You are on page 1of 0

Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data

Spina, P. R., Venturini, M.


1
GAS TURBINE MODELING BY USING
NEURAL NETWORKS TRAINED
ON FIELD OPERATING DATA
ECOS 2007
June 2528, 2007, Padova, Italy
P. R.
P. R.
Spina
Spina
,
,
M. Venturini
M. Venturini
Energy System Research Group
ENDIF - Engineering Department in Ferrara
University of Ferrara (ITALY)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
2
INTRODUCTION
Monitoring and diagnostics techniques are useful as a support for an optimized
management of gas turbine operation (high levels of availability in the most
efficient operating conditions)
Different methodologies for gas turbine diagnostics have been developed
Linear Gas Path Analysis, Linear Estimators, Non-Linear Gas Path Analysis
Artificial Intelligence based methods (Neural Networks, Fuzzy and Neuro-Fuzzy
Systems and Expert Systems)
In the field application of Gas Path Analysis methodologies, the main problems are
the availability of a program for the thermodynamic cycle calculation (Cycle Program)
which should accurately reproduce the actual machine behavior (modeling approximations,
lack of accurate data for CP fine tuning on the particular unit)
the available measured data which are affected by measurement errors and, sometimes,
may be also corrupted by bias errors due to sensor malfunctions
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
3
need for limited knowledge of the physics of the problem to be modeled
capability of learning different typologies of information
high robustness in the presence of poor and incorrect input data
high computational speed, which allows real-time calculations
The main limit of black box models is high prediction error when
they operate outside the field in which they were trained
Among black box models, Neural Networks (NNs) have proved
to be flexible and robust in simulating and diagnosing the
behavior of energy systems
Reasons for adopting black box models
instead of physical models
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
4
GAS TURBINE DIAGNOSTIC SYSTEM
Boundary
conditions
(p
amb
, T
amb
, )
+
Load conditions
(P, N, IGV)
Measurable
quantities along
the gas path
(p
oC
, T
oC
, M
f
, T
oT
)
'
*
m
m
m
Q
Q
Q =
Baselines
estimator
Health
Indices
estimator
The relative deviations of the Health Indices are
evaluated starting from the relative deviations of
the measurements with respect to their expected
values (baselines) since, in this manner, errors in
estimating the Health Indices can be reduced
Efficiencies
Corrected mass
flow rates
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
5
Tools for gas turbine diagnostics
Already developed Work in progress
Diagnostic
module
Tool Specific tool Training data
Thermodynamic
program
Cycle program
(direct mode)
Statistics
Regression,

Simulated Field
Neural
Network
Simulated Field
Baselines
estimator
Black-box
Neuro-Fuzzy
Network
Simulated Field
Thermodynamic
program
Cycle program
(inverse mode)
Neural
Network
Simulated only
Neuro-Fuzzy
Network
Simulated only
Health Indices
estimator
Black-box
Fuzzy Logic
System
Rules defined by the
operator
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
6
Few applications reported
in literature
Development of NN models for the simulation of gas turbines (baseline
estimator) by using field operating data
Data used for both training and testing the NNs are collected
during normal operation over a period of about seven months
on a FIAT Avio 701F single shaft gas turbine
Focus on the choice of the data to be used for NN training and testing by
evaluating their number and significance
The best NN model is tested over all the available field data and also
against a Cycle Program calibrated on the same gas turbine
OBJECTIVES
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
7
FIAT AVIO 701F
- single-shaft
- variable Inlet Guide Vanes
GAS TURBINE CONSIDERED
C
T ~
CC
fuel
ID
air
ED
iT
oT
oC
fAH
cool
P
GT
= 255 MW

