Hamid Asgari1

Mem. ASME
Department of Mechanical Engineering,
University of Canterbury (UC),
Modeling and Simulation
Christchurch 8140, New Zealand
e-mail: hamid.asgari@pg.canterbury.ac.nz of the Transient Behavior
Mauro Venturini
Mem. ASME
of an Industrial Power Plant
Dipartimento di Ingegneria,
Universita degli Studi di Ferrara,
Via G. Saragat,
Gas Turbine
Ferrara 1-44122, Italy
e-mail: mauro.venturini@unife.it
This study deals with modeling and simulation of the transient behavior of an Industrial
Power Plant Gas Turbine (IPGT). The data used for model setup and validation were
taken experimentally during the start-up procedure of a single-shaft heavy duty gas tur-
XiaoQi Chen bine. Two different models are developed and compared by using both a physics-based
Mem. ASME
and a black-box approach, and are implemented by using the MATLAB# tools including
Department of Mechanical Engineering,
Simulink and Neural Network toolbox, respectively. The Simulink model was constructed
University of Canterbury (UC),
based on the thermodynamic and energy balance equations in MATLAB environment. The
Christchurch 8140, New Zealand
nonlinear autoregressive with exogenous inputs NARX model was set up by using the
e-mail: xiaoqi.chen@canterbury.ac.nz
same data sets and subsequently applied to each of the data sets separately. The results
showed that both Simulink and NARX models are capable of satisfactory prediction, if it
Raazesh Sainudiin is considered that the data used for model training and validation is experimental data
Department of Mathematics and Statistics,
taken during gas turbine normal operation by using its standard instrumentation. [DOI:
University of Canterbury (UC), 10.1115/1.4026215]
Christchurch 8140, New Zealand
e-mail: r.sainudiin@math.canterbury.ac.nz

1 Introduction Besides, the research in this area may be classified based on the
gas turbine type and configuration. The model under investigation
Modeling and simulation of industrial systems, such as gas tur-
can be a micro gas turbine (MGT), an aero gas turbine, or an
bines (GTs), is a conventional methodology for optimization of
IPGT including heavy-duty industrial gas turbines, which are the
design, manufacturing, performance, and trouble shooting. A gas
subject of this study. In the field of modeling and simulation of
turbine model can be employed to clarify details of design strat-
IPGTs, using white-box approach, one can refer to the efforts car-
egies, manufacturing procedures, operating maneuvers, and even
ried out by Rowen [2,3], Najjar [4], Hannett et al. [5], Bettocchi
maintenance guidelines.
et al. [6], Ricketts [7], Bianchi et al. [8], Mostafavi et al. [9], Nag-
Gas turbine models can be categorized into two main groups
pal et al. [10], Kaikko et al. [11], Klang and Lindholm [12], Cam-
including white-box and black-box models. In some cases, the
poreale et al. [13], Al-Hamdan and Ebaid [14], Zhu and Frey [15],
expression of gray-box model may also be used as a combination
Mantzaris and Vournas [16], Yee et al. [17], Lazzaretto and Tof-
of the above mentioned methods. Each of these approaches has its
folo [18], Khosravy-el-Hossani and Dorosti [19], Tavakoli et al.
own characteristics, benefits, and limitations. White-box models
[20], Roldan-Villasana et al. [21], Yadav et al. [22], and Shalan
are used when there is enough information about the physics of
et al. [23]. These researchers employed a variety of methodolo-
the system. They make use of dynamic equations of the system
gies, linearization techniques, and professional software such as
which are usually coupled and nonlinear [1]. In this case, dynamic
MATHEMATICA and Simulink/MATLAB to solve coupled and nonlinear
equations, thermodynamic relationships, energy balance, and lin-
dynamic equations of the system, in order to make reliable and
earization methods are employed to model the system. On the
accurate models.
contrary, black-box models such as neural networks are used
Using black-box approach as a branch of artificial intelligence
when there is not enough knowledge about the physics of the sys-
has opened a new horizon to the area of modeling and simulation
tem. In this case, there is no need to struggle with the complicated
of industrial systems. Black-box methodology is used to disclose
dynamic equations of the system [1]. To make a reliable black-
the relationships between variables of the system using the meas-
box model, different kinds of neural networks can be trained
ured operational data or data generated by means of a simulation
based on the values of different parameters of the system for the
tool. Artificial neural network (ANN), as a data-driven model, is
whole operational range. Before making a decision about the
one of the most significant methods in black-box modeling. ANN
modeling methodology, it is necessary to survey the whole system
includes different approaches such as NARX, adaptive network-
including monitoring system, measureable parameters, system
based fuzzy inference system (ANFIS), feedforward multilayer
history record, sensors health and reliability, accessibility of the
perceptron (MLP), nonlinear autoregressive moving average
system data, technical characteristics, and availability of perform-
with exogeneous inputs (NARMAX), backpropagation neural net-
ance curves. It is also essential to evaluate which methodology is
works (BPNN), radial basis function (RBS), and B-spline. Consid-
more compatible with the research expectations.
erable ANN-based models have been built so far for different
There are many sources regarding modeling and simulation of
kinds of gas turbines including micro gas turbines, aero gas tur-
gas turbines in the literature. As it was mentioned, the GT models
bines, and industrial gas turbines. However, in the area of black-
can be mainly divided into white-box and black-box models.
box models specifically constructed for IPGTs, one can refer to
the research activities carried out by Lazzaretto and Toffolo [24],
1
Corresponding author. Ogaji et al. [25], Arriagada et al. [26], Basso et al. [27], Bettocchi
Contributed by the Controls, Diagnostics and Instrumentation Committee of
ASME for publication in the JOURNAL OF ENGINEERING FOR GAS TURBINES AND POWER.
et al. [28–31], Spina and Venturini [32], Simani and Patton [33],
Manuscript received November 27, 2013; final manuscript received December 8, Yoru et al. [34], Fast et al. [35–38], Fast and Palme [39], and Fast
2013; published online January 9, 2014. Editor: David Wisler. [40].

