Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220372259

Prediction of power output of a coal-fired power plant by artificial neural


network

Article  in  Neural Computing and Applications · July 2010


DOI: 10.1007/s00521-009-0331-6 · Source: DBLP

CITATIONS READS

33 607

5 authors, including:

Jure Smrekar Mohsen Assadi


JS energy Ltd. University of Stavanger (UiS)
19 PUBLICATIONS   261 CITATIONS    115 PUBLICATIONS   1,444 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Absorption heat transformer View project

EU H2-IGCC View project

All content following this page was uploaded by Jure Smrekar on 24 November 2016.

The user has requested enhancement of the downloaded file.


Neural Comput & Applic (2010) 19:725–740
DOI 10.1007/s00521-009-0331-6
Author's personal copy
ORIGINAL ARTICLE

Prediction of power output of a coal-fired power plant by artificial


neural network
J. Smrekar • D. Pandit • M. Fast • M. Assadi •

Sudipta De

Received: 7 January 2009 / Accepted: 16 December 2009 / Published online: 30 December 2009
Ó Springer-Verlag London Limited 2009

Abstract Accurate modeling of thermal power plant is been discussed along with the expected prediction accu-
very useful as well as difficult. Conventional simulation racies and validation of models with real plant data. The
programs based on heat and mass balances represent plant interpolation and extrapolation capability of ANN models
processes with mathematical equations. These are good for for the plant has also been studied, and observed results are
understanding the processes but usually complicated and at reported.
times limited with large number of parameters needed. On
the other hand, artificial neural network (ANN) models Keywords Power plant  Coal-fired boiler 
could be developed using real plant data, which are already Steam turbine  ANN model  Real plant data 
measured and stored. These models are fast in response and Extrapolation  Interpolation
easy to be updated with new plant data. Usually, in ANN
modeling, energy systems can also be simulated with fewer List of symbols
numbers of parameters compared to mathematical ones. AI Artificial intelligence
Step-by-step method of the ANN model development of a ANN Artificial neural network
coal-fired power plant for its base line operation is dis- B Bleed steam data, pressure (kg/cm2) and
cussed in this paper. The ultimate objective of the work temperature (°C)
was to predict power output from a coal-fired plant by CHP Combined heat and power
using the least number of controllable parameters as inputs. C Concentration (%)
The paper describes two ANN models, one for boiler and MLP Multi layer perceptron
one for turbine, which are eventually integrated into a MRE Mean relative error (%)
single ANN model representing the real power plant. The m Mass flow rate (t/h)
two models are connected through main steam properties, p Pressure (kg/cm2)
which are the predicted parameters from boiler ANN stg1 Group of parameters (mcoal, UB,out, pfw, tfw)
model. Detailed procedure of ANN model development has optimized after first stage sensitivity analysis
t Temperature (°C)
J. Smrekar  M. Assadi Greek symbol
Department of Mechanical and Structural Engineering U Valve opening (°)
and Materials Science, University of Stavanger,
Stavanger, Norway Subscripts
1–6 Steam extractions from turbine for bleeding
D. Pandit  S. De (&)
Department of Mechanical Engineering, B Boiler
Jadavpur University, Kolkata, India CN Condenser
e-mail: de_sudipta@rediffmail.com coal Coal
FG Flue gas to stack
M. Fast  M. Assadi
Department of Energy Sciences, fw Feed water
Lund University, Lund, Sweden in At the inlet

123
726 Author's personal copy Neural Comput & Applic (2010) 19:725–740

out At the outlet between involved parameters. To develop the mathematical


oxygen Oxygen model of a boiler is a challenging job. Several such devel-
p Pressure opments of mathematical models have been reported [1–10]
s Steam in literature. Similar works are also reported for the mathe-
SA Secondary air to the boiler furnace matical modeling of turbines [11–20]. Different aspects of
t Temperature the development of ANN models for gas turbines have been
reported also by several authors [21–23]. However, ANN
models for steam turbines using real plant data are not
1 Introduction available.
Each of these mathematical models reported in literature
Modeling a power plant has been a challenging issue. It is developed with some simplifying assumptions, as it is
involves many components with large number of involved almost impossible to incorporate all aspects of operation of
parameters. Mathematical models try to express actual such a complex system in a single model. Thus, results
processes by representative mathematical equations. from these models show some agreement with real plant
Though these are developed with better understanding of data only for some conditions of operations. Moreover,
the processes, it is very difficult to develop such mathe- these models have limitations also as mentioned earlier.
matical models for a complete power plant. This is mainly Introduction of artificial intelligence (AI)-based tools for
due to the large number of components/subcomponents of the modeling and simulation of real life complex systems
the plant and inherent complexity in modeling them. may be a suitable alternative [24–28]. ANN is one of these
Integrating models for these components for the prediction AI-based tools that could be used for the modeling of
of the overall performance of the plant also creates some energy systems for which operational data (either real plant
difficulties as equations for several components have or simulated) are available. The development of ANN
mutual dependence through some common interdependent models for the steam process of a biomass and coal co-fired
parameters. Thus, iterative solution is the only way, which combined heat and power plant using real plant data has
makes it difficult for real time use. Moreover, the accuracy already been reported by the same group [29, 30]. How-
of the developed mathematical model will decrease over ever, there are a few reported works [31–35] on the mod-
time if the inherent ‘degradation’ of the plant is not taken eling of conventional coal-fired boilers using ANN, and
into consideration. But it is again usually very difficult. only one of them [35] used real plant data.
On the other hand, real data for large number of Similarly, different aspects of the development of ANN
parameters of an operating plant are always captured and models for gas turbines have been reported by several
stored as database for monitoring and other purposes. authors [21–23, 36]. However, ANN models for steam
Hence, real data for any operating power plant are always turbines using real plant data are not available.
available without any additional effort or investment. These In this study, the step-by-step development of ANN
already stored data can be used for the development of model for the steady-state operation of a coal-fired power
artificial neural network (ANN) model for the components plant of India is reported. ANNs were trained with real
or the whole power plant. If properly trained with this real plant data. Separate ANN models were developed for the
plant data, ANN can predict the values of desired parame- two most important components of the plant, i.e., the boiler
ters with good accuracy. However, the prediction of the and the steam turbine. Selection of data from the available
trained ANN is also expected to deviate from the measured plant database for training and determination of ‘optimum’
value over time due to the ‘degradation’ of the plant. This number of input parameters for decided output parameters
deviation can be used to estimate the plant degradation, and through sensitivity analysis has been discussed. The
measures may be adopted accordingly. Moreover, a new developed models were also tested with data from the
ANN may be trained with latest data for more accurate plant, which were not used during training. This indicated
prediction of plant parameters if this degradation cannot be the expected accuracy by the developed model, i.e., the
avoided. Thus, it is easy to update the ANN model with validation of the model for real life use. Finally, the models
current status of the plant. As trained ANNs produce results for the boiler and the turbine were integrated into a single
without any iteration, they are very fast in response and can model to predict the overall power output of the plant by
be used for ‘online’ applications also. providing the values of only a few number of input
Boiler is undoubtedly one of the most important equip- parameters. The integration of models was done by using
ment of a power plant that needs precise monitoring and the predicted values of output parameters of the boiler
control. Models to simulate the operation of boilers are hence model as some of the inputs to the turbine model rather
very important requirement. However, the operation of the than using real plant data for the prediction of its power
boiler is very complex with numerous interdependences output. Thus, the required measured data for the prediction

