet these control input ranges are used to operate the VCS for 2, 000
temperature and flow rate. Finally, the gaseous refrigerant passes seconds to achieve the steady-state operating point. Using the
into a flow junction, which merges the parallel streams and en- method to linearize and extract a design model discussed in [12],
ters the compressor, then returning to the condenser to restart the the linear model is then extracted. The equilibrium component
cycle. inlet conditions (except where noted) of this simulation are
provided in Table 1, with the pressure-enthalpy diagram of the
cycle in Figure 2.
VCS Linear Model Derivation
For the above VCS, we attempt to find a steady-state Once the nonlinear model has reached its steady state oper-
operating point to obtain the linear model and to design the ating condition, the linear model was generated using the Lin-
MPC. The steady-state operating point is found by considering earize function in the MATLAB Linear Analysis Tool Box [13].
the range of its control inputs. The compressor speed range For the current VCS, the resultant linear model had 132 states, 4
desired is from 7, 500 rpm to 8, 500 rpm, while the expansion outputs made up of Evaporator 1 oil exit temperature, Evaporator

3 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


Component Temperature Pressure Outlet Mass Flow Control Setting
Quality

Refrig. 71.2 ◦ C 1379.1 kPa 0 .1224 kg/s N/A


Condenser
Secondary 26.7 ◦C 248 kPa N/A .3025 kg/s N/A
Refrig. 10.27 ◦ C 418.2 kPa .9675 .0612 kg/s N/A
Evaporators
Secondary 65.6 ◦ C 248 kPa N/A 0.075 kg/s N/A
Expansion N/A 1379.1 kPa N/A N/A 34.6%
Valves
Compressor 10.27 ◦ C 418.2 kPa N/A N/A 8000 rpm

TABLE 1. System equilibrium conditions used for linear model derivation. N/A indicates that a given initial condition is not applicable to the listed
component

2 oil exit temperature, Compressor inlet pressure, Compressor Models Responses to Compressor Command at 800 second
outlet pressure. The 3 control inputs are the compressor’s speed 0.5
ATTMO Model
in revolutions per minute (rpm), Expansion Value 1 percentage Linear Model
opening, Expansion Valve 2 percentage opening. Essential steps
0
have to be taken to derive a model of the VCS that is conducive
toward developing the MPC controller [12]. These steps are dis-
Evaporator Temperature

cussed briefly below. −0.5


1. Scalings of Inputs and Outputs: Due to the different mea-
surement units employed for the inputs and outputs of the VCS,
large discrepancies in magnitudes between the control inputs ex- −1
ist. The differences can be seen between the large compressor
speed (from 7, 500 to 8, 000 rpm) and the small expansion valves −1.5
openings (from 29.6% to 39.6%). To avoid the numerical is-
sues that may arise in the linearization process we make use of
the use of scaling factors. A technique on input scalings given −2
in [12] was applied to extract a linear model.
2. Model Discretization and Reduction Model predictive control
−2.5
design is a iterative, finite-time horizon optimization of the se- 600 800 1000 1200 1400 1600 1800 2000 2200
Time (s)
lected performance index. A discrete-time model of the VCS is
therefore needed to describe the plant dynamics over this speci-
FIGURE 3. Comparison of Evaporator 1 response of the linear VCS
fied time horizon, called the prediction time horizon Tp , so that
model versus the ATTMO VCS Model
optimal control input signals can be calculated for the perfor-
mance index. Open loop analysis of the VCS system suggests
that with the sampling period Ts = 15 second, and the prediction
time horizon Tp = 17 seconds, the system dynamics and con- model.
trol bandwidth can adequately respond to input perturbations. A comparison of the Evaporator 1 temperature responses
As mentioned earlier, the VCS linear model has 132 states. A of the reduced-order linear model against the full-order linear
comparision of the linear and nonlinear models can be seen in model is shown in Figure 4.
Figure 3. With such high number of states, the computational
cost for the optimal control signals can be prohibitive, making it
impractical for hardware implementation. As such, model order Model Predictive Controller Design
reduction is necessary to speed up the calculation for the con- The model predictive controller is a finite horizon, optimal
trol law. In this design, a technique suggested in [12] is adopted controller design characterized by a performance index with re-
to reduce the model order from 132 to 16, after discretizing the ceding time horizons.
The performance index J is a scalar value to be minimized