GT
= 36.0 %
T
oT
= 889 K
M
oT
= 574 kg/s
p
oC
= 1408 kPa
T
oC
= 655 K
M
f
= 13.2 kg/s
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
8
DATA SELECTION
The selection of data used for the NN training phase can be
performed according to two different choices:
NN training patterns can be chosen through a heuristic procedure,
based on experience and knowledge of the process to be modeled
(passive learning)
NN training patterns can be obtained by means of an optimization
procedure, which selects data patterns in order to minimize a cost
function (active learning)
In the paper, the two procedures are combined:
experience and knowledge of the process are used to identify the
base patterns
the optimization procedure is adopted to select new training
patters to optimize NN performance
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
9
Available Measured Data
Operating period of about seven months (from 3
rd
June to 23
rd
December 2003), for a total amount of 14593 available
measured patterns. Data sampling time is equal to 15 minutes
Temperature control logic (TCL) is usually adopted for normal
operation at base load. Thus, only data patterns referring to TCL
are considered for NN training and testing: these are 10150 data
patterns out of the 14593 originally available measured data
Gas turbine thermodynamic measurements only are considered
(steam section measurements and vibrations excluded)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
10
Available measured quantities and range of variation
for data patterns (temperature control logic)
12/20/03
1.15 - 2.00 p.m.
-
From 06/03/03 to
12/23/03
- Date and time
Temp. control
mode
- Binary code - Gas turbine control mode index
834.9 0.90
(*)
[828.0; 858.0] K Turbine outlet temperature T
oT
12.54 1.50
(*)
[11.12; 13.61] kg/s Fuel mass flow rate M
f
692.0 0.23
(*)
[621.6; 725.9] K Compressor outlet temperature T
oC
1 518 1.43 [1402; 1582] kPa Compressor outlet pressure p
oC
218 522 2.0
(*)
[192591; 235603] kW Gas turbine power P
GT
2.86 1.0
(*)
[-2.43; 3.97] IGV angular position A
IGV
49 619 3.0
(*)
[47459; 50016] kJ/kg Fuel lower heating value LHV
379.2 1.0 [352.0; 391.1] K Temperature of the fuel after heating T
fAH
461.0 1.0 [456.6; 468.9] K
Cooling flow temp. after the
exchanger
T
cool
3001 30.0 [2995; 3005] rpm Rotational speed N
2.14 0.10 [1.98; 2.66] kPa Pressure drop at the turbine outlet p
ED
1.98 0.10 [1.67; 2.25] kPa Pressure drop at the compressor inlet p
ID
77 0 2.0 [21; 100] % Relative humidity RH
288.7 0 0.3 [277.7; 308.9] K Ambient temperature T
amb
102. 5 0.630 [99; 103] kPa Ambient pressure p
amb
Value for CP
calibration
Measurement
uncertainty [14]
Variation range
for available data
Unit Quantity
(*) indicates % of the reading
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
11
Trend over time of some measured quantities
500
525
550
575
600
6/1/03 8/10/03 10/19/03 12/28/03 day
T
ot
[C]
Load control
Temperature control
0
10
20
30
40
6/1/03 8/10/03 10/19/03 12/28/03 day
T
amb
[C]
150
175
200
225
250 Ambient temperature
Net electric power
P
[MW]
1300
1400
1500
1600
6/1/03 8/10/03 10/19/03 12/28/03 day
p
oc
[kPa]
11
12
13
14 Compressor outlet pressure
Fuel mass flow rate
M
f
[kg/s]
For temperature control logic, turbine
outlet temperature T
ot
follows T
amb
P
GT
increases as T
amb
decreases in
cold season
T
amb
decrease and P
GT
increase make
both M
f
and p
oc
increase
Data refer to Temperature Control Logic Data refer to Temperature Control Logic
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
12
Input
layer
Hidden
layer/s
Output
layer
2
1 1
1
patt
o
n
n
ij ij
j i
o patt
mse t y
n n
= =

=


Performance function
Multi-layer perceptron
(feed-forward neural network)
ARTIFICIAL NEURAL NETWORKS
One hidden layer
Continuous sigmoid activation function
60 neurons in the hidden layer (good compromise between NN
accuracy and computational time required for NN training)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
13
For the NN training the back-propagation algorithm was used
The adopted stopping criterion for the NN training phase is the
minimization of the MSE on the whole training set
[ ]

= = = =

=
o
patt
o
patt
1 1
2
patt o
1 1
2
patt o
1 1
n
j
n
i
ij ij
n
j
n
i
ij
y t
n n
e
n n
MSE
t
ij
: target output
y
ij
: NN computed output
Additional stopping criteria:
attainment of 10
6
epochs
early stopping method: the
training process is stopped when
the MSE on validation data does
not decrease for 100 epochs
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
14
o
2
1
patt
1,..., ,
1
patt
n j
t
y t
n
RMSE
j
n
i
i
i i
=


=

=
RMSE on the whole set of test patterns
on each output (p
oC
, T
oC
, M
f
, T
oT
)
EVALUATION OF NN PERFORMANCE
RMSE
ov
on the whole set of test patterns
in the calculation of all outputs
( )