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 727

of power from the plant reduced to a few only, though the and tube, slow speed pulverizers with gravimetric-type raw
accuracy of prediction was good. Thus, the feasibility of coal feeders for each coal mill. Three primary air fans were
ANN model development for coal-fired power plants using in operation, one for each pulverizer. Each mill was sup-
already stored data has been established. plied with one seal air fan, which takes suction from
The whole work covers the modeling of a real power atmosphere and supplies seal air to coal mills for sealing
plant consisting of the boiler and steam turbine models. In system, i.e., preventing the pressurized dust coal from
addition, issues related to the integration of these two coming out of the mill. The superheater steam system had
developed models are discussed. ANNs trained with data mainly three sections. The low temperature superheater
for a particular range of power and used for prediction of was arranged in the second pass of the unit. The radiant
the same beyond this range is also covered in the paper. platen superheater was located at the outlet section of the
Thus, ‘interpolation’ and ‘extrapolation’ capability of ANN furnace, and the final superheater was located in the hori-
for such plants was also tested. zontal pass of the boiler. The reheater was located in the
This work is one of the series of studies [29, 30, 35–37] horizontal pass of the boiler in between the radiant platen
made on ANN modeling of coal-based power and CHP superheater and final superheater sections. The econo-
plants. These studies explored the aspects of generalization mizer, a single block unit, was of the continuous loop, plain
of ANN modeling of such plants, if any, and established a tubular, drainable and horizontal, in line arrangement with
road map for such studies in future. water flow upward and gas flow downward.
The turbine was manufactured by Bharat Heavy Elect-
ricals Limited, India. It was a 210 MW, three-cylinder
2 Brief description of the power plant condensing-type 50% reaction turbine with high-, medium-
and low-pressure stages. Rated steam pressure before
The power plant was a 210 MW unit designed on the emergency stop valve was 150 kg cm-2 at full load. Rated
concept of unit system where a single boiler supplies steam steam temperature before emergency stop valve and before
to a single turbine, coupled to an AC Generator. The basic IP connector valve was same and was 535°C at full load.
schematic of the plant is shown in Fig. 1. The boiler was a There were totally six extractions or bleeds of steam for
radiant, reheat, natural circulation, single drum, dry bottom the regenerative heating of feed water to the boiler. The
and semi outdoors type. Boiler was provided with three ball regenerative feed water heating was done by heating the

Fig. 1 Schematic of the 210 MW coal-fired power plant

123
728 Author's personal copy Neural Comput & Applic (2010) 19:725–740

condensate in ejectors, gland steam cooler, low-pressure desired accuracy, the model is ready to predict the outputs
heaters, deaerator and high-pressure heaters. Extracted for data not presented to the ANN during training, i.e.,
steam from various points of the turbine was utilized to ‘unknown’ to it.
heat the condensate in these heaters. The condenser cooling
water system was a closed cycle. Three cooling water
pumps of 15,000 m3/h capacity were provided for each unit 4 Data from the plant and selection of ANN training
with suction from the common cooling water basin and data
supplied water to the twin condenser of each unit. The hot
return water was taken to the cooling tower of the unit. Data for a large number of parameters are usually mea-
After being cooled in the induced-draft-type cooling tower, sured in a power plant. These are used for the monitoring
the water was returned to the cooling water basin through a of the plant and stored for possible future use. However, for
concrete tunnel. To make up the drift and evaporation ANN model, only a few of these are good enough as input
losses, clarified water was supplied. parameters for the prediction of specified output parame-
The rated output of the generator was 210 MW (with ters. Thus, selection of proper input parameters is an
150% overload capacity). The HP and IP rotors were important issue as discussed in Sect. 5.1. On the other
connected by a rigid coupling and had a common bearing. hand, the accuracy of a trained ANN can never be better
A semi-flexible coupling connected the IP and LP rotors. than that of the measured data with which it is trained.
The common bearing of HP and IP rotor was combined Hence, the proper selection of data for the training of the
journal cum thrust bearing. ANN is also crucial. The large database measured and
stored in a power plant is expected to contain some erro-
neous data also. This may be due to faulty sensors or
3 Basics of ANN recording system or due to some human errors. How these
erroneous data, called ‘outliers’, were removed for this
Artificial neural network mimics the neural structure of the model is discussed in Sect. 4.1. Moreover, selection of data
human brain [38]. The brain basically learns from experi- for training after removal of outliers is discussed separately
ence. ANN consists of a number of interconnected artificial in Sect. 4.2.
neurons with linear or non-linear transfer functions. It can
identify and learn the relationships between the inputs and 4.1 Data pre-processing before training
the outputs of a non-linear multi-dimensional system (i.e.,
from many inputs to many outputs), and therefore, they Data for 33 parameters from the plant were obtained for an
have become very popular for solving problems that are interval of 30 min during a 1-month period. Parameters
difficult to solve by traditional means. There is no included those related to air, fuel (coal), feed water and
impediment having more than one hidden layer, but it has main steam. It was almost impossible to identify all erro-
been proved that one layer of hidden neurons is enough to neous data from such a large database. However, identifi-
approximate any continuous non-linear function if it has cation and removal of outliers as much as possible is
only a sufficient number of neurons [38]. desired. For this purpose, values of parameters were plotted
In this study, feed-forward multi-layer perceptron against time. By careful observation of these plots, most of
(MLP) networks were used. The training was done with the the outliers could be identified. Some outliers could be
back-propagation learning algorithm. MLP consists of an identified instantaneously. For normal operation, continu-
input layer, one or more hidden layer/s and an output layer. ous variation of values of any parameter with respect to
Once the inputs are presented to the network, they will be neighboring points was expected. Thus, any data point,
multiplied by their adjustable weights. Then, they are which is substantially deviating from the general trend of
summed and transferred in the processing elements (neu- variations of its neighboring points, could be suspected as
rons) in order to produce an output. The data used as inputs outliers.
are transmitted through the network, layer by layer, and a Similar investigations were carried out for data com-
set of output/s is obtained. Neural Network has a strong prehending both boiler and turbine parameters. For the
modeling environment for a user to test and explore sim- boiler model, in total, 695 rows (out of 1,440 available
ulated models faster and easier than ever before. Training rows) of data representing variations in operation (as dis-
of the model is done with available data. The input and cussed in Sect. 4.2) were selected for the training of the
output data are introduced to the neural network, and the ANN. In total, 229 rows of data were removed as either
network is trained by using a neural network program. One outliers or for shutdown period, and the rest 466 rows of
such commercially available program for training ANN is data were used for the training of the ANN. During the
NeuroSolutions [39]. When the training is finished with selection of training data variations of different parameters

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 729

related to steam turbine was also considered. Eventually, be used as if ‘real life’ data from the plant to check the
the training data set for the turbine model included the accuracy of the developed ANN model, i.e., validation of
same rows of data as those used for the boiler modeling. the model.
This can be expected since dynamics of boiler and turbine
operation is interrelated through load, main steam proper-
ties and power as presented later. 5 Development of ANN model for power plant

