IFAC and

Control ofCIGRE/CIRED
TransmissionWorkshop on
and Distribution Smart Grids
IFAC and
Control
October ofCIGRE/CIRED
Transmission
11-13, Workshop
Czechon
and Distribution
2016. Prague, Smart Grids
Republic
Control of Transmission and Distribution Smartonline
Available
October 11-13, 2016. Prague, Czech Republic Grids at www.sciencedirect.com
October 11-13, 2016. Prague, Czech Republic

ScienceDirect
IFAC-PapersOnLine 49-27 (2016) 419–424
Data-driven Predictive Control of Micro Gas Turbine Combined Cooling
Data-driven Predictive Control of Micro Gas Turbine Combined Cooling
Data-driven Predictive Heating
Control and
of Micro
PowerGas Turbine Combined Cooling
system
Heating and Power system
Heating and Power system
Xiao Wu*, Jiong Shen*, Yiguo Li*, Junli Zhang*, Kwang Y. Lee**
Xiao Wu*, Jiong Shen*, Yiguo Li*, Junli Zhang*, Kwang Y. Lee**
Xiao Wu*, Jiong Shen*, Yiguo Li*, Junli Zhang*, Kwang Y. Lee**

* Key Laboratory of Energy Thermal Conversion and Control of Ministry of Education, Southeast University, Nanjing,
* Key Laboratory of Energy
210096, China Thermal
(e-mail: Conversion shenj@seu.edu.cn;
wux@seu.edu.cn; and Control of Ministry of Education,230079407@seu.edu.cn).
lyg@seu.edu.cn; Southeast University, Nanjing,
* Key Laboratory of Energy
210096, China Thermal
(e-mail: Conversion shenj@seu.edu.cn;
wux@seu.edu.cn; and Control of Ministry of Education,230079407@seu.edu.cn).
lyg@seu.edu.cn; Southeast University, Nanjing,
210096, China**** Department
(e-mail: of Electrical
wux@seu.edu.cn; and Computer Engineering,
shenj@seu.edu.cn; Baylor
lyg@seu.edu.cn; University,
230079407@seu.edu.cn).
One Bear Department of Waco,
Electrical and Computer Engineering, Baylor University,
One Bear** Place
Place
#97356,
Department of Waco,
#97356,
TX 76798-7356,
Electrical and Computer
TX 76798-7356,
USA (e-mail: Kwang_Y_Lee@baylor.edu)
Engineering,
USA Baylor University,
(e-mail: Kwang_Y_Lee@baylor.edu)
One Bear Place #97356, Waco, TX 76798-7356, USA (e-mail: Kwang_Y_Lee@baylor.edu)
Abstract: Micro gas turbine-based combined cooling, heating and power (MGT-CCHP) system is in an
important Micro
Abstract:
Abstract: direction
Micro
gastoward
turbine-based
gastoward
turbine-based
combined cooling,
the development of smartheating
combined cooling,
and power
buildings
heating
(MGT-CCHP)
and district
and power
system iswhich
energy systems,
(MGT-CCHP)
in an
system iswhich
in an
important direction
provides a direction
clean, highly the
efficient development
and reliable meansof smart buildings
of producing and
energy district energy
for multiple systems,
use. However, the
important
provides aMGT-CCHP toward
clean, highlysystem the
efficient development
and reliable dueof
means smart buildings
of behavior and
producingsuch
energy district energy
for multiple systems,
use. However, which
the
control of
providesofaMGT-CCHP
clean, highlysystem is
efficient a challenge, to its as large thermal inertia, and strong
control
couplingofamong multi-variables. aand
is For reliable due
challenge, means
this reasons, to of behavior
thisits
producingsuch
paper
energy
develops as for multiple
large thermal
a data-driven
use. However,
inertia,
predictive and the
strong
controller
control
coupling MGT-CCHP system
among multi-variables. is a challenge, due to its behavior such as large thermal inertia, and strong
for the MGT-CCHP
coupling system toFor
among multi-variables.
this reasons,
improve this paper performance.
its operating develops a data-driven predictive
The technique of controller
subspace
for the MGT-CCHP
identification is utilized system
to toForimprove
construct
this reasons,
the predictor
this paper performance.
its operating
directly
develops a data-driven
from the The data,
input-output
predictive
technique
which
controller
ofcansubspace
be used
for the MGT-CCHP
identification is future system
utilizedbehavior to
to construct improve
thesystem. its
predictor operating
directly performance.
fromcontroller The
the input-output technique
data, which of subspace
can be used
to estimate
identification the of the The predictive is then designed to regulate the
to estimate theisMGT-CCHP
multi-variable
utilizedbehavior
future to construct
of the
system
thesystem.
under
predictor
the Thedirectly
predictive
input-constraints.
fromcontroller
the input-output
The
data, which
is then designed
effectiveness of the
can be used
to regulate
proposed the
control
to estimate theMGT-CCHP
multi-variable future behavior of the
system undersystem. The predictive controller
the results
input-constraints. is then designed to regulate the
approach is demonstrated
multi-variable MGT-CCHP through
system simulation
under the on an 80kwThe
input-constraints. The
effectiveness
MGT-CCHP
effectiveness
of the proposed
simulator.
of the proposed
control
control
approach is demonstrated through simulation results on an 80kw MGT-CCHP simulator.
© 2016, IFAC
approach
Keyword (International
s:isMicro
demonstrated Federation
through
gas turbine-based of Automatic
simulation
cooling, results
heating Control)
onand Hosting
an 80kw
power by Elsevier
MGT-CCHP
system, Ltd. All
simulator.
predictive rights reserved.
control, data-driven
Keyword
predictives: Micro
control, gas turbine-based
subspace cooling,
identification. heating and power system, predictive control, data-driven
predictive Micro gas
Keywords: control, turbine-based
subspace cooling, heating and power system, predictive control, data-driven
identification.
predictive control, subspace identification. 