Journal of Engineering for Gas Turbines and Power JUNE 2014, Vol. 136 / 061601-1
C 2014 by ASME
Copyright V

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms

 Table 1 Gas turbine specifications

GT Number of Rotational Pressure TIT TOT Air flow Power HR Efficiency

type shafts speed (rpm) ratio ( C) ( C) rate (kg/s) (kW) (kJ/kWh) (%)

GE 9351FA 1 3000 15.8 1327 599 648 259500 9643 37.3

Table 2 Time series data sets for different maneuvers

Operational range of the inputs

Maneuver Data sets Number of time-series data sets T01 (K) P01 (Pa) m_ f (kg/s) W_ load (MW)

M1 [0 0 0 1] 1336 [296.48; 301.50] [99,570; 99,909] [3.74; 4.60] ’ 0.3

M2 [0 1 0 1] 1165 [297.04; 303.15] [99,570; 99,670] [3.99; 4.50] [3.30; 18.70]
M3 [0 1 0 1 a] 1506 [290.37; 295.37] [101,940; 102,040] [4.42; 4.80] [4.42; 4.76]
M4 [0 1 0 1 b] 3296 [299.26; 302.59] [100,950; 101,070] [4.27; 6.90] [18.20; 23.70]

As it can be seen from the literature review, a majority of both
white-box and black-box models have been built based on the
steady-state operation of gas turbines when GTs have already
passed the start-up procedure and run in a stable mode. Start-up is
the operating period before the gas turbine reaches stable combus-
tion conditions. Gas turbine start-up procedure can be divided into
four phases including dry cranking, purging, light-off and acceler-
ation to idle [41,42]. In the dry cranking phase, the engine shaft is
rotated by the starting system without any fuel feeding. In the
purging phase, residual fuel from previous operation or failed start
attempts is purged out of the fuel system. In this phase, the rotat-
ing speed is kept constant at a value which ensures a proper mass
flow rate through the combustion chamber, the turbine, and the
heat recovery steam generator. During light-off, fuel is fed to the
combustor, and igniters are energized. This causes ignition to start
locally within the combustor, followed by light-around of all the
burners. Finally, in the acceleration to idle phase, the fuel mass
flow rate is further increased and the rotational speed increases
towards idle value [41,42].
Unfortunately, the literature lacks enough sources about model-
ing and simulation of GT transient behavior during start-up, espe-
cially for IPGTs. Among the limited number of studies covering
this topic, the reader can refer to the works by Agrawal and Yunis
[43], Balakrishnan and Santhakumar [44], Peretto and Spina [45],
Beyene and Fredlund [46], Kim et al. [47–49], Shin et al. [50],
Davison and Birk [51], Huang and Zheng [52], Xunkai and Ying-
hong [53], Sanaye and Rezazadeh [54], Corbett et al. [55], Alo-
baid et al. [56], Daneshvar et al. [57], and Sarkar et al. [58]. An
approximate expression for part-load performance of a microtur-
bine combined heat and power system heat recovery unit was
identified by Rachtan and Malinowski [59]. There are also very
useful white-box and black-box methodologies regarding simula-
tion of transient behavior of individual main components of gas
turbines such as compressors which can be effectively applied to
gas turbines. For instance, one can refer to neural network techni-
ques employed by Venturini [60,61], to explore transient behavior
of compressors. Similar efforts were carried out by Venturini [62]
and Morini et al. [63,64] by using white-box methods.
Because of the importance of transient behavior of gas turbines
during start-up and its direct effect on GT performance and life
time, extensive research is still necessary to fill the existing infor-
mation gaps. To the authors’ best knowledge, there is no Simulink
or NARX model in literature developed to simulate gas turbine
transient behavior during start-up and near full-speed operation.
One of the few examples of such simulation models is docu-
mented by the authors in another paper [65], where a multi-input
single-output NARX model was set up and optimized for the sim-
ulation of the transient behavior of the same gas turbine consid-
ered in the current paper by using data which cover the whole
operational range of the gas turbine during the start-up procedure.
Moreover, another challenge issued in this paper is the use of
field data for model development and testing. Therefore, the set
up and application of these models can help in understanding and
analyzing transient behavior of gas turbines.
For this reason, in this paper, two separate simulation models
using both white-box and black-box methods are built to simulate
very low power operating region for an IPGT. The modeling and
simulation is carried out on the basis of the experimental time-
series datasets obtained from an IPGT located in Italy. The specifi-
cations of the GT are described in Sec. 2. The subject of the Sec. 3
is data acquisition and preparation. Sections 4 and 5 present the
physics-based modeling approach in Simulink and the set up of a
black-box model by using NARX modeling, respectively. The
comparison of all the significant measured and predicted variables
and concluding remarks are presented in Secs. 6 and 7,
respectively.
2 Gas Turbine Specifications
The gas turbine modeled in this research is the General Electric
PG 9351FA, which is a heavy-duty single-shaft gas turbine used
for power generation. The main specifications of this IPGT are
summarized in Table 1.
3 Available Field Data
The data sets used for model set up and verification were taken
experimentally during several start-up maneuvers. The data sets
cover the range 420–3000 rpm. Power is also very low (less than
24 MW, according to Table 2), compared to the nominal power
approximately equal to 260 MW. Therefore, these data are repre-
sentative of the operating conditions during start-up and also
account for all the conditions related to this type of transient oper-
ation (e.g. bleed valve opening, IGV control, etc.).
In general, the data sets during start-up can be categorized as
follows:
• cold start-up: the gas turbine was shut down some day before
start-up
• warm start-up: the gas turbine was shut down some hours
before start-up
• hot start-up: the gas turbine was shut down just few hours or
less before start-up
The data sets used in this paper refer to cold start-up. Moreover,
each of the data sets may fall into different combinations of the
following conditions:
• if the starter is on or off: 1 or 0
• if the gas turbine is connected to the grid or not: 1 or 0
• if customer trip happens or not: 1 or 0
• if the flame is on or off: 1 or 0

061601-2 / Vol. 136, JUNE 2014 Transactions of the ASME

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms

 In this research, the operating conditions sketched below are
considered:
• the starter is off: 0
• the gas turbine is connected/disconnected to/from the grid: 1
or 0
• customer Trip does not happen: 0
• the flame is on: 1
For instance, the maneuver [0 1 0 1] refers to the situation,
when the starter is off, the gas turbine is connected to the grid,
customer trip does not happen, and the flame is on.
In this study, the measured time-series data sets [0 1 0 1] and [0
0 0 1] are called M1 and M2 and are used for Simulink model tun-
ing. The two other data sets including [0 1 0 1 a] and [0 1 0 1 b]
which were used for verification of the models, are specified as
M3 and M4. Table 2 shows more details of these data and the
operational range for the input parameters. The time step for the
data acquisition is one second. As Table 2 indicates, maneuver
M4 is the longest maneuver. Load for M1 is very low and nearly
constant. Figure 1 shows the variations of load for other maneu- Fig. 1 Variations of load for different maneuvers
vers. With respect to M2 and M4, the changes of the load of M3
are smaller.