4.2 Selection of training data The structure of ANN was decided from previous experi-
ence as well as by some trial and error. Next step for the
As mentioned in Sect. 4.1, at the beginning, total 695 rows model development was to select suitable input parameters
of data from the plant were selected for the training of the for a decided set of output parameters by different trials.
ANN. This training data set should include data from all The final set of input parameters was decided by ‘opti-
operational conditions. The operational condition of a mizing’ between smaller number of input parameters and
boiler and turbine mostly varies with the load to the boiler. desired accuracy in prediction by the ANN through sensi-
Other operating parameters are set to their values according tivity analysis. Finally, the accuracy of the developed
to the load. In order to choose training data to represent all model was validated by using data set, which was not used
possible ranges of operations, parameters were plotted during training and thus was ‘unknown’ to the trained
against time and their variations were observed. Training ANN. This gave an assessment of the expected perfor-
data were selected from two periods covering the opera- mance of this model in real life use. These steps of the
tional range of plant as marked by dashed vertical lines in model development are discussed later.
Fig. 2 showing mass flow rate of fresh steam and temper- The final objective of this work was to develop a unified
ature of feed water. In these periods, most of the parame- model of the boiler and the steam turbine to represent the
ters showed good variations simultaneously. It should be whole power plant. The unified model would predict the
noted that all these variations were shown for a period of main steam properties (mass flow rate, pressure and tem-
25 days approximately. Data for about rest 5 days were perature) and power output of the plant with inputs to the
removed as those were during the two shutdown periods. boiler model and some other additional input parameters to
Finally, suspected outliers were removed from these two the turbine model as discussed in the following paragraphs.
selected zones to obtain the final training data set that Predicted values with boiler model were used for the pre-
comprised 466 rows of data. Remaining 745 rows of data diction of the power output by the turbine model rather
(outside the selected regions in Fig. 2) were preserved for than using measured values of these.
the validation of the trained ANNs. These data were not Input parameters for a decided set of output parameters
processed by removing suspected outliers as these would are automatic choice in mathematical modeling by the

Fig. 2 Variations of mass flow 720


rate of steam and temperature of
700 training
feed water with time and
selection of training data sets 680
m (t/h)

660
validation
s

training
validation

640 validation
620

600
0 5 10 15 20 25
time (days)
230
validation

220
210 training
(°C)

200
190
fw

validation
t

180 training validation


170
160
0 5 10 15 20 25
time (days)

123
730 Author's personal copy Neural Comput & Applic (2010) 19:725–740

equations describing corresponding physical processes. decided based on system knowledge of the authors and
However, the selection of input parameters in ANN mod- their previous experiences with ANN modeling [29] that
eling emerges differently, since the choice of input and parameters of the main steam (ms, ps, ts), condenser pres-
output parameters is not constrained by the ‘‘cause and sure (pCN) and those related to bleeds from the turbine
effect’’ relations like mathematical modeling. However, the would be included in the initial set of input parameters. The
experience and expert knowledge of underlying relations final set of input parameters for turbine model emerged
between parameters help to initially decide input parame- from this set through two-stage sensitivity analysis.
ters for a set of output parameters, though the input
parameters are usually ‘optimized’ on the basis of com- 5.2 ANN structure and training
promise between the number of parameters and the desired
accuracy of prediction by the ANN, called sensitivity A feed-forward ANN model with back-propagation learn-
analysis as discussed in Sect. 5.3. As a result, the number ing algorithm was used for this work. It was a fully con-
of input parameters for ANN model is usually lesser than nected multi-layer perceptron (MLP) with one hidden
that required for a mathematical model of the same system. layer. Based on previous experience [29] and subsequent
This also has the advantage of avoiding sensor errors with a trials, hyperbolic tangent transfer function was found to be
greater probability as lesser number of measured input the most suitable one and was used for these ANNs. The
parameters are required for ANN modeling. It makes training of the ANN was carried out with commercially
possible also to develop ANN model without some available software, called NeuroSolutions. Data were pre-
parameters as inputs that have consistent errors in their randomized before divided into different groups, viz.,
measured values as discussed in Sect. 5.3.1. This is an training data set, cross validation data set and test data set.
important advantage relative to mathematical modeling for While dividing the data into these sets in NeuroSolutions,
which it is never possible. some initial rows were used for the training (60%), some
following rows were for cross validation (15%) and the rest
5.1 Selection of output and input parameters were test data (25%). The normalization rate of transfer
functions was [-0.9, 0.9]. The training of the boiler ANN
It was known from previous experience [29] that mass flow model was performed with a variation of 1–17 neurons and
rate of fuel, pressure and temperature of feed water were 4,000 epochs. The training was repeated three times with
expected to influence the prediction of main steam prop- the same pre-randomized data set but with randomly cho-
erties. So, they were included as input parameters in boiler sen set of weights in each run in order to avoid local
model. Moreover, the valves before and after the boiler minima. In final model, the best solution converged at 9
were also independent control equipment. According to neurons and 4,000 epochs.
expert knowledge, valve positions of both valves before The training of the turbine ANN model was performed
and after the boiler are expected to influence the output with a variation of 1–18 neurons and 10,000 epochs.
parameters [35]. However, due to unavailability or con- Similarly, the training was repeated three times with the
stancy of data for some of these input parameters, the same pre-randomized data set. In final model, the best
initial input parameters chosen were mass flow rate of coal solution converged at 10 neurons and 10,000 epochs.
(mcoal), valve openings at the outlet of the boiler (UB,out),
pressure of the feed water (pfw) and temperature of the feed 5.3 Sensitivity analysis
water (tfw). These were the input parameters for the stage
one of the trial as discussed in Sect. 5.3.1. Data for some The objective of sensitivity analysis was to decide the set
other parameters related to air flow and flue gas were also of input parameters for the finally accepted model out of
available. Effects of these parameters on the prediction of those initially assumed. During this analysis, the prediction
output parameters were carried out at the second stage of accuracies of output parameters for the initially assumed
the sensitivity analysis as discussed in Sect. 5.3.2. The final input parameters were used as reference standard. Then,
set of input parameters for the boiler model was decided on one input parameter is removed each time, and the ANN
the basis of these two-stage sensitivity analysis and final was retrained with identical structure and the same pre-
validation of the model as discussed later. randomized data set. By using the same randomized data
The selection of suitable set of input parameters for the set, it was assured that the same data was used in training,
prediction of the power output with the turbine model was cross validation and testing data sets for all ANNs. This
not obvious. In fact, a lot of parameters are involved with provided the common platform for the comparison of
the operation of steam turbines. In order to select suitable results for different cases.
input parameters, all available parameters, for which data To find all possible parameters that influence the pre-
were available from the plant, were examined. It was diction accuracy, it is recommended to include large