=
=
o
n
j
j
o
ov
RMSE
n
RMSE
1
2
1
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
15
METHODOLOGY FOR SELECTING
NN TRAINING DATA
The selection of data patterns to be used for NN training and
testing in case of field operating data represents a key problem
once the training set is chosen, the NN will reproduce the particular
condition which is identified by this specific training set
NN performance also depends on patterns chosen for testing.
In fact, if training and testing patterns are consistent with each
other, NN performance can be reasonably expected to be good
- careful selection of NN training data, in order to cover all the possible
range of variation of the considered process variables
- the NN model should not operate outside the field in which it was
trained (extrapolation not advisable)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
16
Field data patterns for NN training
Number of available data patterns (good field data availability may
be limited and/or not statistically significant)
Patterns should be representative of a given health condition for the
machine (e.g. new and clean conditions)
Patterns should cover the entire expectable range of variation for
all input quantities
Test patterns should refer to a machine health state close to that of
training patterns. Thus, NN capability in reproducing a given health
condition can be assessed, and, at the same time, the NN can be used
to evidence a health condition different from that of training data
Stationary or transient conditions should be considered (In the paper,
gas turbine stationary behavior only is considered)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
17
Active learning to select training patterns
Iterative process to let
the NN reach a desired
level of accuracy
Guess training set to
start the procedure
(base patterns)
Objective function to
be minimized
(in the paper it is the
Mean Square Error)
Available data patterns
(Training data)
(I)
NN
(I)
= NN trained on (Training data)
(I)
Simulation through NN
(I)
Calculation of F
ob
(I)
F
ob
(I)
< F
ob
(I-1)
Discard (Training data)
(I)
NO
YES
UPDATE
training data
(I)
=
(I+1)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
18
1
st
step of the Active Learning procedure
- base patterns identification -
The base patterns should include patterns from winter and
summer seasons (extended range of ambient temperatures)
Four patterns for low ambient temperatures (20
th
December)
These same patterns will be used for Cycle Program calibration
Four patterns to consider also high T
amb
(end of July)
Both groups of data (July and December) refer to a similar health
state (4 months of operation)
The NN model obtained by using these eight base patterns is
identified as NN
(0)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
19
Identification of the period for searching training and testing data
Training: the weeks before the considered base patterns (one
week in July and one week in December) ought to be searched
for further candidate patterns
Testing: test patterns should also belong to the same two weeks
considered above
2
nd
step of the Active Learning procedure
- searching for other patterns -
Sixteen patterns taken on the same two weeks as validation data,
to apply the early stopping method (ESM)
Both choices for training and test data find their reason in setting
up a simulator for the machine, i.e. an NN model highly
specialized in reproducing a particular machine health state
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
20
Identification of NN
(final)
, whose training set is composed of only 10 patterns
These patterns allow the extension of the range of variation for some quantities
(e.g. LHV and M
f
)
NN
(final)
actually proves better for all quantities
The quantity mainly improved is compressor outlet pressure
Overall error is just 1.7 %, obtained with 10 training patterns only!
Identification of the best NN model
Simulation of all
available data from
June to December
0.0
1.0
2.0
3.0
poc Toc Mf Tot Overall
RMSE
[%]
11 nor a6a
p
oc
T
oc
M
f
T
ot
Overall
NN
(final)
NN
(0)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
21
Use of validation data
0.0
1.0
2.0
3.0
poc Toc Mf Tot Overall
RMSE
[%]
11 nor 11 val a6a a6a val
p
oc
T
oc
M
f
T
ot
Overall
NN
(final)
ESM NN
(final)
NN
(0)
NN
(0)
ESM
Simulation of all
available data from
June to December
The use of validation data (early stopping method ESM) confirms that
the best NN model is actually NN
(final)
In fact, the effect of ESM can not be appreciated since the NN model is
already optimized through the active learning procedure
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
22
GPA model calibration
The Cycle Program allows the calculation of the most significant quantities along