4 Physics-Based Model of IPGT by Using

MATLAB-Simulink
The Simulink model of the gas turbine was built by using the
operational time-series data sets and by employing the thermody-
namic equations for its components. The data sets were used to
approximate the correlations between corrected parameters in the
compressor and turbine. The approximations were obtained by
using the curve fitting tool in MATLAB. The setup of the Simulink
model consisted of modeling the four main components including
compressor, combustion chamber, turbine and rotational parts dy-
namics. Each component was considered as a single block, stack-
ing dynamic behavior of all the individual stages into a single
block with only the inlet and exit conditions of the component.
The main thermodynamic equations used in the physics-based
model are reported below. The modeling assumptions and the pa-
rameter values will be also discussed in the following:
• Compressor
2 3
 ðcair 1Þ
T01 4 P02 cair
T02 ¼ T01 þ 15
gc P01
cair ¼ CPair =Cvair
W_ c ¼ m_ air ðCPair 02 T02  CPair 01 T01 Þ
4.1 Measured Parameters. The parameters which were
measured directly from the IPGT include: rotational speed (N),
generator power (W_ load ), ambient pressure (P00 ), ambient tempera-
ture (T00 ), compressor inlet stagnation pressure (P01 ), compressor
inlet temperature (T01 ), compressor outlet stagnation pressure
(P02 ), compressor outlet temperature (T02 ), turbine outlet tempera-
ture (T04 ), and fuel flow rate (m_ f ).
4.2 Calculated or Estimated Parameters. To formulate the
correlation between corrected parameters of the compressor and
turbine components of the GT to be used in the Simulink models,
calculation or estimation of some unmeasured parameters was
unavoidable. These parameters can be extracted by employing
thermodynamic relationships or general experimental results
about gas turbines.
4.2.1 Turbine Inlet Stagnation Pressure (P03 Þ. As the Eq. (8)
shows, in practical applications, P03 can be approximated by con-
sidering a linear decrease with respect to P02 . The loss pressure in
combustion chamber (ncc Þis about 3%.
(1)
4.2.2 Turbine Outlet Stagnation Pressure (P04 ). As a practical
routine, turbine outlet stagnation pressure is estimated according
(2) to Eq. (10). C was assumed a constant value and is approximated
(3) as: C ’ 0:03 bar

P04 ’ P00 þ C (10)

• Turbine
2 3 4.2.3 Specific Heat of Air and Gas (Cp and Cv ). Specific heat
 ðcgas 1Þ
P04 c gas of air and gas at constant pressure was calculated based on the
T04 ¼ T03  T03 gt 41  5 (4) fact that it is a function of average temperatures during the com-
P03
pression and expansion processes in the compressor and turbine.
cgas ¼ CPgas =Cvgas (5) After determination of Cp , specific heat of air and gas at constant
volume was calculated using the following equations:
W_ t ¼ m_ gas ðCPgas 03 T03  CPgas 04 T04 Þ (6)
c ¼ Cp =Cv (11)
• Combustor R ¼ Cp  Cv (12)
1 gcc LHV
¼ m_ a =m_ f ¼  1 (7) 4.2.4 Turbine Inlet Temperature ðT03 Þ and Mass Flow Rate of
F Cpgas 03 T03  Cpair 02 T02 Air (m_ a ). When W_ load is approximately constant (i.e., acceleration
P03 ¼ P02 ð1  ncc Þ (8) is zero), Eq. (9) can be written as follows:

• W_ t ¼ W_ c þ W_ load (13)
Equilibrium (balance) equation
  Then, by replacing W_ c from Eq. (3), W_ t can be determined and
W_ t  W_ c  W_ load ¼ ð2p=60Þ2 I NðdN=dtÞ (9) Eq. (6) can be written as follows:

Journal of Engineering for Gas Turbines and Power JUNE 2014, Vol. 136 / 061601-3

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms

 Fig. 2 Block diagram of the Simulink model of the IPGT

Fig. 3 Simulink model of the IPGT

2 3
CP 04 W_ t  ðcairc 1Þ
T03 ¼ T04 þ (14) P 02 air
CPgas 03 m_ gas CPgas gc ¼ T01  4 15=ðT02  T01 Þ (17)
03
P01
In which 2 0 13
 ðcgas 1Þ
P 04
cgas
gT ¼ ðT03  T04 Þ=4T03  @1  A5 (18)
m_ gas ¼ m_ f þ m_ a (15) P03

4.3 Model Architecture. Figure 2 shows the block diagram

Besides, Eq. (7) can be written as
of the gas turbine system. It includes four inputs and four outputs.
The inputs are compressor inlet temperature (T01 ), compressor
CPair 02 gcc  LHV inlet stagnation pressure (P01 ), fuel flow rate (m_ f ), and network
T03 ¼ T02 þ   (16)
CPgas 03 CPgas 03  1 þ m_ a =m_ f load (W_ load ). The outputs consist of rotational speed (N), compres-
sor pressure ratio (PrC ), Compressor outlet temperature (T02 ), and
turbine outlet temperature (T04 ). Figure 2 outlines model inputs
Finally, the equation systems of Eqs. (14) and (16) can be solved and outputs, while Fig. 3 shows the block diagram of the Simulink
for T03 and m_ a . For this purpose, a computer code was written and model in MATLAB environment, to highlight the information flow.
run in MATLAB, and these parameters were calculated for the avail-
able data sets.
4.4 Discussion on Physics-Based Modeling Approach. As
4.2.5 Efficiency and Corrected Parameters of the Compressor outlined in the previous equations, one equation only is used to
and Turbine. The code already written for calculation of T03 and account for system dynamics, i.e., torque balance, since this is
m_ a was developed to calculate efficiency and corrected parameters usually recognized as the most influencing factor on transient
for both the compressor and the turbine. Compressor efficiency behavior. Instead, all the other equations represent steady-state
(gc ) and turbine efficiency ðgT Þwere calculated using the follow- correlations among the different thermodynamic quantities, calcu-
ing equations: lated by using the performance maps obtained by fitting the