1. INTRODUCTION  In Yang (2009), PID control loops are designed for an MGT-
1. INTRODUCTION In Yangunit,
CCHP (2009), PID control
in which the "trialloops
and are designed
error" approach for is
anused
MGT- to
1. INTRODUCTION In Yangunit,
CCHP (2009), PID control
in which the "trialloops
and are designed
error" approach for is
anused
MGT- to
Micro gas turbine-based cooling, heating and power (MGT- tune the controller parameters at the given operating point.
Micro gas turbine-based cooling, heating and power CCHP unit, in which the "trial and error" approach is used to
CCHP)
Micro system
gas is a
turbine-basedpromising
cooling,energy generating
heating and powerunit,(MGT-
which tune the controller
However, when parameters
the set-point atchanges,
the givenstrongoperating point.
oscillation
CCHP) system is a promising energy generating unit,(MGT-
which tune the
However, controller parameters at the given operating point.
has
CCHP) the ability
system to provide
is aprovide
promisinghot water,
energy absorbing cooling
generating cooling and
unit, which occurs
However, in when the set-point changes,
the output-variables,
when the set-point changes, which strong
may reduce
strong
oscillation
the
oscillation
has the
electricity ability to
simultaneously. hot water,
Because the absorbing
thermal energy and
stored occurs in the output-variables, which may reduce the
has the ability to provide Because
hot water, absorbing cooling and efficiency
occurs of
in ofthethe plant and threat
output-variables, the safety
which of
may many devices.
reduce the
electricity
in the exhaust simultaneously.
gas of gas turbine the
can bethermal energy
continually stored
used as efficiency the plant and threat the safety of many devices.
electricity simultaneously. Because the thermal energy stored As a direct
efficiency of approach
the plant to improve
and threat thethe conventional
safety of many PI/PID
devices.
in thesource
exhaust gas ofheater
gas turbine canthebe energy
continually used ofas As a direct aapproach to improve the conventional
heat
in the exhaust of water
gas of gas or chiller,
turbine can be efficiency
continually used as controller,
As a direct aapproachfrequency-domain
to improve the design method PI/PID
conventional based
PI/PID
heat
the source of water
integrated MGT-CCHPheater orsystem
chiller,isthe
up energy
to 80% efficiency
(Colombo, of controller, frequency-domain design method based
heat source of water heater or chiller, the energy efficiency of decoupling
controller, PI PI algorithm
a algorithm is proposed
frequency-domain in Zhang
design (2015),
method which
based
the integrated&MGT-CCHP
Armanasco, Perego, 2007), system
and is up to 80% (Colombo, decoupling is proposed in Zhang (2015), which
the integrated&MGT-CCHP
Armanasco, Perego, system
2007), and ismoreover,
up to 80%
moreover,
although
(Colombo,
although
the
the
can achieve PI
decoupling smaller overshoots
algorithm is and faster
proposed in responses.
Zhang (2015), which
investment and maintenance cost of MGT-CCHP unit is can achieve smaller overshoots and faster responses.
Armanasco, & Perego, 2007),costand of moreover, although the can achievesincesmaller overshootsofand
investment
higher and
compared maintenance
with conventional MGT-CCHP
internal unit
combustion is However, the dynamics thefaster responses.
MGT-CCHP system has
investment
higher and maintenance
compared with cost of MGT-CCHP
conventional internal unit is
combustion However, since the dynamics of the MGT-CCHP system has
engine, compared
the MGT-CCHP system is much small combustion
in size and complex
However, properties such
since the dynamics as large
of thermal
the MGT-CCHP inertia,
systemstrong
has
higher
engine, theflexibly
MGT-CCHP with conventional
system in is much internal
small in size and complex properties such as large thermal inertia, strong
thus can be installed a small residential or coupling
complex among
properties multi-variables
such as and thermal
large unknowninertia,
disturbances,
strong
engine, theflexibly
MGT-CCHP system in is much small in size and coupling
thus can
commercial district. be installed
Therefore, for a
the small residential
purpose of energyor the PID among
coupling approaches
among
multi-variables
which areand
multi-variables and
unknown
devised
unknown on thedisturbances,
basis of
disturbances,
thus can flexibly
commercial district.bereduction
installed for
Therefore, in athesmall residential
purpose of energy or the PID approaches which are devised on the basis of
saving, consumption and environmental protection, separate
the single-input,
PID single-input,
approaches which single-output
are devised(SISO) loops
on loops are no
the basis of
commercial
saving, district. reduction
consumption Therefore,and forenvironmental
the purpose protection,