4 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


Models Responses to Expansion Valve 4 Unit % Openning Command at 800 Second
1.5
Reduced Model
Full Linear Model  CT Qy C 0 x̂(k)
  
ŷT (k)Qy ŷ(k) = x̂T (k) x̂iT (k)

1 | {z } 0 Qi x̂i (k)
xT (k)
| {z } | {z }
Evaporator Temperature

a
Q xa (k)

0.5
In Eq. (1), we can rewrite the first four terms of the summation
as:
0

˜ = xaT (k)Qxk + uT (k)Ru(k) + δ u(k)T R2 δ u(k)


J(k) (2)
−0.5

For the MPC design, the performance index J and the closed-
−1
loop response of the VCS depend heavily on the designer’s
600 800 1000 1200 1400 1600 1800 2000 2200 choices for the design parameters P, Qy , Qi , Ru , and Rδ u . Taking
Time (s)
the terms in order in Eq. (2), the first term defines the cost of the
FIGURE 4. Reduced-Order Model Response to Unit Step Expansion current state, where xa is the state vector x augmented with two
Valve integrator states associated with two temperature tracking errors
e1,2 = Tc1,2 − T1,2 , and Q is a symmetric, positive definite matrix
of weights pertaining to each of the system states. The second
term similarly defines the cost of the control action, with u being
the control vector and R the symmetric, positive definite weight
with the control input signal to the plant. For the VCS controller,
matrix. Finally, the last term penalizes the rate of the control
the performance index J is as follows:
actions, preventing the controller from attempting to move actu-
ators beyond their hardware limits.
k=N p
Setting the weights of the cost function is largely a matter of
J = x̂T (N p )Px̂(N p ) + ŷT (k)Qy ŷ(k) + x̂iT (k)Qi x̂i (k) + · · · determining which control actions are preferable to others. For
∑ example, changing the expansion valve opening requires much
k=1
uT (k)Ru u(k) + δ u (k)Rδ u δ u(k) + sT (k)Ss(k)
T
(1) less energy than changing the compressor speed. Therefore, the
relative weight of the valves would be considerably lower than
that of the compressor. This intuitive approach works for the
second two terms, as all of the components in the vectors have a
The performance index shown in Eq. (1) has a quadratic form,
physical meaning; Eq. (3) displays the weights set for the cost
similar to the linear quadratic regulator formulation. The first
function based on this approach, where ε = 2−52 .
term of the performance index J, x̂T (N p )Px̂(N p ), includes the
final values of the estimated controlled system states x̂(N p ) to
ensure the system states x are well-behaved at the final sample of    
the prediction horizon N p . All the terms within the summation 10 0 0 0 5 0 0 0
are to ensure the output y(k), the integrator states xi (k), control  0 .001 0 0 0 5 0 0 
R= ,R =  (3)
0 2 0

input u(k), and the changes in the control inputs δ u(k) provide 0 0 ε 0 .001 0 
small temperature tracking error, stable system states, small con- 0 0 0 ε 0 0 0 .001
trol signals, and small changes in the control signals respectively.
The parameters s(k) is a slack parameter vector of the same di- This approach becomes more difficult when determining the
mension as the output y(k). The slack parameter s(k) soften the weighting for the state vector because not all of the states have a
constraints on y(k) so that violations are allowed with the penalty physical, intuitive meaning. It would be beneficial, therefore, to
on the performance index J.The designer’s choice of S influences have a way to systematically find a cost function weighting that
the availability of a feasible control u(k). bring about the desired response.
An automatic method to choose these design parameters can fa-
cilitate the overall MPC design process and can be seen as fol-
lows. At each sampling instance K in Eq. (1), we can rewrite the Approach Since the Q matrix must be symmetric, pos-
first term of the summation as: itive definite, a baseline matrix was chosen which would have