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 731

number of input parameters initially. So influence of most operators that data for mcoal were measured from the feeder
of the available parameters on the prediction accuracy was speed. It was assumed that coal flow was always propor-
tested. However, as the assumed initial input parameters tional to the feeder speed. At 100 rpm, 7.5 t/h of coal were
were many, sensitivity analysis was carried out in two supplied to the boiler. The mill level was maintained by
stages for both boiler and turbine model. Optimum set of monitoring the mill sound level. When mill became empty,
input parameters was selected through the first stage of sound level went high. The feeder speed was adjusted
sensitivity analysis by ‘top-down’ approach as discussed in accordingly to maintain the desired flow rate of coal. Thus,
Sect. 5.3.1. The sensitivity analysis for the second stage the measured values of coal flow were not accurate, though
started with the optimum solution emerged from this first it served the purpose of running the plant. Moreover,
stage and then adding more input parameters to test their though the basic fuel for the plant was coal, there was some
relevance, i.e., by ‘bottom-up’ approach. Results of these support of oil firing when there were some problems in
two-stage sensitivity analyses are reported in Sects. 5.3.1 maintaining the flame inside the furnace either due to
and 5.3.2. quality of coal or due to some other operational conditions.
Four oil guns thus fired occasionally and no data could be
5.3.1 First stage sensitivity analysis found for duration and rate of this oil firing. Thus, the used
values of mass flow rate of fuel as obtained from the plant
The results of first stage sensitivity analysis are summa- were grossly erroneous, which was confirmed by plotting
rized in Fig. 3. In this figure, mean relative errors (MREs) the measured mass flow rate of fuel with mass flow rate of
(in vertical axis), of ANN predictions of steam properties steam on the same time scale. It was decided that the mass
and power output, respectively, are plotted in two diagrams flow rate of fuel would be dropped, as it did not show any
for different combinations of input parameters. It is obvi- sensitivity. Finally, the set of three parameters (UB,out, pfw,
ous from first diagram that pressure and temperature of tfw) emerged as the ‘optimum’ input set after the first stage
feed water had maximum sensitivity and should be inclu- of sensitivity analysis. The accuracy of prediction (second
ded. On the other hand, absence of outlet valve openings as set of points from the left in first diagram) was acceptable
input parameter affected the prediction of mass flow rate of for this set of three parameters. This showed the great
steam significantly with respect to the reference case, i.e., advantage of ANN model which would predict accurately
including all four parameters (left most set of points in first even without values of some parameters that could not be
diagram). Most unexpectedly, the mass flow rate of fuel did measured due to some reasons.
not influence the prediction accuracy as shown by the Results of the first stage of sensitivity analysis for tur-
second set of points from left. To explain this apparently bine model are shown in the second diagram of Fig. 3. The
unexpected result, the accuracy of data for mass flow rate left most point shows the MRE in power prediction rep-
of fuel was investigated. It was informed by the plant resenting reference point as this corresponds to all initially

Fig. 3 Results of first stage 1.4 p Boiler ANN model


s
sensitivity analysis for the boiler 1.2
MRE (%)

ts
and turbine ANN models 1 m
s
0.8
0.6
0.4
0.2
fw
fw
fw

fw
fw

,p
,t
,t

,t
,t

t
ou

t
ou
fw
fw

fw

B,
,p
,p

B,
,p
t

t
ou

ou


al
B,

B,

co
m

al

al
co
m

co
m
al
co
m

0.21 Turbine ANN model


Power
MRE (%)

0.19
0.17
0.15
0.13
6

5
6

6
6

6
5,

3,
5,

3,
5,

3,
5,
3,

2,
2,

2,
3,

3,

,B
2,

1,
1,

1,
2,

1,
1,

,B

,B
,B

,B
,B

N
,B

C
t
,p
ou
N

N
N

N
N
C

C
C

C
,p

,p
,p

,p
C
N

,p

B,
t

t
t

t
t
ou

ou
ou

ou
C

ou
t
,p
ou


B,

B,
B,

B,
B,
B,

m


s

,
m

m
m

m
s

s
s
m

,t
s

s
m
s

,
,

,
,

p
s
,t

,t
,t

,t
,

s
s
,t

s
s
,t
s

p
p

p
s

s
s
p

s
s
p
s

123
732 Author's personal copy Neural Comput & Applic (2010) 19:725–740

selected input parameters. Rest points correspond to trials increased accuracy. At this stage, it was a bit confusing to
carried out during the sensitivity analysis. It should be decide the final set of input parameters for boiler model, as
mentioned that no bleed was noted for the extraction point the ‘optimization’ was not obvious. It was decided that the
4 and thus not included in the trial. It was concluded from final set of input parameters would be selected by com-
results of this analysis that bleed data corresponding to paring the prediction accuracy for these combinations of
extraction point 3 was most important and that for input parameters with plant data that were not used during
extraction point 6 was also significant to some extent. training, i.e., by validation of the developed model as
Thus, from this first stage of sensitivity analysis, it was discussed in Sect. 5.5. This seemed most logical, as it
resolved that bleed data of two extraction points, i.e., B3,6 would indicate the expected accuracy of the model in real
would be included as input parameters. life.
In the case of turbine model, second stage of sensitivity
5.3.2 Second stage sensitivity analysis analysis was carried out with two-bleed steam data always
included as input parameters. At this stage, intention was to
In second stage of sensitivity analysis for the boiler model, examine the sensitivity of other five (ps, ts, ms, Ut,in, pCN)
the ‘optimized’ set of three input parameters of the first parameters as shown in the second diagram of Fig. 4. The
stage was always included. In addition, other new param- sensitivity of each was tested by removing one of these five
eters related to air and flue gas were included to check their at a time as an input parameter and checking the corre-
sensitivity. Exhaustive search was carried out with differ- sponding accuracy in prediction. MREs in predicting the
ent possible combinations of these new parameters at power output for a combination of input parameters (shown
second stage. The results of this stage of sensitivity anal- in x-axis) are represented by each data point. It is obvious
ysis are summarized in Fig. 4 for both boiler and turbine from this figure that mass flow rate of steam (ms) and the
ANN models. In the first diagram, the reference case condenser pressure (pCN) were the two most significant
(optimized solution of the first stage) of boiler model is parameters identified at this stage of sensitivity analysis.
shown as the left most set of points, named collectively However, sensitivities of other three parameters (ps, ts,
‘stg1’. Only results for significant combinations of new UB,out) were not that significant. Unlike mathematical
parameters with previous three are presented in that dia- modeling, the number of input parameters was judiciously
gram. It was obvious that all these combinations of new ‘optimized’ between the number of parameters and corre-
input parameters with ‘stg1’ improved the prediction sponding accuracy. These three parameters (ps, ts, UB,out)
accuracy, though the improvements were not very signifi- could be dropped as inputs if this developed turbine model
cant. However, in ANN modeling, number of input would be used independently with measured values of all
parameters has to be ‘optimized’ with respect to the input parameters. However, the final objective was to

Fig. 4 Results of second stage 0.7 Boiler ANN model


sensitivity analysis for the boiler 0.6
ps
MRE (%)

t
and turbine ANN models s

0.5 m
s

0.4
0.3
0.2
FG

FG

FG
FG

SA
FG
g1

,m

,t
,m

,t
,m

,m
st

SA
en
FG

,m
FG
FG

en

ig
,t

,t
,t

ig

ox

g1
,C
ox
en

SA
en

,C
,t

st
ig

ig

SA
ox

ox

SA

,t
SA
,C

,C

,m

,t

SA
g1
SA

,m
g1

SA
,t

,m
st

g1
st
SA

g1
,m

st
st
g1
st

0.3 Turbine ANN model


Power
MRE (%)