of energy separate single-output (SISO) are no
employing the MGT-CCHP devices has been promoted as the longer
separate sufficient in meeting
single-input, performance
single-output (SISO) specifications.
loops are no
saving, consumption reduction and environmental protection, longer sufficient in meetingareperformance
employing
primary the MGT-CCHP
direction of leading devices
energy has been promoted
developments (Xu as
et the
al., Advanced
longer control techniques
sufficient called for tospecifications.
in meetingareperformance improve the
employing
primary the MGT-CCHP
direction of leading devices
energy has been promoted
developments (Xu as
et the
al., Advanced
operation ofcontrol
MGT-CCHPtechniques
system. called for tospecifications.
improve the
2010,
primary Wu, & Wang,
direction 2006).
of 2006).
leading energy developments (Xu et al., Advancedofcontrol techniques are called for to improve the
2010, Wu, & Wang, operation MGT-CCHP system.
Current
2010, Wu,studies
& Wang, of the MGT-CCHP unit mainly focus on the
2006). operation
Predictiveofcontrol
MGT-CCHPhas been system.
extensively used in the area of
Currentsystem
static studiesconfiguration
of the MGT-CCHP unit mainly
and operation focus on the
optimization, in Predictive control has been extensively usedhasinshown
the area of
Currentsystem
static studiesconfiguration
of the MGT-CCHP and unit mainly
operation focus on the
optimization, in power generating
Predictive unit
control unit control
has control recently, and
been extensively usedhasinshown
the area to be
of
which,
static the
system capital cost,
configurationoperation profit,
and profit,
operation energy consumption
optimization, in power
effectivegenerating
in overcoming recently,
the problems and
ofand
multi-variable, to be
large
which,
and the capital
environmental cost, operation
protection indices energy
areenergy consumption
taken consumption
into account power
effectivegenerating unit
in constrained control
overcoming system recently,
the problems has shown to be
which,
and the capital
environmental cost, operation
protection profit,
indices are2009).
taken Although
into account inertia
effectiveand in constrained
overcoming system (Wu etof
the problems
multi-variable,
ofal., 2015). Underlarge
multi-variable, the
large
(Savola,
and & Fogelholm,
environmental 2007, Shin
protection et al.,
indices are2009).
taken Although
into account the inertia and
traditional design framework, (Wu et
modeling al., 2015).
is the Under
first the
and
(Savola,
dynamic & Fogelholm,
characteristics 2007,
and Shin
modeling et al.,
of MGT-CCHP the
system inertia and constrained
traditional design system (Wu
framework, et al., 2015).
modeling is the Under
first the
and
(Savola, &
dynamic Fogelholm, 2007,
characteristics and Shin et al.,
modeling of 2009). Although
MGT-CCHP the
system foremost
traditional step in predictive
design framework, controller
modeling design.
is Because
the first the
and
have
dynamic beencharacteristics
studied in theand pastmodeling
few years of (Anvari
MGT-CCHP et al.,system
2015, foremost
input-output stepdata
in of
predictive
a systemcontroller design.
can be easily Because
acquired the
during
have
Rey been studiedthe in the pastapproaches
few years (Anvari et al., 2015, foremost stepdata
in of
predictive controller design. Because the
haveet
Rey
al., 2015),
etbeen
al., studiedthe
2015),
control
in the pastapproaches
control
for the MGT-CCHP
few years (Anvari et al., 2015,
forconventional
the MGT-CCHP
input-output
the operation,data
input-output of
a
a
system
data-driven
system
can
models
can
be
are
be
easily
employed
easily
acquired
acquired
during
for most
duringof
system
Rey are
et al., still
2015), staying in
the control the stage
approaches offorconventional
the MGT-CCHP PID the operation, controllers.
the predictive data-driven However,
models arethe employed
modelingfor most of
procedure
system
control. are still staying in the stage of PID the operation,
thestill
predictive data-driven models are employed for most of
system are still staying in the stage of conventional PID
control. is
the complexcontrollers.
predictive controllers.
However,
and the modeling
However,
the modeling
mismatch
the modeling
procedure
that will greatly
procedure
control. is still complex and the modeling mismatch that will greatly
is still complex and the modeling mismatch that will greatly
Copyright © 2016 IFAC 419
Copyright
2405-8963 © 2016
2016, IFAC
IFAC (International Federation of Automatic Control)419 Hosting by Elsevier Ltd. All rights reserved.
Copyright
Peer review©under
2016 IFAC 419Control.
responsibility of International Federation of Automatic
10.1016/j.ifacol.2016.10.769
2016 IFAC CTDSG
420
October 11-13, 2016. Prague, Czech RepublicXiao Wu et al. / IFAC-PapersOnLine 49-27 (2016) 419–424