5 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


these properties. The simplest version of this would be the iden- A gradient descent method (Eq. (8)) is then used to alter
tity matrix; however, during construction of the system model, the gain vector. However, since the system in question is not an
two states were added representing the integral of the error in analytic expression, F is unknown. In order to estimate ∇F, a
temperature output. For the system involved, the two output finite difference is calculated between the current point and the
temperatures affect one another, meaning it is appropriate to add previous point, as seen in Eq. (9).
cross-terms to these two states.The final structure of the baseline
matrix can be seen in Eq. (4).
wnew = wold − γ∇F (8)
 
100 ... 000
0 1 0
 ... 0 0 0 
0 0 1 ... 0 0 0
  wnew − wold
Qbase =  ... ... ... .. .. .. ..  (4) ∇F = (9)

 . . . .  Pnew − Pold
0 0 0
 ... 1 0 0 
0 0 0 ... 0 2 1
Care must be taken in establishing a value for γ, as allowing
000 ... 012
the second term of Eq. (8) to become too large can cause the sim-
ulation to simply crash. In order to prevent this, γ was selected
In order to set a final value for Q, Qbase can be multiplied by as the minimum of either the inverse of the magnitude of ∇F or
a diagonal matrix, W composed of gains w1 ...wn . In this case, 1. This limits the gradient descent to taking no more than a unit
since the final 2 states’ cost are cross coupled, the final 2 gains step in any direction. Plugging this into Eq. (8) results in Eq.
must be identical to maintain the symmetric, positive definite re- (10).
quirement on Q. This leaves the gain structure as seen in Eq.
(5).
1
wnew = wold − min( , 1)∇F (10)
k∇Fk
W = diag(w)
 
w1 This method still has a few flaws. Most problematic is the
 w2  fact that it can allow w to have values less than 0, violating the
  (5)
w =  ...  positive definite constraint on Q. To prevent this, a check was
 

wn−1 
 established. If any value of wnew goes below 0, it is replaced
wn−1 with half of the previous (positive) value.
Because of the way ∇F is estimated, the vector w essentially
moves back and forth along a line in Rn−1 . This means that occa-
sionally the figure of merit can increase for a number of iterations
before slowly working its way back down. In order to save time,
Q = QbaseW (6) the vector resulting in the best-so-far value of P is saved. If the
algorithm rises 3 times in a row, w is returned to the best-so-far
Gain Tuning Once an initial guess for w is selected, the value, and a new direction is randomly selected for testing. The
model predictive controller is generated via the YALMIP toolbox final overall algorithm can be seen in Eq. (11).
for Matlab. Testing the controller involves running a full nonlin-
ear simulation with a commanded step change in Evaporator 1’s
temperature setpoint. From the resulting data, a figure of merit
(
1
wold − min( k∇Fk , 1)∇F normally
is established. By combining the percent overshoot and settling y= N
time into a vector and taking the 2-norm of the result as seen in wbest + kNk if rising P
( (11)
Eq. (7), a single quantity can be used to represent the quality of yi normally
the controller response. wnew,i =
wold /2 if yi < 0

where N is a random vector in Rn−1 . This process is run


 
% Overshoot ∗ 100
P=k k (7)
Ts iteratively until some stop condition is met. In this case, the stop