the gas path by using component performance maps
Data for the calibration are obtained as an average value of four data patterns over
45 minutes of operation (20
th
December 2003, 1.15 p.m. - 2.00 p.m.)
These data can be considered representative of quasi-ISO conditions (machine
operation assumed stable and stationary)
C
T
~
CC
fuel
ID
air
ED
iT
oT
oC
fAH
cool
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
23
Comparison between the best NN model and
the GPA Cycle Program
The performance of the NN model obtained through the active learning procedure
is tested against
- all available data (7 months)
- Cycle Program response
Boundary
conditions
(p
amb
, T
amb
, )
+
Load conditions
(P, N, IGV)
Measured
quantities along
the gas path
(p
oC
, T
oC
, M
f
, T
oT
)
Cycle Program
NN
(final)
Expected
value Q
NN
Expected
value Q
CP
ected exp
measured *
Q
Q
Q =
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
24
Analysis of the results
For all quantities, the normalization through NN
(final)
provides trends more
scattered than those obtained through the CP
For p
oC
and M
f
, the relative deviation is about 5 %
The relative deviation for T
oC
is about 3 %
T
oT
is better reproduced by both models ( 2 %)
CP response seems to be more centered around the value 1 (expected value
if GT health state remains unchanged)
0.90
0.95
1.00
1.05
1.10
6/1/2003 8/10/2003 10/19/2003 12/28/2003
day
p
oc
*
Active 6a Cycle Program
NN
(final)
Cycle Program
0.90
0.95
1.00
1.05
1.10
6/1/2003 8/10/2003 10/19/2003 12/28/2003 day
T
oc
*
Active 6a Cycle Program NN
(final)
Cycle Program
0.90
0.95
1.00
1.05
1.10
6/1/2003 8/10/2003 10/19/2003 12/28/2003 day
M
f
*
Active 6a Cycle Program NN
(final)
Cycle Program
0.90
0.95
1.00
1.05
1.10
6/1/2003 8/10/2003 10/19/2003 12/28/2003 day
T
ot
*
Active 6a Cycle Program NN
(final)
Cycle Program
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
25
CONCLUSIONS
In the paper, Neural Networks for the simulation of gas turbines
are developed in order to reproduce gas turbine behavior
The field data used for both training and testing the NNs refer to
a single-shaft GT and were taken during normal operation
Active learning procedure leads to an NN model
trained by means of only 10 patterns, taken on 3 days only
with a good accuracy, i.e. overall error over all the available
field data equal to 1.7 %
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
26
CONCLUSIONS
Finally, the NN model behaves in a more scattered way with
respect to the Cycle Program
However
Very limited information about physical system is required and,
so, an NN model can be easily implemented for a whatever system
The development and calibration of a cycle program requires a
significant amount of time and the availability of specific engine
data, such as the performance maps of main components
An NN model allows an acceptable simulation of the machine with
a low computational and time effort
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
27
GAS TURBINE MODELING BY USING
NEURAL NETWORKS TRAINED
ON FIELD OPERATING DATA
ECOS 2007
June 2528, 2007, Padova, Italy
P. R.
P. R.
Spina
Spina
,
,
M. Venturini
M. Venturini
Energy System Research Group
ENDIF - Engineering Department in Ferrara
University of Ferrara (ITALY)
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
28
Measurement errors (standard instrumentation category)
a
Percentage of the reading
Non-dimensional quantities are the ratio between two quantities: the measured value
and the corresponding value expected in new and clean conditions
at the same boundary and load conditions.
Thus, their uncertainties are obtained from absolute uncertainties in
Bettocchi et al. (2004) and evaluated by using the uncertainty propagation equation.
Quantity
Measurement
uncertainty
Quantity
Measurement
uncertainty
p
amb
0 0.600 [kPa] LHV 1.0 [%]
a
T
amb
0 0.3 [K] A
IGV
1.0 []
RH 0 3.0 [%] P
GT
0.8 [%]
a
p
IC
0 0.10 [kPa] p
*
oC
1.50 [%]
a
p
ED
0 0.10 [kPa] T
*
oC
1.03 [%]
a
N 10.0 [rpm] M
*
f
1.94 [%]
a
T
fPH
0 1.0 [K] T
*
oT
1.33 [%]
a
Gas Turbine Modeling by Using Neural Networks Trained on Field Operating Data
Spina, P. R., Venturini, M.
29
[4] Measurements
along the gas path
Neural Network inputs (11) and outputs (4)
INPUTS
p
amb
Ambient pressure
T
amb
Ambient temperature
RH Relative humidity
p
ID
Pressure drop at the compressor inlet
p
ED
Pressure drop at the turbine outlet
T
fPH
Temperature of the fuel after heating
T
cool
Cooling flow temp. after the exchanger
LHV Fuel lower heating value
N Rotational speed
A
IGV
IGV angular position
P
GT
Gas turbine power

OUTPUTS
p
*
oC
Compressor outlet pressure
T
*
oC
Compressor outlet temperature
M
*
f
Fuel mass flow rate
T
*
oT
Turbine outlet temperature

[8] Ambient conditions
[3] Operating point

You might also like