Waste Gas Hot Water

Cryogen
Steam
Valve

Recuperator Exhaust Gas

Combustor

—
High-
Fuel Flow Temp
Heater Condensor
Regenerative
Heat Fan
Air Flow Gas Valve Cooling
Water
Generator
~ Electricity High-
Temp
——— Low-
Temp
Compressor Turbine Heat Evaporator Heat
Exchange Exchange

Micro Gas Turbine

Absorption Refrigerating
Machine
Fig. 1. Schematic diagram of the MGT-CCHP process.

degrade the control performance is unavoidable. On the other
hand, since the models are developed from the data, data
itself contains more information than the model developed,
which implies that data can be directly used to build the
controller.
For these reasons, this paper proposes to develop a predictive
controller for the MGT-CCHP system directly from the
input-output data. The subspace identification (SID)
approach is utilized to construct the predictor from the data
without developing the plant model (Kadali, Huang, &
Rossiter, 2003, Wu et al., 2013, 2014). Therefore, the
modeling effort for the conventional model predictive control
(MPC) and resulting model mismatches can be avoided.
The proposed data-driven predictive controller (DDPC) is
implemented in an 80kw MGT-CCHP simulator. The
remainder of this paper is organized as follows: Section II
introduces the MGT-CCHP system and its dynamics. The
DDPC is presented in Section III. Simulation results are
given in Section IV and conclusions are drawn in Section V.
2. SYSTEM DESCRIPTION
The MGT-CCHP system under consideration is composed of
an 80kw regenerative micro gas turbine and a 425 kw double-
effect lithium bromide absorption refrigerating machine. The
compressed air is preheated in the recuperator with the
exhaust gas from gas turbine and then piped into the
combustor, where the mixed fuel gas and air is burned,
producing flue gas in high pressure and temperature. The flue
gas is expanded to drive a micro gas turbine to produce
electrical energy and then sent into the absorption
refrigerating machine as the heat source of cryogen after
heating the compressed air. The cryogen heats the water in
the high temperature heater for hot water supply and then is
chilled in the condenser. After leaving the condenser, the
condensed cryogen absorbs the heat from the water in the
evaporator for refrigeration and is fed into the high
temperature heater for another cycle. Following this way,
besides generating the electricity, the MGT-CCHP system is
also capable of producing 60℃-80℃ hot water and 7℃-
12℃ cooling water from the exhaust gas of micro gas turbine.
The model of this MGT-CCHP unit is developed from the
lumped parameters modular modeling approach based on the
first principles. It is then simplified and used as a simulator
validated in the MATLAB environment. The output variables
of the system are: power output y1 (kw), cold water
temperature y2 ( ℃ ), and hot water temperature y3 ( ℃ ),
control inputs into the system are valve actuator positions
that control the flow rate of fuel, represented as u1;
regenerative heat load, represented as u2; and flow rate of
cryogen steam, represented as u3.
The behavior of the MGT-CCHP system is complex due to
the strong couplings among the multi-variables and the
relatively large thermal inertia property of the absorption
refrigerating machine, which is demonstrated in Fig. 2 for a
step response test when the three inputs are applied
sequentially. Therefore, advanced control techniques are
needed in place of the conventional PI/PID controllers.
3. DATA-DRIVEN PREDICTIVE CONTROL OF THE
MGT-CCHP SYSTEM
Consider the objective function:
T T
J  ( yˆ f  r f ) Q f ( yˆ f  r f )   u f R f  u f (1)