0.25

0.2

0.15
6
6

6
6

6
3,
3,

3,
3,

3,

3,

3,
,B
,B

,B
,B

,B

,B

,B
t
ou
N
N
N

B,
N

N
C
C

,p
C

,p

C
,p

C
,p

,p

,p
t
t

t
ou
ou

ou

ou


,m
s

m
B,

s
B,

B,

B,

,m
s

t,

,t

s
,t
s
p
s
m
,m

,m

,t
s
s

p
s
t,

p
s
,t

p
s
s

s
p
s

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 733

combine the boiler model with that of the turbine. Values


of two of these three parameters (ps, ts) would be supplied
by the boiler model to the turbine model, rather than using
their measured values. The other parameter (UB,out) would
be used as an input to the boiler model as reported previ-
ously. It seemed logical to include it as an input to the
turbine model also as its measured values were required for
the boiler model and hence available for its use in the
turbine model too. Hence, these three parameters (ps, ts,
UB,out) were included as inputs as they improved the pre-
diction accuracy (second diagram in Fig. 4), though not
significantly. Finally, a set of seven input parameters was Fig. 5 Integration of boiler and turbine ANN models for the overall
used for the prediction of power output. power plant

5.4 Integration of the boiler and turbine models


model (pCN, B3,6). Thus, the model would be useful both
Values of three input parameters (ps, ts, ms) would pass online and offline. However, it should be noted that value
from the boiler model to the turbine model in the integrated of UB,out is used by both models. Moreover, bleed data
model for the power plant as discussed earlier. Measured include both pressure and temperature of bleed steam and
values of the valve opening (UB,out) in the main steam line hence B3,6 indicates four parameters in total, i.e., pressures
between the boiler and the turbine would be used as inputs and temperatures of bleed steam for extraction points 3 and
to either of these models as it was an independent control 6. Thus, in the integrated model, the power output of the
parameter for both of them. power plant would be predicted by specifying values of
In previous sections, development of boiler and turbine only eight input parameters.
ANN models is discussed. These are independent models.
Predictions of output parameters could be obtained by 5.5 Validation of the power plant model
specifying values of corresponding input parameters.
However, in actual operation of a power plant, both boiler The developed ANN model is useful if it predicts the
and turbine operate with a final common objective of values of output parameters accurately with the real life
delivering power. Steam is generated from the feed water values of input parameters. To check this aspect of the
and reheated again in the boiler. The main steam as well as trained ANN, it was provided with real measured values of
the reheated steam from the boiler is used in the steam input parameters from a data set, which was not used
turbine to generate power. Thus, the main output from the during the training of the ANN (i.e., validation data sets in
boiler (i.e., steam) is the required input to the steam turbine Fig. 2). Thus, values of these sets of input and output
for its delivery of power. The operation and control of the parameters were not already ‘known’ or ‘seen’ by the
overall power plant are thus decided mostly on the basis of ANNs. To ensure a situation closer to the real life, these
the combined functions of these two main components sets of data were not filtered for outliers. By comparing the
toward the generation of power. In actual power plants, predictions of the developed ANNs with the actual mea-
values of many parameters are measured for its monitoring, sured values of output parameters, an assessment of
and control strategies are decided accordingly. However, it expected performance for real life applications can be
may be noted that some of the input parameters of the obtained. Figure 6 presents validation of boiler model with
steam turbine model are nothing other than output param- corresponding errors for some combination of input
eters of the boiler model. The proposition of the integration parameters. Validation of additional models with different
of these two models is through these common parameters. inputs was performed due to hard decision on final set of
Instead of providing measured values of all input param- ‘optimum’ input parameters that emerged suitable after the
eters to the steam turbine model, values for steam prop- two stages of sensitivity analyses as described in Sect. 5.3.
erties (ps, ts, ms) for the turbine model could be those Only those with comparable accuracy are shown in Fig. 6.
predicted by the boiler model. This integration of two Best prediction was obtained for the set of input parameters
independent models toward a unified model for the overall (UB,out, pfw, tfw), which emerged as ‘optimum’ set after the
plant is shown in Fig. 5. In this integrated model, the power first stage of sensitivity analysis. Since the boiler and tur-
output of the plant would be predicted by specifying values bine ANN models can be applied independently as well as
of three input parameters for the boiler model (tfw, pfw, one integrated model, the results of validation for each
UB,out) and two other input parameters for the turbine option are shown in Table 1.

123
734 Author's personal copy Neural Comput & Applic (2010) 19:725–740

Fig. 6 Boiler ANN model


1 ps
validation results with data not

MRE (%)
ts
used during training and final 0.8
m
decision on ‘optimum’ set of s
0.6
input parameters
0.4

fw
fw
fw

fw
,t
,t
,t

,t
fw

fw
fw

fw
,p
,p
,p

,p
t
t
t

t
ou

ou
ou

ou
B,

B,
B,

B,
φ



l
l

SA
a
a

,t
co
co

,m
m

SA
m
SA
,t
SA
m
Table 1 Validation results of the boiler and turbine ANN models as Since the validation data set was not filtered, the max-
individual ones and the validation results of the integrated ANN imum errors were additionally investigated. It was con-
model
firmed that most of the errors for all observed parameters
ANN model Error \1 1–2 2–3 3–4 [4 MRE Max. higher then 2.5% were outliers in either input or output sets
(%) (%) (%) of parameters that were measured. Therefore, the MREs
Boiler ps 736 9 0 0 0 0.32 1.47 are more representative for final decision on ANN model
(individual) ts 494 191 49 11 0 0.86 3.93 applications rather than error distribution results. Further, it
ms 628 101 9 3 4 0.60 8.39 should be emphasized that accuracy of ANN models (both
Turbine Power 595 86 49 9 6 0.49 6.85 individual ones and integrated) could be significantly
(individual) improved if better representation of power plant operation
Integrated Power 551 143 44 4 3 0.77 6.41 was recorded. Namely, the data used for training were
30 min snapshot readings that do not represent plant’s
operation well, especially in transients. Furthermore, in this
There were 745 rows of data available from the plant case, the quality of sensor readings in terms of sensor
that were used for the validation of the developed models. installations and uncertainties was also unknown.
From Table 1, it can be seen that the error distribution of The integrated model can serve the purpose of some
this validation study was satisfactory. The MREs are 0.32, additional convenience. It is useful for general-purpose
0.86 and 0.60% for pressure (ps), temperature (ts) and mass assessment of expected power output from the plant for
flow rate (ms) of fresh steam, respectively. steady base line operation as well as for the training of
However, the accuracy of prediction of power output by staff. The accuracy in prediction of power by this inte-
the integrated model was expected to differ from that grated model, as shown in Table 1, is also quite good and
predicted by the independent turbine model as the pre- hence acceptable. Moreover, the option for using inde-
dicted values of steam properties (ps, ts, ms) by the boiler pendent boiler and turbine models for more accurate pre-
model could be different from the measured values of these dictions still remains as and when necessary.
parameters. The integrated model is useful if the accuracy
of it is acceptable. Comparison of measured power with
that predicted by the integrated model is summarized in 6 Interpolation and extrapolation studies of power
Table 1. It shows that errors in predicting power are rela- plant ANN model
tively higher than those predicted by the independent tur-
bine model. In the case of integrated model, the MRE of During training of an ANN, the underlying relations
predicted power is 0.77%, whereas independent turbine between values of input and output parameters are cap-
model yields 0.49%. In independent turbine model, values tured. ANN trains itself with known values of input and
of all input parameters were actual measured values from corresponding output parameters according to chosen
the plant. On the other hand, in integrated model values of learning algorithm. This is done on the basis of numerical
steam properties were provided to the turbine model by the values of input and output parameters without having any
boiler model, rather than directly measured values. Thus, relevance with the physical aspects. However, real life
errors in predicted steam property values additionally systems have different operational ranges for which types
affect the prediction of power. Therefore, the coupling of of operations can be quite different, e.g., full and part load
two independent models into a unified model for the power operations of any power plant are different from opera-
plant led to greater inaccuracy in the prediction of power. tional viewpoint. For ANN, only the underlying relations