6 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


condition dictated that none of the values in w changed more than performance index J in Eq. (1). Without the soft constraints
±10−4 . This is a rather tight condition that, in the future, could setup for the output y(k), the MPC optimization algorithm
be relaxed. The final gains achieved from this method can be will return a no feasible control input u(k) solution when
seen in Eq. (12). y(k) is not within the Ymin and Ymax limits. In Eq. (1), the
positive-semidefinite matrix S is the design parameter that
can be varied to change the amount excursion outside of the
 T allowable range of the output.
11.238 1.4986 .82476 1.5051 1.5232 .84537 . . .
w =  .95575 1.4059 1.5468 1.5542 .79597 1.5601 . . .
1.5464 .93945 6.0408 .80819 .83817 .83817 Ymin − s(k) ≤ y(k) ≤ Ymax + s(k) (14)
(12) Umin ≤ u(k) ≤ Umax (15)
For this VCS, we design the MPC using two receding time hori-
zons:
1. Prediction Horizon N p consists of 17 sampling periods, with CONTROL COMPARISON
each sampling period being 15 seconds long. The control As discussed earlier, the MPC controller is designed and
signal u(k) is calculated to minimize the quadratic perfor- simulated for the nonlinear model of the VCS hardware. For
mance index J in Eq. (1) for N p = 12, subject to the the sys- a baseline comparison, the MPC controller is compared against
tem dynamics constraints, input constraints, input rate con- a cycle-optimized PID controller designed and implemented by
straints, and output constraints: Michalak et al. [3] for this VCS hardware. This proportional-
2. Control Horizon Nc is smaller than the prediction horizon integral-derivative (PID) controller was optimized to reduce the
N p , and consists of 15 sampling periods. It is over the con- tracking errors when the input commands to the oil temperatures
trol horizon Nc that the optimal control input signal u(k) of the evaporators 1 and 2 are given by changing the percentage
found for 0 < k ≤ Nc . For Nc < k ≤ N p , u(k) is set equal opening of the expansion valves 1 and 2. With the PID con-
to u(Nc ), i.e. u(k) = u(15) for k = 16, and 17. By having troller, the desired inlet pressure of the compressor is maximized
Nc < N p by 2 sampling steps, we allows the transient effects by modulating its rotational speed, while the desired outlet dis-
of the last control signal to propagate 2 time steps into the charge pressure is achieved by adjusting the condenser cooling
future. fluid flow. Four PID controller loops were designed separately
3. Constraints: The constraints used on the controller design to track the temperature and pressure commands. For this pa-
synthesis come from differenct sources. The first set of con- per, the performance of the optimized PID controllers serves as a
straints is the controller design synthesis model Σa , shown baseline against which the MPC controller is evaluated. Further-
in Eq. (13), which is composed of the VCS plant dynamics more, the performance of an LQG controller for the same VCS,
augmented with two integral states for the temperature track- detailed in [12] is also used as a basis for comparison to high-
ing errors T4cmd − T4 and T5cmd − T5 respectively. In Eq. (13), light the improvement in performance achieved with the MPC
xa (k) is the augmented VCS model discrete state [x xi ]T at controller whereas the inclusion of the LQG results is to high-
sample time k, Aa is the augmented state matrix, Ba is the light the MPC controller’s ability to look ahead and prepare for
augmented control input matrix, Ca is the augmented output commands that are known apriori, and under constrained control
matrix, Da is augmented feed-through control input matrix, inputs.
Bd is the perturbation input matrix, Dd is the feed through
perturbation input matrix. Tracking Performance

xa (k + 1) = Aa xa (k) + Ba u(k) + Bd v(k) The improvement in tracking performance by the MPC con-
Σa : (13)
ya (k) = Ca xa (k) + Da u(k) + Dd v(k) troller comes from the controller design formulation in which the
computation of the current control input into the plant is made
The second set of constraints, as shown in Eqs. (14 and based on the closed-loop tracking responses of the plant over the
15), is associated with the maximum and minimum values prediction horizon N p for the next 17 sampling periods under the
imposed on outputs ya (k) and inputs u(k) respectively. To constraints defined in Eqs (13, 14, and 15). As such, control
make the output constraints Ymin , and Ymax soft constraints, activities of the expansion valves and the compressor are initi-
a slack variable s(k) is included in Eqs. 14. With such soft ated before the two simultaneous temperature command of 1.5◦
constraints, the ouput y(k) is allowed to go outside of its and 1◦ are given to the evaporators 1 and 2 respectively from
maximum or minimum values so that the MPC optimization t = 500 seconds to t = 1800 seconds in the simulation. In Fig-
algorithm is able to find a feasible control input u(k) solu- ure 6, the compressor can be seen to begin changing its rotational
tion under the penalty on the performance index J. In such speed at around t = 425 seconds in preparation for the step tem-
case, the term sT (k)Ss(k) imposes a non-zero term on the perature commands to come at t = 500 second. Similarly, the