420
 2016 IFAC CTDSG
October 11-13, 2016. Prague, Czech RepublicXiao Wu et al. / IFAC-PapersOnLine 49-27 (2016) 419–424 421

where Q f  Q Tf  0, R f  R Tf  0 are weighting matrices of Detailed procedure for identifying the subspace matrices Lw
and Lu using the SID can be found in the Appendix.
future output and input, respectively, and
T Since the subspace matrices Lw and Lu are constructed
rf   r 
T T T
r r is the desired output
 k 1 k 2 kNy  directly from the input-output data, prediction of the output
trajectory. can be simply achieved by the predictor (2) without further
developing the plant model, therefore the remaining
Using the past output-input data modeling procedure for the MPC and resulting model
T
wp   y
T
y ;
T
u
T
u  , the predictive
T mismatch can be avoided.
 k  N 1 k k  N 1 k
T To achieve an offset-free tracking performance of the data-
output yˆ f   yˆ kT 1 T
yˆ k  2 yˆ k  N  can be estimated by
T
driven predictive controller, an integral action can be
 y 

Kadali, Huang, & Rossiter (2003): included by using a difference operator   1  z  1 in the
output prediction (2):
yˆ f  l w w p  lu u f (2)
 yˆ f  l w  w p  lu  u f (3)
following the basic idea of SID,
T Thus, we can have:
where u f   u kT  1 T
uk 2 uk  N 
T
u 
is the future control
yˆ f  y k   l w  w p   lu  u f (4)
input. In the predictor (2), l w  L w (1 : lN y ,:) and
l u  L u (1 : lN y ,1 : m N u ) are prediction matrices in MATLAB I 0 0
 
T I I 0
expression (i.e., lw is the first lNy rows of Lw and lu is the first where y k   y kT T
yk yk 
T
 and    .
lNy rows and first mNu columns of Lu), where Lw and Lu are  
matrices constructed through SID, Ny and Nu, Ny ≥ Nu are,  
I I I
respectively, the prediction horizon and the control horizon,
m and l, are the dimensions of the output and input, and N is The input magnitude constraint ( u m in , u m ax ) as well as the
the number of row blocks in the data Hankel matrix wp .
input rate constraint (  u min ,  u max ) can be imposed as:
80
I  I 
Power Output (kw)

   
I I
  (u  u )    u    (u  uk ) (5)
m in k f m ax
75    
   
I  I 

70 I  I 
7.3
   
I I
  u   u f     u m ax
Cold Water Temperature (℃ )

m in
(6)
   
7.2    
I  I 
7.1 Substituting (4) into the objective function (1), and at every
sampling time, by minimizing (1) subject to (5) and (6),
7  u f can be calculated, and the value of u k 1 can be obtained
and applied to the plant. The minimization of the objective
function can be solved efficiently through the computational
Hot Water Temperature (℃ )

tools such as quadratic programming (QP).