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 735

for transformation of values of input parameters to corre- capability of the ANN for this plant. However, for dem-
sponding values of output parameters are relevant. A onstration, data were collected for almost 2 months for a
trained ANN with data for a range of operation may predict different period than that used above with shorter mea-
output parameters for a different range also. The advantage suring intervals (10 min). Total 7,621 rows of data were
of this is that the ANN does not need data for all ranges of selected for all the parameters involved. The power output
operations for its training, though it would be able to variation for this data set is shown in the first diagram of
predict for those ranges too. This is called the ‘interpola- Fig. 7. It shows that this plant practically operated at steady
tion’ or ‘extrapolation’ capability of a trained ANN as it is load though there are some variations of power output as it
similar to numerical interpolation or extrapolation, was expected in real life operation.
respectively. Interpolation is the prediction by an ANN for To check the interpolation capabilities of the ANN for
a range of operation lying in between the ranges of oper- this plant, all rows of selected data were divided into two
ation of training data. Extrapolation is the similar predic- groups. Data for power above (including) 210 MW and
tion of parameters beyond the range of operation of below (including) 190 MW were used for the training of
training data set, either for lower range (called ‘downward the ANN (as shown in the first diagram of Fig. 7). On the
extrapolation’) or for upper range (called ‘upward extrap- other hand, data for power in between these two limits
olation’). These aspects of an ANN are very significant for were preserved for the validation of the interpolation
prediction of power from a plant. A power plant runs at capability of the trained ANN. In total, 4,343 rows of data
different loads and predictions are possible for all loads by were used for training and rest 3,278 rows of data were
the same ANN if the trained ANN has the interpolation used for the validation study. Training data set was ran-
and extrapolation capability. A study was carried out to domized, and the same randomized pattern of data set was
check these aspects for the boiler and turbine models and used for the training of ANNs for both the boiler and the
for the integrated model too. Observed results are presented turbine. The training of the boiler ANN was performed
later. with a variation of 2–20 neurons and 5,000 epochs and one
hidden layer. The training was repeated three times with
6.1 Interpolation results the same pre-randomized data set. The best solution con-
verged at 14 neurons and 5,000 epochs. Table 2 shows the
For this specific plant of India, the principle of operation observed accuracies for the boiler ANN after training
was to maintain steady power output, as far as possible, at procedure as well as for the validation of trained ANNs. It
the economic load. As this plant supplies power to a is obvious from this table that interpolation capability of
national grid, this is a better option to always run it at the ANN for the boiler is acceptable. Namely, the MREs of
maximum efficiency. Thus, the data available from this validation data set that contained approximately 25 days of
plant were always within a small variation of power output data for all predicted parameters (i.e., ps, Ts, ms) are less
values. This led to the difficulty for testing the interpolation than 1%.

Fig. 7 Selection of training and 230


validation data for the study of training
220
interpolation capability of
Power (MW)

210
integrated ANN model and its
200
validation validation
190
180
170 training
160
0 5 10 15 20 25 30 35 40 45 50 55
time (days)
error
reference 4
predicted 3.5
210
Power (MW)

3
error (%)

2.5
200
2
1.5
190 1
0.5
180 0
0 5 10 15 20
time (days)

123
736 Author's personal copy Neural Comput & Applic (2010) 19:725–740

Table 2 Training and validation results of boiler, turbine and inte- 0.9], and it was found the best at [-0.3, 0.3] in all cases.
grated ANN models for interpolation study Observed results are discussed later.
Boiler ANN model
6.2.1 Upward extrapolation results
Training Validation
Error (%) ps ts ms ps ts ms For upward extrapolation, all available data for power less
MRE 0.51 0.65 0.43 0.71 0.76 0.45 than (including) 207 MW were used for training and rest
Max. 1.95 4.43 4.77 2.14 6.11 3.05 data, i.e., for power above 207 MW were used for the
validation of the trained ANNs as shown in the first dia-
Turbine ANN model
gram of Fig. 8. As a result, total 2,012 rows of data were
Training: Power Validation: Power used for training and 5,609 rows of data were used for the
validation of trained ANN. The training of the boiler ANN
MRE 0.20 0.29
was performed with a variation of 2–20 neurons,
Max. 1.27 1.63
8,000 epochs and one hidden layer. The training was
Integrated ANN model repeated three times with the same pre-randomized data set
Validation results: Power
but with a different set of weights in order to avoid local
minima. The best solution converged with 18 neurons and
MRE 0.47 8,000 epochs. Observed accuracy for this boiler ANN is
Max. 3.19 summarized in Table 3. The MREs for validation data set
are 0.64, 1.06 and 0.47% for ps, ts and ms, respectively. The
prediction for ts is a bit higher in comparison with other
results, but overall performance of the boiler model is
Similar training and validation procedure were carried acceptable and the capability of ANN for upward extrap-
out for turbine model also. Observed accuracy is summa- olation can be confirmed.
rized in Table 2. The MRE of validation data set for this Similarly, the training and validation of turbine model
turbine ANN is only 0.29%. Acceptable interpolation were also carried out with the same data sets. The best
capability for the turbine model was thus also proved. solution converged with 16 neurons and 8,000 epochs.
In the last case, the integrated model for the overall Observed accuracy for this turbine ANN is summarized in
plant was studied with the same validation data set. Results Table 3. The MRE in power prediction for validation data
are shown in the second diagram of Fig. 7 for all data set is 0.56%, and hence the turbine model has also proved
points, and the MREs are summarized in Table 2. The good capability in upward extrapolation.
accuracy of integrated model for the power plant was The upward extrapolation study was then extended for
expectedly deteriorated in comparison with independent the integrated model with same set of validation data.
turbine model as the boiler model inaccuracy had also its Results for the power plant model are shown in the second
effect. However, the MRE of integrated model is still under diagram of Fig. 8 for validation data set, and it is also
0.5%, which can be considered as very good. The inter- summarized in Table 3. Expectedly, 0.94% MRE for
polation capability of the integrated model has been power prediction by integrated ANN model is higher than
proven successful, and thus it is acceptable for real life in the case of independent turbine model, but still
applications. acceptable.