7 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


Overshoot Settling Time
Cntrl MPC LQG PID MPC LQG PID
Evaporator 1 Oil Exit Temperature
15
MPC
EXV1 0.20% 0.60% 3.96% 30 s 68 s 104 s
14.5 LQG
PID
14 EXV2 0.10% 2.49% 3.58% 40 s 48 s 104 s
13.5
Temperature (/ C)

13 Press. 0.50% 0.50% -0.32% 38 s 36 s 106 s


12.5

12 TABLE 2. Controllers Performance Comparison


11.5

11

10.5

10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Actuator Control Activities
Time (s)
From Figure 6, the MPC and LQG controllers appear to have
the speed when the temperture commands are initiated. How-
15
Evaporator 2 Oil Exit Temperature ever, under the MPC controller, the compressor begins speeding
14.5
MPC
LQG
up before the temperature commands start, allowing for better
PID
14 tracking with little overshoot. When compared to the PID con-
13.5
troller, there is a difference of 1600 rpm in the compressor speed.
Temperature (/ C)

13
The modulation of the compressor speeds during the temperature
12.5

12
commands are similar. Looking at the expansion valves time re-
11.5
sponses in Figure 7, there is also an offset by around 2% at the
11 begining under the PID controller vs. the MPC and LQG con-
10.5 trollers at t = 500 seconds. Two step temperature commands of
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
1.5◦ C and 1◦ C to the VCS cause the three controllers to reduce
Time (s)
the openings of the expansion valves as seen in Figure EXVAc-
FIGURE 5. Oil Exit Temperatures Under Different Controllers tion. As expected, the MPC and the LQG controllers regulate
the expansion valves in similar manners while the PID controller
commands to the expansion valves to a different opening posi-
tion. However, all controllers command the expansion valves to
increase their openings by the same amount. Once the step tem-
openings of the Expansion Valves 1 and 2 also begin to change perature commands are removed, both expansion valves return
around t = 425 seconds in anticipation of the step temperature to their steady-state values. Figure 8 shows the pressures of
commands. These control actions by the compressor and the ex- the compressor on the condensor side and evaporator side under
pansion valves allow for the closed-loop VCS under the MPC the three controllers. For the MPC and LQG controllers, these
controller to track the temperature commands with very small pressure time histories are very similar since they have the same
overshoot and undershoot. A comparison of the tracking per- initial trim values and both controllers have the same objective
formances of the three controller can be seen in Figure 5. Un- functions. However, when these pressure responses are com-
der the PID controller, the VCS has more overshoot in tracking pared with that of the PID-controlled plant, there are small differ-
the temperature commands. Temperature overshoots also occur ences in the inlet and outlet pressures of the compressor (418.8
when the VCS is controlled by the LQG controller. Very lit- kPa vs. 386.1 kPa for inlet pressure, and 1362.4 kPa vs. 1378.9
tle overhshoot is observed when the system is controlled by the kPa for the outlet pressure) because the PID controller is tuned
MPC controller. The differences in temperature shoots are due to optimize the inlet pressure. From the time of the initiation of
to the facts that the PID controller is a single-input, single output the temperature commands at t = 500 seconds to the termination
controller with sequential loop closure for each input-output pair of the commands at t = 1750 seconds, the MPC controller and
while the LQG controller is a multi-input, multi-output (MIMO) LQG controller are similar to each other, but different to that of
controller and thus it is able to cross coupling betwen inputs and the PID due to different initial trim state values. A more interest-
outputs. The MPC controller, on the other hand, has the advan- ing feature of the MPC controller is that the pressures at the inlet
tage of a MIMO controller and the ability to include known fu- and outlet of the compressor begin to change in anticipation of
ture disturbances in the current control input computations. Ta- the temperature commands. These anticipatory compressor ac-
ble 2 summarizes the overshoots and undershoots of their time tivities can be seen before t = 500 seconds in Figure 8. After
responses to these temperature step commands. the temperature command inputs return to zero at t = 1750 sec-