80
Remark 3.1: As a finite horizon predictive controller, the
79 DDPC proposed in this paper cannot guarantee the stability
even for the un-constrained linear time invariant (LTI)
78 system and the stability depends on a good tuning of
controller parameters, such as the weighting matrices Qf and
0 500 1000 1500 2000 2500 3000 3500 Rf and predictive and control horizons Ny and Nu.
Time (Second)
However, considering the advantages of the DDPC, it is still
Fig. 2. Step response of the MGT-CCHP system. Three a competitive controller for the boiler–turbine unit. How to
outputs show strong coupling when three inputs (fuel flow, develop a stable data-driven predictive controller for the
regenerative heat load, and cryogen steam flow valves) are MGT-CCHP system will be the future research.
decreased by 10% sequentially at 100, 1300, and 2500
second, respectively. 4. SIMULATION RESULTS

421
2016 IFAC CTDSG
422
October 11-13, 2016. Prague, Czech Republic Xiao Wu et al. / IFAC-PapersOnLine 49-27 (2016) 419–424

This section demonstrates the data-driven predictive
controller design for the MGT-CCHP system using SID. The
proposed controller is tested and compared with the regular
MPC and conventional PID controller.
For the proposed predictive controller, the sampling time is
set as 2s, and a prediction horizon Ny =40s and control
separate SISO loops still cannot consider the interactions
of the different variables in the system very well.
Therefore, it is difficult to attain a desired control for all
the output variables.
80

Power Output (kw)

horizon Nu=40s are adopted. The weighting diagonal matrices
75
Q f , R f for all four controllers are given as:

70
 Q1 0 0 0   R1 0 0 0 
   
0 Q2 0 0 0 R2 0 0
Qf   ,R   
 0 0  0 0 

Cold Water Temperature (℃ )

f
0 0
    8
0 0 0 QN   0 0 0 RN 
 y  u 

7.5
with diagonal blocks:
7
 30 0 0  1 0 0
   
Qi  0 50 0 , Rj  0 1 0 , 6.5
   
 0 0 50   0 0 1  Hot Water Temperature (℃ )
80

in which i=1, 2,..., Ny, j=1,2,...,Nu. The input constraints are

u1,2,3,max=1; u1,2,3,min=0 due to the physical limitations of 75

valves. Considering the fast dynamics of the valve, the rate

limitations of control input are ignored. 70

The simulation is devised to show the overall performance of

0 40 80 120 160 200 240 280
the controllers for the load tracking. At t=50s, the set-points
Time (Second)
of power output, cold water temperature, and hot water
Fig. 3. Performance of the MGT-CCHP system: Output Variables (solid red:
temperature are changed from (80kw, 7℃, 80℃) to (70kw,
DDPC; dotted black: decoupling PI; dashed blue: MPC; dotted-dashed
8℃, 70℃) simultaneously, which reflects the most stressed
green: reference).
operation.
Two other controllers are used for comparison: 0.5
Fuel Flow Valve

(1) Regular MPC with the same parameters (the model is 0.4
developed using SID with the same input-output data);
0.3
(2) Decoupling PI controller (Zhang, 2015), with parameters:
0.2
 0.03 
 1  5
GT  0.93  0.84  s 
 
Regenerative Heat Load Valve

K    1
  2.48  7.85  
1
0.1    0.5 
 s 
0.8
 0.05 
1
RM 1 1  s    0.3 
K      0.6
1  1  0.05   5
1
 s  0.4
Cryogen Steam Flow Valve

for the gas turbine and refrigerating machine controllers, 0.56

respectively.
0.52
The simulation results in Figs. 3 and 4 show that both
predictive controllers (DDPC and MPC) have satisfactory
0.48
control performance, which can track the power output and
temperature set-points quickly.
0.44
For the PI controller, although the frequency-domain based 0 40 80 120 160 200 240 280
decoupling approach is utilized in the controller design, the Time (Second)

422
 2016 IFAC CTDSG
October 11-13, 2016. Prague, Czech RepublicXiao Wu et al. / IFAC-PapersOnLine 49-27 (2016) 419–424 423