6.2 Extrapolation results 6.2.2 Downward extrapolation results

To check extrapolation capability, studies were carried out The study of downward extrapolation was also carried out.
for both ‘upward extrapolation’ and ‘downward extrapo- All available data for power greater than (including)
lation’. When using ANN in extrapolation applications, the 200 MW were used for training and rest data, i.e., for
normalization rate should be considered [37]. If the oper- power less than 200 MW were used for the validation of
ational range is fully covered, the normalization rate can be the trained ANNs as shown in the first diagram of Fig. 9.
close to [-1, 1] (i.e., the transfer function boundaries). On As a result, total 6,863 rows of data were used for training
the contrary, when the data partially covers the operating and 758 rows of data were used for the validation of
range, there is a need for extrapolation capability and the the trained ANNs. The training of the boiler ANN was
data should be normalized within a smaller range to pro- performed with a variation of 2–20 neurons and
vide room for extrapolation. In this study, the normaliza- 8,000 epochs. The best solution converged with 8 neurons
tion rates were tested in a range from [-0.2, 0.2] to [-0.9, and 8,000 epochs. Observed accuracy for this boiler ANN

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 737

Fig. 8 Selection of training and 230


validation data for the study of validation
220
upward extrapolation capability

Power (MW)
210
of integrated ANN model and
200
its validation
190
180 training
170
160
0 5 10 15 20 25 30 35 40 45 50 55
time (days)

225 error 4
reference
220 predicted

Power (MW)

error (%)
210
2
200

190

180 0
0 5 10 15 20 25 30 35 40
time (days)

Table 3 Training and validation results of boiler, turbine and integrated ANN models for extrapolation study
Upward extrapolation Downward extrapolation
Boiler ANN model
Training Validation Training Validation
Error (%) ps ts ms ps ts ms ps ts ms ps ts ms

MRE 0.55 0.67 0.54 0.64 1.06 0.47 0.57 0.70 0.41 0.93 0.67 1.81
Max. 2.26 5.70 3.67 2.06 5.50 2.28 2.58 4.61 2.20 2.85 6.29 11.45
Turbine ANN model
Training: Power Validation: Power Training: Power Validation: Power

MRE 0.25 0.56 0.22 1.34


Max. 1.41 2.25 1.38 9.98

Integrated ANN model


Validation results: Power Validation results: Power

MRE 0.94 2.51


Max. 3.19 14.79

is summarized in Table 3. The MREs for validation data accuracy for this turbine ANN is summarized in Table 3.
set are 0.93, 0.67 and 1.81% for ps, ts and ms, respectively. With the MRE of 1.34% for validation data set, the accu-
The error in prediction of ms is more than 1% and, there- racy is higher in comparison with upward extrapolation.
fore, not within desired limit. It is interesting to note that The downward extrapolation study was then extended
the MREs are small for all parameters both during training for the integrated model with same set of validation data.
and validation of the model. The possible causes for higher Results for the power plant model are shown in the second
errors are discussed in the following section. diagram of Fig. 9 for validation data set, and the summary
The training and validation of turbine model were also can be found in Table 3. The MRE for this integrated
carried out with the same data sets. The best solution model is 2.51%, which is expectedly higher than in the case
converged with 16 neurons and 8,000 epochs. Observed of individual turbine model.

123
738 Author's personal copy Neural Comput & Applic (2010) 19:725–740

Fig. 9 Selection of training and 230


validation data for the study of 220 training

Power (MW)
downward extrapolation 210
capability of integrated ANN
200
model and its validation
190
180
validation
170
160
0 5 10 15 20 25 30 35 40 45 50 55
time (days)
error
reference 15
predicted
200

Power (MW)

error (%)
10

180

160

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5
time (days)

6.2.3 Discussion on extrapolation results smaller than by validation, a similar conclusion can be
made.
Due to some higher MREs, especially in model predictions Due to higher MREs in downward extrapolation, which
of validation data sets, the data were additionally exam- comprehend wider region of operation, a question about
ined. Firstly, validation data set was not filtered in order to missing mass flow rate of fuel arises. The correlation
simulate real life monitoring. It was confirmed under closer between mass flow rate of fuel and of fresh steam is, in
inspection of the data that operational points, where the general, very high, i.e., close to 1. Therefore, it could be
highest errors occurred, are mostly due to outliers. Hence, expected that having accurate data of mass flow rate of
in the process of validation, the MREs are more repre- fuel, the prediction of steam properties in extrapolation
sentative to look at. study could be improved.
However, in the case of downward extrapolation, the Another interesting feature can be noticed that despite
MREs are significantly higher in comparison with upward well-represented training data set (6,863 rows) in down-
extrapolation. It was found that these errors, in most of the ward extrapolation in comparison with upward extrapola-
cases, cannot be related to outliers. On one hand, it was tion (2,012 rows), the accuracy is still better in upward
observed that high MREs are mainly due to rather opti- extrapolation with much less data used for training. Con-
mistic selection of training and validation data sets. In the clusion can be drawn that extrapolation capability of ANN
case of downward extrapolation, the boundary between is not dependent on largeness of amount of training data
training and validation data set is at 200 MW that makes but rather on careful selection of data that comprehend
extrapolation range rather large, i.e., down to approxi- adequate operational range with sufficient representation of
mately 165 MW. On the contrary, in upward extrapolation, transients.
the boundary is set at 207 MW and the extrapolation
reaches approximately 222 MW. In latter case, the range of
15 MW is much narrower than at downward extrapolation 7 Conclusions
that is 35 MW. It follows that in downward case, the
selected training data set also comprehends relatively Results of an extensive study of artificial neural network
narrow range of operation, i.e., from 200 to 220 MW, modeling of a coal-based power plant with real plant data
whereas in upward extrapolation, it is much wider, i.e., 165 have been reported in this work. This study is one of some
to 207 MW. From the comparison of Figs. 8 and 9, it can similar studies carried out for different coal-fired power
be seen that extrapolations, whether upward or downward, plants. The objective is to investigate the feasibility of
are comparable in smaller ranges considering also the level ANN modeling for coal-fired power plants. Developed
of extrapolated/predicted values. Therefore, the extrapola- ANN model by using already measured data is useful for
tion capability of ANNs can be confirmed taking into different purposes. Separate models were developed for the
account proper data selection for training. Looking at boiler and steam turbine. The step-by-step model devel-
accuracies of models by training, which are significantly opment for both the boiler and the turbine has been