061601-4 / Vol. 136, JUNE 2014 Transactions of the ASME

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms

 Fig. 4 Block diagram of complete NARX model of the IPGT
Fig. 5 NARX model of the IPGT
experimental data used for model tuning, through the curve-fitting
tool in MATLAB. This is a key step for tuning the model, which
allows reproducing actual gas turbine behavior, in the considered
operating region. Fitting the experimental data also allows to
smooth measurement uncertainty, which affects the measured
data sets, which, in this paper, were taken by using standard gas
turbine sensors.
Another innovative aspect of the developed model is the itera-
tive procedure adopted for estimating turbine inlet temperature
and inlet mass flow rate, as discussed in Sec. 4.2.4. This procedure
assures that, at any time point, the energy balance is satisfied and
it is in agreement with the instant measured values. Finally, it has
to be stressed the fact that this dynamic model has been developed
and will be validated in the next sections by using field data.
5 NARX Model of IPGT
NARX is a recurrent dynamic network, commonly used in
time-series modeling. NARX includes feedback connections
enclosing several layers of the network. The defining equation of
the NARX model is as follows [66]:
 
yðtÞ ¼ f ðyðt  1Þ; yðt  2Þ; …; y t  ny ;
 uðt  1Þ; uðt  2Þ; …; uðt  nu ÞÞ (19)
in which y is the variable of interest, and u is the externally deter-
mined variable. The next value of the dependent output signal yðtÞ
is regressed on previous values of the output signal and previous
values of an independent (exogenous) input signal. NARX models
can be implemented by using a feedforward neural network to ap-
proximate the function f [65]. NARX networks have many appli-
cations, e.g., they can be used for nonlinear filtering of noisy input
signals or prediction of the next value of the input signal. How-
ever, the most significant application of NARX network is to
model nonlinear dynamic systems [65].
In this study, the neural network toolbox in MATLAB was
employed to tune the NARX models by using measured time-
Journal of Engineering for Gas Turbines and Power
Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms
Fig. 6 Performance of the trained NARX model
series data sets. A NARX model was trained separately for M1 to
predict outputs for [0 0 0 1] start-up condition. For the whole
time-series data sets related to M2, M3, and M4 maneuvers,
another NARX model was trained and the final NARX models
were obtained after trial-and-error efforts for getting reliable and
accurate models in terms of accuracy of the trends and root mean
square errors (RMSE) for output parameters. The resulting model
was tested against each of the M2, M3, and M4 maneuvers sepa-
rately. Inputs and outputs of the NARX models are the same cor-
responding parameters as in the Simulink model. Figures 4 and 5
show the block diagram and the closed-loop structure of the
NARX model in MATLAB environment. As it can be seen from the
Fig. 5, the best result for both NARX models is related to net-
works with one hidden layer with nine neurons, using Leven-
berg–Marquardt backpropagation (trainlm) as the training
function, and a tapped delay line with delays from 1 to 2 at the
input. This means that the NARX model makes use of the
regressed outputs y(t–1) and y(t–2) at time points (t–1) and (t–2).
Figure 6 shows the details of the final trained network. Perform-
ance of the NARX for training, validation, and test is also shown
in the figure. As it can be seen, thirteen iterations were required so
that the validation performance error reached the minimum. The
mean square error (mse) of the performance at this point was quite
low (less than 0.014). The training continued for six more itera-
tions before the training stopped. The term “Epoch” in the figure
is a neural network term for iteration in a training process. The
number of epochs shows the number of times that all the patterns
JUNE 2014, Vol. 136 / 061601-5
 Fig. 7 Variations of rotational speed for the maneuver M1 for Fig. 11 Variations of rotational speed for the maneuver M2 for
the real system, Simulink model, and NARX model the real system, Simulink model, and NARX model

Fig. 8 Variations of compressor pressure ratio for the maneu-

ver M1 for the real system, Simulink model, and NARX model Fig. 12 Variations of compressor pressure ratio for the maneu-
ver M2 for the real system, Simulink model, and NARX model

Fig. 9 Variations of compress outlet temperature for the ma- Fig. 13 Variations of compress outlet temperature for the ma-
neuver M1 for the real system, Simulink model, and NARX neuver M2 for the real system, Simulink model, and NARX
model model

Fig. 10 Variations of turbine outlet temperature for the maneu- Fig. 14 Variations of turbine outlet temperature for the maneu-
ver M1 for the real system, Simulink model, and NARX model ver M2 for the real system, Simulink model, and NARX model

061601-6 / Vol. 136, JUNE 2014 Transactions of the ASME

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms

 Fig. 15 Variations of rotational speed for the maneuver M3 for Fig. 19 Variations of rotational speed for the maneuver M4 for
the real system, Simulink model, and NARX model the real system, Simulink model, and NARX model

Fig. 16 Variations of compressor pressure ratio for the maneu- Fig. 20 Variations of compressor pressure ratio for the maneu-
ver M3 for the real system, Simulink model, and NARX model ver M4 for the real system, Simulink model, and NARX model

Fig. 17 Variations of compress outlet temperature for the ma-

neuver M3 for the real system, Simulink model, and NARX Fig. 21 Variations of compress outlet temperature for the ma-
model neuver M4 for the real system, Simulink model, and NARX
model

Fig. 18 Variations of turbine outlet temperature for the maneu- Fig. 22 Variations of turbine outlet temperature for the maneu-
ver M3 for the real system, Simulink model, and NARX model ver M4 for the real system, Simulink model, and NARX model