Fig. 4. Performance of the MGT-CCHP system: Manipulated Variables ( y 0 , y1 , y 2 N  j  2 ) , where N and j are respectively the row
(solid red: DDPC; dotted black: decoupling PI; dashed blue: MPC). p f
and column block numbers of Y and Y . We should
choose N larger than the order of the system n , and number
The two predictive controllers have almost the same of column blocks j should be sufficiently large
performance. However, for the regular MPC, additional (typically j  m ax( m N , lN ) ) to reduce noise sensitivity.
computation for model parameter estimation is needed and,
The input and noise data Hankel matrices, U and E, can be
because the state variables in the model do not have physical
constructed in the similar format.
meanings and is thus unmeasurable, additional observer must
be designed to estimate the state values. Assume that the data is generated by an innovation form of
the state-space model given by:
The simulation clearly demonstrate the effectiveness of the
proposed DDPC which can enhance the operating level of the x k 1  A x k  B u k  K e k
MGT-CCHP system. (A1)
y k  C xk  D u k  ek
5. CONCLUSIONS
where the innovation term e k is assumed to be zero-mean
As an important direction towards the development of smart white noise, and we also assume that the system is
buildings and district energy systems, the MPT-CCHP observable.
system provides a clean, highly efficient and reliable way to
produce heating and cooling and electrical power Then by stacking up the model (A1) with input-output data
simultaneously. In order to overcome the control issues of the for a number of steps, these Hankel matrices can be used to
MPT-CCHP system and improve its operation performance, a develop the following subspace matrix equations (Favoreel et
data-driven predictive control strategy is proposed in this al., 2000; Overschee & Moor , 1995; Qin, 2006):
paper. The technique of SID is utilized to construct the
predictor from the input-output data, thus modeling Y
f
 N X
f
 H Nd U
f
 H Ns E
f
(A2)
procedure for the conventional model predictive control
p p p p
(MPC) can be avoided. The advantages and effectiveness of Y  N X  H Nd U  H Ns E (A3)
the proposed controller design are demonstrated through X
f
 YY
p
 UU
p
 ( A) X
N p
(A4)
the simulations of an 80kw MGT-CCHP system.
T
N 1 T
where  N   ( C ) T (C A )
T
(C A )  ,

ACKNOWLEGEMENT
The authors acknowledge the National Natural Science  D 0 0 0
Foundation of China (NSFC) under Grant 51506029, Grant  
CB D 0 0
51576041 and Grant 51576042, and the Natural Science  
Foundation of Jiangsu Province, China under Grant H Nd   CAB CB D 0 ,
 
BK20150631 and Grant BK20141119 for funding this work.  
C A N 2 B CA
N 3
B CA
N 4
B D 
APPENDIX: IDENTIFYING THE SUBSPACE MATRICES 

The first step of the SID is to build the following data Hankel  I 0 0 0
matrices using the consequent input and output  
CK I 0 0
measurements of the system:  
H Ns   CAK CK I 0,
 
 y0 y1 y j 1   
  C A N 2 K I 
N 3 N 4
y1 y2 yj  CA K CA K
 
 
   Y  ( A)
N 1
K ( A)
N 2
K AK K ,
Y p
  y N 1 y N yN  j2   
Y     .
f
y y N 1 y N  j 1   U  ( A)
N 1 N 2
B
Y   N  
B ( A) B AB

y
 N 1 y N 2
yN  j 
  in which A  ( A  K C ) , B  ( B  K D ) .
 
 y 2 N  1 y2 N y2 N  j2  The state matrix X are defined as:
Here, the output data Hankel matrices Y is partitioned into the X p
  x0 x1 x j 1 
past (upper-half block Y p ) and the future (lower-half block X   f   .
f X   x N x N 1 x N  j  1 
Y ) block matrices and is composed of all the sampled data

423
2016 IFAC CTDSG
October
424 11-13, 2016. Prague, Czech RepublicXiao Wu et al. / IFAC-PapersOnLine 49-27 (2016) 419–424