8 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


Compressor Speed vs Time Expansion Valve 4 Percent Open vs Time
10000 50
MPC
LQG
PID
MPC 45
LQG
PID

Opening (%)
9500 40

35

9000 30

25
0 500 1000 1500 2000 2500
Speed (rpm)

Time (s)
8500

Expansion Valve 5 Percent Open vs Time


50
MPC
LQG
PID
8000 45

Opening (%)
40

7500 35

30

7000 25
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (s) Time (s)

FIGURE 6. Control Activities of Compressor Speed FIGURE 7. Control Activities of Expansion Valves

onds, the closed-loop plant returns to its acquiescent state under controllers are very similar and are also higher than that of the
all three controllers. closed-loop system under the PID controller. The higher system
COP under MPC and LQG is desirable and is an indirect result of
Coefficient of Performance an optimal controller whose performance index includes control
It is noted that the closed-loop VCS’s coefficient of perfor- signal norms as seen in Eq. (1).
mance (COP) is not directly maximized in the control synthesis
for all three controllers. Instead, tracking performance is opti-
mized by minimizing the resultant errors. The COPs under dif- Conclusion
ferent controllers are analyzed as a fall-out results from the con- In this paper, we have applied the Model Predictive Con-
trollers. The steady-state COP is given as: trol design technique on a constrained linear model of a vapor
compression system. The controller design synthesis include
Q̇loads ṁ1C p (∆T1 ) + ṁ2C p (∆T2 ) a linear model with limited control inputs, limited rates of the
COP = = (16) control inputs, limited system outputs. The MPC controller is
Ẇcomp (ṁc )(h1 − h5 )
a discrete controller with a finite prediction horizon (Np = 17
sampling periods) and a finite control horizon (Nc = 15 sam-
where Q̇loads is the heat rate dissipated by the evaporators, Ẇcomp pling periods). The performance index J includes the norms of
is the power used by the compressor, ṁc is the total refrigerant state values, output values, input values, rates of input values,
flow rate through the compressor, and h1 and h5 are the refrig- and output slack varlues. The linear model of the VCS are rep-
erant enthalpy values at the compressor outlet and inlet, respec- resented as constraints along with other constraints such as the
tively. Using Eq. (9), the COP of the closed-loop VCS under the input constraints, output constraints, and input rate constraints.
MPC, LQG and PID controllers are computed and shown in Fig- While the input constraints are hard constraints, the output con-
ure 9. The COP of the closedloop VCS under the MPC and LQG straints are soft constraints, made possible by the inclusion of the

9 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


Condensor Side Pressure Coefficient of Performance
1400 5

1390 MPC
LQG
1380 PID
4.8
1370
Pressure (kPa)

1360

1350 4.6

1340

1330
4.4
1320 MPC
LQG
1310 PID

4.2
1300
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (s)

COP
4

Evaporator Side Pressure


460
3.8
MPC
450 LQG
PID

440
3.6
Pressure (kPa)

430

420 3.4

410

400 3.2

390

380 3
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (s) Time (s)