Journal of Engineering for Gas Turbines and Power JUNE 2014, Vol. 136 / 061601-7

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms


Fig. 23 RMSE (%) of the Simulink and NARX models for four
main selected outputs of the all maneuvers
are presented to the neural network. The parameter used for the
comparison of the measured data ym to the predictions of the mod-
els, i.e., y is the RMSE defined according to Eq. (20),where nd is
the number of data sets for each curve (see Table 2).
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u nd 
u 1 X ym  y 2
RMSE ¼ t (20)
nd i¼1 ym
6 Results
The trend over time of the prediction of the two simulation
models (physics-based model developed in Simulink and NARX
model) for the four outputs (rotational speed, pressure ratio, com-
pressor outlet temperature, turbine outlet temperature) is com-
pared to the trend over time of measured data. The comparison is
made both for the “training” curves M1 and M2 (Figs. 7–10 and
11–14, respectively) and for the curves M3 and M4 used to assess
the generalization capability of the simulation models (curve M3
in Figs. 15–18 and curve M4 in Figs. 19–22).
Figure 23 summarizes the results in terms of RMSE for the all
maneuvers, to allow a synoptic view.
It should be noted that, during the simulation phase, the NARX
model is fed with the regressed outputs at time points y(t–1) and
y(t–2) estimated by the NARX model itself at antecedent time
steps.
As it can be seen, the both models reproduce M1 very accu-
rately. The maximum RMSE of Simulink and NARX models are
2.2 and 0.96% (both for T04 ), respectively. M2 is also reproduced
with a satisfactory prediction. The maximum RMSE of Simulink
and NARX models for this maneuver are 4.3 and 2.1% (both for
PrC ). M3 is also simulated with acceptable accuracy. The maxi-
mum RMSE of Simulink and NARX models are 3.9 and 2.8%
(both for PrC ), respectively. The results for M4 are also satisfac-
tory enough for prediction of GT dynamics. The maximum errors
of Simulink and NARX models for this maneuver are 4 and 1.7%,
both for PrC .
Overall, the results show that both Simulink and NARX models
can simulate and predict dynamic behavior of the gas turbine with
acceptable accuracy. However, in this study, the NARX model
showed more accuracy compared to the Simulink model. It can be
also noticed that despite higher errors at the beginning of simula-
tion until the stabilization of the response, the RMSE seemed sat-
isfactory in the Simulink model in all cases for all the outputs.
7 Conclusions
In this study, Simulink and NARX models of a heavy-duty sin-
gle-shaft power plant gas turbine were set up for simulating two
different transient maneuvers in the very low power operating
region. The main objective of the research was to explore and pre-
dict transient behavior of IPGTs. Thermodynamic and energy bal-
061601-8 / Vol. 136, JUNE 2014
Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms
ance equations were employed to model the gas turbine in
Simulink/MATLAB environment. Correlations between corrected
parameters of the gas turbine components (compressor and tur-
bine) were investigated by using measured data sets and by using
curve-fitting tool in MATLAB. The same time-series data sets were
employed to build NARX model for the IPGT. To verify the mod-
els, the resulting Simulink and NARX models were tested against
two other time-series data sets. For this purpose, four important
outputs from the IPGT models and their corresponding values
from the measured data sets were compared and the related results
were captured and figured. These outputs were rotational speed,
compressor pressure ratio, compressor outlet temperature, and tur-
bine outlet temperature. The results showed that both Simulink
and NARX models successfully captured dynamics of the system.
They provided satisfactory prediction of the dynamic behavior of
the gas turbine for the studied maneuvers. The results can be con-
sidered satisfactory since the data used both for model training
and validation is experimental data taken during GT normal oper-
ation by using its standard instrumentation. It can be noticed that
despite higher errors at the beginning of simulation until the stabi-
lization of the response, the RMSE were satisfactory in the Simu-
link model. The RMSE of rotational speed, pressure ratio,
compressor outlet temperature, and turbine outlet temperature for
all maneuvers were equal or less than 0.8, 4, 1.6, and 3.1%,
respectively. The maximum error of the Simulink and NARX
models were, respectively, 4.3 and 2.8%.
The simplicity of the thermodynamic model developed above is
one of strong points of this research. In fact, accurate modeling
would have required the knowledge of bleed flows and IGV con-
trol. Since the information about bleed flows is usually unknown
(they are confidential manufacturer’s data) and IGV control logic
is not always known (in this case, it was not available, since this