Owing to the
stability of the Kalman filter, Colombo, L. P. M., Armanasco, F., & Perego, O. (2007).
N N
( A )  ( A  K C )  0 as N   ; thus for a large N , Experimentation on a co-generative system based on a
micro turbine. Applied Thermal Engineering, 27(4), 705-
(A4) converges to:
711.
X
f
 LN W
p
(A5) Favoreel, W., Moor, B.D., & Overschee, P.V. (2000),
p
Subspace state space system identification for industrial
where subspace matrices L N and past data matrices W are processes. Journal of Process Control, 10(2-3), 149–155.
defined as: L N    Y  U  and Kadali, R., Huang, B., & Rossiter A. (2003). A data driven
 subspace approach to predictive controller design.
T
W
p
  (Y )
p T
(U ) 
p T
. Control Engineering Practice, 11, 261-278.
  Liu, X. J., & Kong, X. B. (2013). Nonlinear fuzzy model
Substituting (A5) into (A2), we have: predictive iterative learning control for drum-type
boiler–turbine system. Journal of process control, 23(8),
f p f f
Y  L wW  Lu U  Le E (A6) 1023-1040.
Moon, U., & Lee, K. Y. (2009). Step-response model
with the subspace matrices defined by L w   N L N , development for dynamic matrix control of a drum-
type boiler–turbine system. IEEE Trans. on Energy
L u  H N d and L e  H N s . Conversion, 24(2), 423-430.
Overschee, P.V., & Moor, B.D. (1995), A unifying theorem
With the conditions that (i) u k is uncorrelated with e k , (ii) for three subspace system identification algorithms.
u k is persistently exciting in the order of 2 N , and (iii) the Automotica, 31(12), 1853–1861.
number of measurements is sufficiently large, i.e., j   , Qin, S. J. (2006). An overview of subspace identification.
Computers and Chemical Engineering, 30 (10-12),
the data Hankel matrices developed can be decomposed by 1502–1513.
the QR-decomposition as follows : Rey, G., Ulloa, C., Cacabelos A., & Barragáns B. (2015).
Performance analysis, model development and validation
W p
  R11 0 0   Q1  with experimental data of an ICE-based micro-CCHP
 f    
U    R 21 R 22 0  Q2  (A7) system. Applied Thermal Engineering, 76, 233-244.
Y f  R R 32 R 33   Q 3  Savola, T., & Fogelholm, C. (2007). MINLP optimization
   31   model for increased power production in small-scale
CHP plants. Applied Thermal Engineering, 27(4), 89-99.
By expanding this equation and comparing it with (A6), the
Shin, Y., Seo, J. A., Cho, H. W., Nam, S. C., & Jin, H. J.
subspace matrices L   L w L u  can be calculated as:
 (2009). Simulation of dynamics and control of a double-
effect LiBr-H2O absorption chiller. Applied Thermal
R 0 
†
Engineering, 29(13), 2718-2725.
L   R 31 R 32   11  (A8) Wu, D. W., & Wang R. Z. (2006). Combined cooling,
  R R 22 
 21 heating and power: A review. Progress in Energy and
where † represents the Moore-Penrose pseudo-inverse. The Combustion Science, 32, 459-495.
Wu, X., Shen, J., Li, Y., & Lee, K. Y. (2013). Data-driven
block matrices are identified as L w  L (:,1 : N ( m  l )) and modeling and predictive control for boiler-turbine unit.
L u  L (:, N ( m  l )  1 : end ) in MATLAB expression, IEEE Trans. on Energy Conversion, 28(3), 470-481.
Wu, X., Shen, J., Li, Y., & Lee, K. Y. (2014). Hierarchical
representing the first N ( m  l ) columns and the remaining Optimization of Boiler-Turbine Unit Using Fuzzy Stable
columns in L, respectively. Model Predictive Control. Control Engineering Practice,
30, 112-123.
Note that (A6) has a potential to be used as a predictor in
Wu, X., Shen, J., Li, Y., & Lee, K. Y. (2015). Steam power
designing a predictive controller, because the future output is
plant configuration, design and control. WIREs Energy
expressed as a function of future input. Furthermore, with the
Environ, 4(6), 537-563.
three conditions aforementioned, a predictive expression can
Xu, J., Sui, J., Li, B., & Yang M. (2010). Research,
be written as:
development and the prospect of combined cooling,
Yˆ  L wW
f p
 Lu U
f
(A9) heating, and power systems. Energy, 35(11), 4361-4367.
Yang, J. (2009). Integration and performance simulation
study of combined cooling heating and power system
REFERENCES based on micro turbine. Ph.D. dissertation, Shanghai Jiao
Tong University, Shanghai, China. (in Chinese)
Anvari, S., Taghavifar, H., Saray, R. K., Khalilarya, S., & Zhang, J. (2015). Integrated Optimization and dynamic
Jafarmadar S. (2015). Implementation of ANN on CCHP characteristic research on micro-turbine-based Cooling,
system to predict trigeneration performance with Heating and Power System. Ph.D. dissertation, School of
consideration of various operative factors. Energy Energy and Environment, Southeast University, Nanjing,
Conversion and Management, 101, 503-514. Jiangsu, China. (in Chinese)

424