FIGURE 8. System Pressures under Different Controllers FIGURE 9. VCS Coefficients of Performance under Different Con-
trollers

slack variables in the performance index J. The controller design


synthesis model includes two additional states corresponding the tion, number AIAA-2010-0287, Jan 2010.
integral states of the tracking errors. The temperature tracking [3] Travis Michalak, Stephen Emo, and Jamie Ervin. Control
performance of the closed-loop system with the MPC controller strategy for aircraft vapor compression system operation.
is compared with those of the same system under an LQG and International Journal of Refrigeration, 48:10–18, Decem-
PID controllers. The MPC-controlled VCS has the best tracking ber 2014.
performance, followed by the LQG-controlled VCS, and PID- [4] Neera Jain, Bin Li, Michael Keir, Brandon M. Hencey, and
controlled VCS respectively. The MPC has a better performance Andrew G. Alleyne. Decentralized feedback structures of
because it is able to look ahead over the prediction horizon to a vapor compression cycle system. IEEE Transactions on
prepare for command input occurring during this time period. Control Systems Technology, 18(1):185–193, January 2010.
[5] Daniel T Pollock, Zehao Yang, John T Wen, Yoav Peles,
and Michael K Jensen. Modeling and control of single and
REFERENCES multiple evaporator vapor compression cycles for electron-
[1] Graham Warwick. Afrl’s invent program tackles aircraft ics cooling. In American Control Conference (ACC), 2013,
system efficiency. Aviation Week & Space Technology, Nov pages 1645–1650. IEEE, 2013.
2012. [6] Bryan P Rasmussen, Andrew G Alleyne, and Andrew B
[2] Eric Walters, Steve Iden, Kevin McCarthy, Marco Am- Musser. Model-driven system identification of transcritical
rhein, Tim OConnell, Brian Raczkowski, Jason Wells, Pe- vapor compression systems. Control Systems Technology,
ter Lamm, Mitch Wolff, Kirk Yerkes, William Borger, and IEEE Transactions on, 13(3):444–451, 2005.
Blane Wampler. Invent modeling, simulation, analysis and [7] Matthew S Elliott and Bryan P Rasmussen. Model-based
optimization. In 48th AIAA Aerospace Sciences Meeting predictive control of a multi-evaporator vapor compression
Including the New Horizons Forum and Aerospace Exposi- cooling cycle. In American Control Conference, 2008,

10 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo


pages 1463–1468. IEEE, 2008.
[8] Neera Jain and Andrew G Alleyne. Thermodynamics-
Based Optimization and Control of Vapor-Compression
Cycle Operation: Control Synthesis. In ASME 2011 Dy-
namic Systems and Control Conference and Bath/ASME
Symposium on Fluid Power and Motion Control, pages
827–834. American Society of Mechanical Engineers,
2011.
[9] Kevin McCarthy, Patrick McCarthy, Ning Wu, Andrew Al-
leyne, Justin Koeln, Soumya Patnaik, Stephen Emo, and
Joshua Cory. Model Accuracy of Variable Fidelity Vapor
Cycle System Simulations. Technical report, SAE Techni-
cal Paper, 2014.
[10] Anthony Puntel, Stephen Emo, Travis Michalak, Jamie
Ervin, Larry Byrd, Victor Tsao, and Thomas Reitz. Refrig-
erant charge management and control for next-generation
aircraft vapor compression systems. Technical Report
2013-01-2241, SAE Technical Paper, September 2013.
[11] Megan Kania, Justin Koeln, Andrew Alleyne, Kevin Mc-
Carthy, Ning Wu, and Soumya Patnaik. A Dynamic model-
ing toolbox for air vehicle vapor cycle systems. Technical
report, SAE Technical Paper, 2012.
[12] Anthuan Ngo, Joshua Cory, Brandon Hencey, and Soumya
Patnaik. A simulink pathway for model-based control of
vapor compression cycles. In ASME 2015 Dynamic Systems
and Control Conference, volume 1, 2015.
[13] The MathWorks Inc. MATLAB and Linear Analysis Tool-
box Release 2013a. Natick, Massachusetts, United States.

11 Copyright © 2016 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/91131/ on 03/15/2017 Terms of Use: http://www.asme.org/abo

You might also like