information is confidential manufacturer’s data as well), this lack
of information was overcame by implicitly accounting for these
effects (which mainly affect the inlet mass flow rate) by means of
two innovative procedures:
• The performance maps, which relate the corrected parame-
ters, were obtained directly from measured data (only from
the “training” data sets M1 and M2, not from the
“verification” data sets M3 and M4) by using the curve-
fitting tool available in MATLAB. The fine tuning of these cor-
relations represented a key and challenging phase of the ther-
modynamic model set up.
• The inlet mass flow rate and the turbine inlet temperature
were estimated at each time step by means of an iterative
procedure in Eqs (14)–(16). A specific MATLAB routine was
written and dedicated to this calculation.
The choice of developing a Neural Network model to cover this
range of operation goes in the same direction, i.e., developing a
simple model to reproduce a very complicated and usually diffi-
cult-to-model unsteady behavior. The results from this research
indicated that NARX approach modeled gas turbine behavior with
higher accuracy compared to Simulink approach. It was shown
that, despite all controversial issues regarding using artificial intel-
ligence and black-box methods for modeling of industrial systems,
ANN can be still considered a reliable and powerful tool in identi-
fication of systems dynamics. The upcoming research outputs in
modeling and simulation of gas turbines and the start-up proce-
dures can help to design, manufacture, operate and maintain GTs
in desirable conditions. The future efforts in this area can lead to
manufacturing gas turbines with minimum energy consumption
and therefore maximum efficiency.
Acknowledgment
This research is supported by the Premier Scholarship in
Department of Mechanical Engineering at University of Can-
terbury (UC), Private Bag 4800, Christchurch 8140, New
Zealand.
Transactions of the ASME
 The authors gratefully acknowledge Professor Pier Ruggero [9] Mostafavi, M., Alaktiwi, A., and Agnew, B., 1998, “Thermodynamic Analysis
Spina, Professor Michele Pinelli, and Dr. Mirko Morini for their of Combined Open-Cycle Twin-Shaft Gas Turbine (Brayton Cycle) and
Exhaust Gas Operated Absorption Refrigeration Unit,” Appl. Therm. Eng.,
suggestions and helpful comments. 18(9–10), pp. 847–856.
[10] Nagpal, M., Moshref, A., Morison, G. K., and Kundur, P., 2001,
“Experience With Testing and Modeling of Gas Turbines,” IEEE Power En-
Nomenclature gineering Society Winter Meeting, Columbus, OH, January 28-February 1,
Vol. 2, pp. 652–656.
C¼ pressure constant [11] Kaikko, J., Talonpoika, T., and Sarkomma, P., 2002, “Gas Turbine Model for
Cp ¼ specific heat in constant pressure, J/kg K an On-Line Condition Monitoring and Diagnostic System,” Proceedings of the
Cv ¼ specific heat in constant volume, J/kg K Australasian Universities Power Engineering Conference (AUPEC2002), Mel-
f¼ function bourne, Australia, September 29–October 2.
[12] Klang, H., and Lindholm, A., 2005, “Modeling and Simulation of a Gas
F¼ mass ratio of fuel to air Turbine,” Ph.D. thesis, Link€oping University, Link€ oping, Sweden.
HR ¼ heat rate, kJ/kWh [13] Camporeale, S. M., Fortunato, B., and Mastrovito, M., 2006, “A Modular Code
I¼ moment of inertia, kgm2 for Real Time Dynamic Simulation of Gas Turbines in Simuink,” ASME J.
LHV ¼ lower heating value of fuel, J/kg Eng. Gas Turbines Power, 128(3), pp. 506–517.
[14] Al-Hamdan, Q. Z., and Ebaid, M. S. Y., 2006, “Modeling and Simulation of a
M¼ maneuver Gas Turbine Engine for Power Generation,” ASME J. Eng. Gas Turbines
m_ ¼ mass flow rate, kg/s Power, 128(2), pp. 302–311.
n¼ number [15] Zhu, Y., and Frey, H. C., 2007, “Simplified Performance Model of Gas Turbine
N¼ rotational speed, rpm Combined-Cycle Systems,” J. Energy Eng., 133(2), pp. 82–90.
[16] Mantzaris, J., and Vournas, C., 2007, “Modeling and Stability of a Single-Shaft
P¼ stagnation pressure, Pa Combined-Cycle Power Plant,” Int. J. Thermodyn., 10(2), pp. 71–78.
PrC ¼ compressor pressure ratio [17] Yee, S. K., Milanovic, J. V., and Hughes, F. M., 2008, “Overview and Compar-
Q¼ heat power, W ative Analysis of Gas Turbine Models for System Stability Studies,” IEEE
R¼ specific gas constant, J/kg K Trans. Power Syst., 23(1), pp. 108–118.
[18] Lazzaretto, A., and Toffolo, A., 2008, “Prediction of Performance and Emis-
S¼ entropy, J/K sions of a Two-Shaft Gas Turbine From Experimental Data,” Appl. Therm.
t¼ time, s Eng., 28(17–18), pp. 2405–2415.
T¼ temperature, K [19] Khosravi-el-Hossani, M., and Dorosti, Q., 2009, “Improvement of Gas Turbine
TIT ¼ turbine inlet temperature,  C Performance Test in Combine-Cycle,” World Acad. Sci., Eng. Technol., 34, pp.
383–386, available at: http://waset.org/publications/13979
TOT ¼ turbine outlet temperature,  C [20] Bank Tavakoli, M. R., Vahidi, B., and Gawlik, W., 2009, “An Educational
u¼ externally determined variable Guide to Extract the Parameters of Heavy-Duty Gas Turbines Model in