123
Neural Comput & Applic (2010) 19:725–740 Author's personal copy 739

reported. Finally, these two were integrated to a single 160 MW unit. Report TFRT-3192, Lund Institute of Technology,
model where predictions of output parameters of the boiler Sweden
15. Anglart H, Andersson S, Jadrny R (1992) BWR steam line and
model were used as inputs to the turbine model. The fea- turbine model with multiple piping capability. Nucl Eng Design
sibility of ANN modeling for a coal-based power plant has 137(1):1–10
been established. It was concluded that ANN model could 16. Schobeiri MT, Chakka P (2002) Prediction of turbine blade heat
be developed for the prediction of power output from the transfer and aerodynamics using a new unsteady boundary layer
transition model. Int J Heat Mass Transf 45(4):815–829
plant by specifying values of a few input parameters. 17. Bassel WS, Gomes AV (2002) A metastable wet steam turbine
Finally, interpolation and extrapolation (both upward and stage model. Nucl Eng Design 216:113–119
downward) studies were also carried out with plant data for 18. Liu JJ, Cui YQ, Jiang HD (2003) Investigation of flow in a steam
different periods from the same plant. It was concluded that turbine exhaust hood with/without turbine exit conditions simu-
lated. J Eng Gas Turbines Power 125(1):292–299
interpolation accuracy was acceptable. Extrapolation 19. Lampart P, Yershov S (2003) Direct constrained computational
results were also good for the upward extrapolation, though fluid dynamics based optimization of three-dimensional blading
the same for the downward extrapolation was not that good for the exit stage of a large power steam turbine. J Eng Gas
as relatively higher maximum errors for some parameters Turbines Power 125(1):385–390
20. Perez RE, Lopez MAA, Vazquez EM, Littlewood EC, Cruz CA
were noted during validation with data not used during (2004) A comprehensive finite-element model of a turbine-gen-
training. erator infinite-busbar system. Finite Elements Anal Design
40:485–509
Acknowledgments Professor Mohsen Assadi, Department of 21. Boccaletti C, Cerri G, Seyedan B (2001) A neural network sim-
Energy Sciences, Lund University, Sweden and Dr. Sudipta De, ulator of a gas turbine with a waste heat recovery section. Trans
Department of Mechanical Engineering, Jadavpur University, Kolk- ASME J Eng Gas Turb Power 123:371–376
ata, India gratefully acknowledge the financial support for this work 22. Mathioudakis K, Stamatis A, Tsalavoutas A, Aretakis N (2001)
from the Swedish Research Council (Vetenskapsrådet) under Swedish Performance analysis of industrial gas turbines for engine con-
Research Link Program (Grant No.: 348-2006-5349). dition monitoring. Proc Inst Mech Eng Part A J Power Energy
215:173–184
23. Mathioudakis K, Stamatis A, Bonataki E (2002) Allocating the
causes of performance deterioration in combined cycle gas tur-
References bine plants. J Eng Gas Turbines Power 124:256–262
24. Kalogirou SA (2000) Applications of artificial neural-networks
1. Bhambre K, Mitra SK, Gaitunde UN (2007) Modeling of a coal- for energy systems. Appl Energy 67:17–35
fired natural circulation boiler. Trans ASME J Energy Res Tech 25. Kalogirou SA (2003) Artificial intelligence for the modeling and
129:159–167 control of combustion processes: a review. Prog Energy Combust
2. Liu C, Liu J, Niu Y, Liang W (2001) Nonlinear boiler model of Sci 29(6):515–566
300 MW power unit for system dynamic performance studies. 26. Kesgin U, Heperkan H (2005) Simulation of thermodynamic
IEEE Trans Ind Electron 2:1296–1300 systems using soft computing techniques. Int J Energy Res
3. Astrom KJ, Bell RD (2000) Drum boiler dynamics. Automatica 29:581–611
33:363–378 27. Cerri G, Khatry DS (1998) A neural network approach in ther-
4. Adam EJ, Marchetti JL (1999) Dynamic simulation of large modynamic process evaluation. In: Proceedings of the interna-
boilers with natural circulation. Comput Chem Eng 23:1031– tional conference on engineering application of neural network
1040 (EANN-98), Gibraltar (GB), Paper No. 98172
5. Lu S (1999) Dynamic modeling and simulation of power plant 28. Mesbahi E (2000) Artificial neural networks for fault diagnosis,
systems. Proc Inst Mech Eng Part A 213:7–22 modeling and control of diesel engines. Ph.D. thesis, University
6. Lo KL, Song ZM, Marchand E, Pinkerton A (1990) Development of Newcastle Upon Tyne, UK
of a static-state estimator for a power station boiler: part I. 29. De S, Kaiadi M, Fast M, Assadi M (2007) Development of an
mathematical model. Electr Power Syst Res 18:175–179 artificial neural network model for the steam process of a coal
7. Tysso A (1981) Modeling and parameters estimation of a ship biomass cofired combined heat and power (CHP) plant in Swe-
boiler. Automatica 17:157–166 den. Energy 32:2099–2109
8. Usoro PB (1977) Modeling and simulation of a drum boiler- 30. Fast M, Assadi M, Smrekar J (2008) Application of artificial
turbine power plant under emergency state control. Master’s neural network to the condition monitoring and diagnosis of a
thesis, Massachusetts Institute of Technology CHP plant. In: Proceedings of the 21th international conference
9. Kwan HW, Anderson JH (1970) A mathematical model of a on efficiency, cost, optimization, simulation and environmental
200 MW boiler. Int J Control 12(6):977–998 impact of energy systems, Poland, pp 981–988
10. Chein KL, Ergin EI, Ling C, Lee A (1958) Dynamic analysis of a 31. Irwin G, Brown M, Hogg B, Swidenbank E (1995) Neural net-
boiler. Trans ASME 80:1809–1819 work modelling of a 200 MW boiler system. IEE Proc -Control
11. Eklund K (1971) Linear drum boiler–turbine models. Ph.D. thesis Theory Appl 142(6):529–536
TFRT-1001, Lund Institute of Technology, Sweden 32. Lu S, Hogg BW (2000) Dynamic and nonlinear modelling of
12. Brandl P, Reichert K, Vogt W (1975) Simulation of turbogen- power plant by physical principles and neural networks. Electr
erators on steady state load. Brown Bovery Rev 9:444–449 Power Energy Syst 22:67–78
13. Girija Shankar PV (1977) Simulation model of a nuclear reactor 33. Chu JZ, Shieh SS, Jang SS, Chien CI, Wan HP, Ko HH (2003)
turbine. Nucl Eng Design 44(2):269–277 Constrained optimization of combustion in a simulated coal-fired
14. Bell RD, Åström KJ (1987) Dynamic models for boiler–turbine- boiler using artificial neural network model and information
alternator units: data logs and parameter estimation for a analysis. Fuel 82:693–703

123
740 Author's personal copy Neural Comput & Applic (2010) 19:725–740

34. Feretti G, Piroddi L (2001) Estimation of NOX emissions in 37. Fast M (2005) Artificial neural networks for gas turbine modeling
thermal power plants using neural networks. J Eng Gas Turbines and sensor validation. Master Thesis, Lund University
Power 123: 465–471 38. Haykin S (1999) Neural networks, a comprehensive foundation,
35. Smrekar J, Assadi M, Fast M, Kuštrin I, De S (2009) Develop- 2nd edn. ISBN No. 0–13-273350–1. Prentice Hall, Inc, New
ment of artificial neural network model for a coal-fired boiler Jersey, USA
using real plant data. Energy (Oxford) 34:144–152 [Print ed.] 39. Principe JC, Euliano NR, Lefebvre WC (1999) Neural and
36. Fast M, Assadi M, De S (2009) Development and multi-utility of an adaptive systems: fundamentals through simulations. Wiley, New
ANN model for an industrial gas turbine. Appl Energy 86:9–17 York

123

View publication stats

You might also like