V¼ volume, m3 Dynamic Studies Based on Operational Data,” IEEE Trans. Power Syst., 24(3),
W_ ¼ power, W pp. 1366–1374.
[21] Roldan-Villasana, E. J., Vazquez, A., and Jimenez-Sanchez, V. M., 2010,
y¼ output variable “Modeling of the Simplified Systems for a Power Plant Simulator,” Fourth
UKSim European Symposium on Computer Modeling And Simulation (EMS),
Pisa, Italy, November 17–19, pp. 277–282.
Greek Symbols [22] Yadav, N., Khan, I. A., and Grover, S., 2010, “Modeling and Analysis of Sim-
c ¼ ratio of specific heats ple Open-Cycle Gas Turbine Using Graph Networks,” Int. J. Electr. Electron.
Eng., 4(8), pp. 559–567, available at: http://www.waset.org/journals/ijeee/v4/
g ¼ efficiency v4-8-77.pdf
n ¼ pressure loss coefficient [23] Shalan, H. E., Moustafa Hassan, M. A., and Bahgat, A. B. G., 2011, “Parameter
Estimation and Dynamic Simulation of Gas Turbine Model in Combined-Cycle
Power Plants Based on Actual Operational Data,” J. Am. Sci., 7(5), pp. 303–310,
Subscripts available at: http://www.jofamericanscience.org/journals/am-sci/am0705/42_5267
am0705_303_310.pdf
air ¼ air [24] Lazzaretto, A., and Toffolo, A., 2001, “Analytical and Neural Network Models
c¼ compressor for Gas Turbine Design and Off-Design Simulation,” Int. J. Appl. Thermodyn.,
cc ¼ combustion chamber 4(4), pp. 173–182, available at: http://ijoticat.com/index.php/IJoT/article/view/
d¼ data sets 78/70
[25] Ogaji, S. O. T., Singh, R., and Probert, S. D., 2002, “Multiple-Sensor Fault-
f¼ fuel Diagnosis for a 2-Shaft Stationary Gas Turbine,” Appl. Energy, 71(4), pp.
gas ¼ gas 321–339.
load ¼ load [26] Arriagada, J., Genrup, M., Loberg, A., and Assadi, M., 2003, “Fault Diagnosis
m¼ model System for an Industrial Gas Turbine by Means of Neural Networks,” Interna-
tional Gas Turbine Congress (IGTC2003Tokyo), Tokyo, Japan, November 2–7,
t¼ turbine Paper No. TS-001.
01 ¼ compressor inlet [27] Basso, M., Giarre, L., Groppi, S., and Zappa, G., 2005, “NARX Models of an
02 ¼ compressor outlet Industrial Power Plant Gas Turbine,” IEEE Trans. Control Syst. Technol.,
03 ¼ turbine inlet 13(4), pp. 599–604.
[28] Bettocchi, R., Pinelli, M., Spina, P. R., Venturini, M., and Burgio, M., 2004,
04 ¼ turbine outlet “Set Up of a Robust Neural Network for Gas Turbine Simulation,” ASME
References Paper No. GT2004-53421.
[29] Bettocchi, R., Pinelli, M., Spina, P. R., Venturini, M., and Zanetta, G. A., 2006,
[1] Jelali, M., and Kroll, A., 2004, Hydraulic Servo-Systems: Modeling, Identifica-
“Assessment of the Robustness of Gas Turbine Diagnostics Tools Based on
tion, and Control, Springer, New York.
Neural Networks,” ASME Paper No. GT2006-90118.
[2] Rowen, W. I., 1983, “Simplified Mathematical Representations of Heavy-Duty
[30] Bettocchi, R., Pinelli, M., Spina, P. R., and Venturini, M., 2007, “Artificial
Gas Turbines,” ASME J. Eng. Power, 105(4), pp. 865–869.
Intelligence for the Diagnostics of Gas Turbines, Part I: Neural
[3] Rowen, W. I., 1992, “Simplified Mathematical Representations of Single-Shaft
Network Approach,” ASME J. Eng. Gas Turbines Power, 129(3), pp. 711–719.
Gas Turbines in Mechanical Derive Service,” Turbomach. Int., 33(5), pp.
[31] Bettocchi, R., Pinelli, M., Spina, P. R., and Venturini, M., 2007, “Artificial
26–32.
Intelligence for the Diagnostics of Gas Turbines, Part II: Neuro-Fuzzy
[4] Najjar, Y. S. H., 1994, “Performance of Single-Cycle Gas Turbine Engines in
Approach,” ASME J. Eng. Gas Turbines Power, 129(3), pp. 720–729.
Two Modes of Operation,” Energy Convers. Manage., 35(5), pp. 433–441.
[32] Spina, P. R., and Venturini, M., 2007, “Gas Turbine Modeling by Using Neural
[5] Hannett, L. N., Jee, G., and Fardanesh, B., 1995, “A Governer/Turbine Model
Networks Trained on Field Operating Data,” 20th International Conference on
for a Twin-Shaft Combustion Turbine,” IEEE Trans. Power Syst., 10(1), pp.
Efficiency, Costs, Optimization, Simulation and Environmental Impact of
133–140.
Energy Systems (ECOS2007), Padova, Italy, June 25–28.
[6] Bettocchi, R., Spina, P. R., and Fabbri, F., 1996, “Dynamic Modeling of Single-
[33] Simani, S., and Patton, R. J., 2008, “Fault Diagnosis of an Industrial Gas Tur-
Shaft Industrial Gas Turbine,” ASME Paper No. 96-GT-332.
bine Prototype Using a System Identification Approach,” Control Eng. Pract.,
[7] Ricketts, B. E., 1997, “Modeling of a Gas Turbine: A Precursor to Adaptive
16(7), pp. 769–786.
Control,” IEE Colloquium on Adaptive Controllers in Practice ‘97 (Digest No.
[34] Yoru, Y., Karakoc, T. H., and Hepbasli, A., 2009, “Application of Artificial
197/176), Coventry, UK, April 10, pp. 7/1–7/5.
Neural Network (ANN) Method to Exergetic Analyses of Gas Turbines,” Inter-
[8] Bianchi, M., Peretto, A., and Spina, P. R., 1998, “Modular Dynamic Model of
national Symposium on Heat Transfer in Gas Turbine Systems, Antalya, Tur-
Multi-Shaft Gas Turbine and Validation Test,” Proceedings of The Winter An-
key, August 9–14.
nual Meeting of ASME, New York, Vol. 38, pp. 73–81.

Journal of Engineering for Gas Turbines and Power JUNE 2014, Vol. 136 / 061601-9

Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms

 [35] Fast, M., Assadi, M., and De, S., 2008, “Condition Based Maintenance of Gas
Turbines Using Simulation Data and Artificial Neural Network: A Demonstration
of Feasibility,” ASME Paper No. GT2008-50768.
[36] Fast, M., Assadi, M., and De, S., 2009, “Development and Multi-Utility of an
ANN Model for an Industrial Gas Turbine,” J. Appl. Energy, 86(1), pp. 9–17.
[37] Fast, M., Palme, T., and Genrup, M., 2009, “A Novel Approach for Gas Turbine
Condition Monitoring Combining CUSUM Technique and Artificial Neural
Network,” ASME Paper No. GT2009-59402.
[38] Palme, T., Fast, M., and Karlsson, A., 2009, “Gas Turbines Sensor Validation
Through Classification With Artificial Neural Networks,” Appl. Energy,
88(11), pp. 3898–3904.
[39] Fast, M., and Palme, T., 2010, “Application of Artificial Neural Network to the
Condition Monitoring and Diagnosis of a Combined Heat and Power Plant,”
Energy, 35(2), pp. 1114–1120.
[40] Fast, M., 2010, “Artificial Neural Networks for Gas Turbine Monitoring,”
Ph.D. thesis, Lund University, Lund, Sweden.
[41] Walsh, P. P., and Fletcher, P., 1998, Gas Turbine Performance, 2nd ed., Penn-
well Books, Tulsa, OK.
[42] Morini, M., Cataldi, G., Pinelli, M., and Venturini, M., 2007, “A Model for the
Simulation of Large-Size Single-Shaft Gas Turbine Start-Up Based on Operat-
ing Data Fitting,” ASME Paper No. GT2007-27373.
[43] Agrawal, R. K., and Yunis, M., 1982, “A Generalized Mathematical Model to
Estimate Gas Turbine Starting Characteristics,” ASME J. Eng. Gas Turbines
Power, 104(1), pp. 194–201.
[44] Balakrishnan, S. R., and Santhakumar, S., 1996, “Fuzzy Modeling Considerations
in an Aero Gas Turbine Engine Start Cycle,” Fuzzy Sets Syst., 78(1), pp. 1–4.
[45] Peretto, A., and Spina, P. R., 1997, “Comparison of Industrial Gas Turbine
Transient Responses Performed by Different Dynamic Models,” ASME Paper
No. 97-GT-177.
[46] Beyene, A., and Fredlund, T., 1998, “Comparative Analysis of Gas Turbine
Engine Starting,” ASME Paper No. 98-GT-419.
[47] Kim, J. H., Song, T. W., Kim, T., and Ro, S. T., 2002, “Dynamic Simulation of
Full Startup Procedure of Heavy-Duty Gas Turbines,” ASME J. Eng. Gas Tur-
bines Power, 124(3), pp. 510–516.
[48] Kim, J. H., Song, T. W., Kim, T., and Ro, S. T., 2001, “Model Development
and Simulation of Transient Behavior of Heavy Duty Gas Turbines,” ASME J.
Eng. Gas Turbines Power, 123(3), pp. 589–594.
[49] Kim, T. S., Park, H. J., and Ro, S. T., 2001, “Characteristics of Transient Opera-
tion of a Dual-Pressure Bottoming System for the Combined Cycle Power
Plant,” Energy, 26(10), pp. 905–918.
[50] Shin, J. Y., Jeon, Y. J., Maeng, D. J., Kim, J. S., and Ro, S. T., 2002, “Analysis
of the Dynamic Characteristics of a Combined-Cycle Power Plant,” Energy,
27(12), pp. 1085–1098.
[51] Davison, C. R., and Birk, A. M., 2004, “Steady State and Transient Modeling
of a Micro-Turbine with Comparison to Operating Engine,” ASME Paper No.
GT2004-53378.
061601-10 / Vol. 136, JUNE 2014
Downloaded From: http://gasturbinespower.asmedigitalcollection.asme.org/ on 04/06/2014 Terms of Use: http://asme.org/terms
[52] Huang, X. H., and Zheng, X. S., 2005, “Research on Startup Model of Aircraft
Engine Based on Stage-Stacking Method,” Acta Aeronaut. Astronaut. Sinica,
26(5), pp. 524–528, available at: http://en.cnki.com.cn/Article_en/CJFDTOTAL-
HKXB200505003.htm
[53] Xunkai, W., and Yinghong, L., 2006, “Aero-Engine Dynamic Start Model Based
on Parsimonious Genetic Programming,” Sixth World Congress on Intelligent
Control and Automation (WCICA 2006), Dalian China, June 21–23, Vol. 1, pp.
1478–1482.
[54] Sanaye, S., and Rezazadeh, M., 2007, “Transient Thermal Modeling of Heat
Recovery Steam Generators in Combined Cycle Power Plants,” Int. J. Energy
Res., 31(11), pp. 1047–1063.
[55] Corbett, M., Lamm, P., Owen, P., Phillips, S. D., Blackwelder, M., Alt, J. T.,
McNichols, J., Boyd, M., and Wolff, J. M., 2008, “Transient Turbine Engine
Modeling With Hardware-in-the-Loop Power Extraction,” AIAA Paper No.
2008-5932.
[56] Alobaid, F., Postler, R., Strohle, J., Epple, B., and Hyun-Gee, K., 2008,
“Modeling and Investigation Start-Up Procedures of a Combined Cycle Power
Plant,” Appl. Energy, 85(12), pp. 1173–1189.
[57] Daneshvar, K., Behbahani-nia, A., Khazraii, Y., and Ghaedi, A., 2012,
“Transient Modeling of Single-Pressure Combined Cycle Power Plant Exposed
to Load Reduction,” Int. J. Model. Optim. 2(1), pp. 64–70.
[58] Sarkar, S., Mukherjee, K., Sarkar, S., and Ray, A., 2012, “Symbolic Transient
Time-Series Analysis for Fault Detection in Aircraft Gas Turbine Engines,”
American Control Conference (ACC), Montreal, QC, Canada, June 27–29, pp.
5132–5137.
[59] Rachtan, W., and Malinowski, L., 2013, “An Approximate Expression for Part-
Load Performance of a Microturbine Combined Heat and Power System Heat
Recovery Unit,” Energy, 51, pp. 146–153.
[60] Venturini, M., 2005, “Simulation of Compressor Transient Behavior Through
Recurrent Neural Network Models,” ASME J. Turbomach., 128(3), pp. 444–454.
[61] Venturini, M., 2006, “Optimization of a Real-Time Simulator Based on Recur-
rent Neural Networks for Compressor Transient Behavior Prediction,” ASME J.
Turbomach., 129(3), pp. 468–478.
[62] Venturini, M., 2004, “Development and Experimental Validation of a Compres-
sor Dynamic Model,” ASME J. Turbomach., 127(3), pp. 599–608.
[63] Morini, M., Pinelli, M., and Venturini, M., 2006, “Development of a One-
Dimensional Modular Dynamic Model for the Simulation of Surge in Compres-
sion Systems,” ASME J. Turbomach., 129(3), pp. 437–447.
[64] Morini, M., Pinelli, M., and Venturini, M., 2009, “Analysis of Biogas Compres-
sion System Dynamics,” Appl. Energy, 86(11), pp. 2466–2475.
[65] Asgari, H., Chen, X. Q., Sainudiin, R., Morini, M., Pinelli, M., Spina, P. R., and
Venturini, M., 2014, “Modeling and Simulation of the Startup Operation of a
Heavy-Duty Gas Turbine Using NARX Models,” ASME Paper No. GT2014-
25056.
[66] Beale, M. H., Hagan, M. T., and Demuth, H. B., 2011, Neural Network Tool
BoxTM User’s Guide R2011b, The MathWorks, Inc., Torrance, CA.
Transactions of the ASME