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

net/publication/228418236

Power Plant Performance Monitoring Using Statistical Methodology Approach

Article · January 2011

CITATIONS READS

7 2,922

1 author:

Konrad Świrski
Warsaw University of Technology
47 PUBLICATIONS   281 CITATIONS   

SEE PROFILE

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

NARSIS - New Approach to Reactor Safety ImprovementS - Horizon 2020 View project

SARWUT View project

All content following this page was uploaded by Konrad Świrski on 21 October 2016.

The user has requested enhancement of the downloaded file.

 Open Access Journal

Journal of Power Technologies 91 (2) (2011) 63–76

journal homepage:papers.itc.pw.edu.pl

Power Plant Performance Monitoring Using Statistical Methodology

Approach
Konrad Świrski∗
Institute of Heat Engineering, Warsaw University of Technology
21/25 Nowowiejska Street, 00-665 Warsaw, Poland

Abstract
The paper sets out problems experienced in current power plant performance calculation methodology. With
changes in the electricity market and the increased role of computer control and diagnostic system, the previ-
ous methodology is fast becoming obsolete. Several key factors are highlighted (for example, reference values
and corrective curves) which may lead to highly imprecise and inadequate evaluations of plant performance
and, especially, operator behavior. It is important to bear in mind the systemic changes on the energy mar-
ket. The current methodology focuses on controlling factors impacting power generating efficiency whereas
the modern market is more complex and real costs are largely dependent on other variable costs, such as
environmental fees, equipment wearing costs and the energy trading market .
Statistical analysis of data is proposed as a modification of the current methodology. As most heat rate
calculation is done on-line, it is possible to analyze this data in detail and apply, for example, PCA (Principle
Component Analysis) and linear (and nonlinear) regression, thereby enabling a more accurate determination
of the influence of principle process parameters on heat rate deviation. The article presents sample results of
comprehensive analysis (two cases from different plants) of two twin units (heat rate calculation and process
data analysis from a 12-month period) which to demonstrate the clear need to modify and update the old
performance calculation approach.

1. Introduction seems outdated in the context of the current dynamic

deregulation of the power generation industry.
The method of performance monitoring used at
present was developed over 30 years ago for units This older approach is becoming less viable as a
operating in the conditions expected for the power true performance index of plant capability due to the
sector of the time. That methodology corresponded two basic factors – advanced computer technology
to the American and Western Europe standards of enabling widespread use of digital automatic control
the 60s and 70s, which put reliability first. This systems and system changes in the energy market.
method delivered significant advantages in the form Digital computer based automatic control systems
of higher quality performance monitoring, but today have made it possible to provide almost non-stop per-
formance control through direct operator supervision
∗
Corresponding author
and the monitoring of all performance parameters
Email address: swirski@itc.pw.edu.pl (Konrad (and losses) on-line. The increase in the quality of
Świrski ) measurement devices and tools has reduced the role
 Journal of Power Technologies 91 (2) (2011) 63–76

of periodic heat rate testing and warranty measure- as follows: main steam pressure, main steam temper-
ments. The high quality of the DCS automatic con- atures, pressure drop in the superheater, reheat steam
trol connected with the more common application temperature, condenser pressure, feedwater tempera-
of optimization systems (really substituting operator ture, oxygen content in flue gas, flue gas temperature.
actions during normal unit operation) has reduced the Whereas the number of controlled parameters
possibility of a simple improvement of efficiency in- has been expanded many times, the theoretical ba-
dexes. For this reason, the principle role of perfor- sis of this method remains the same. The devia-
mance supervision should be modified for the need tion [kJ/kWh](BTU/kW) was usually calculated to a
to detect possible losses from unit operation [unsure] value of dollars/kWhour for a more visual presen-
in market based generation dispatch. tation of data. Systems based on ASME or simi-
This article proposes using statistical data analysis lar methodology were introduced in practically all
to improve aspects of the current performance mon- power plants, with the modernized automatic control
itoring systems. Due to the special characteristics of systems usually developed into on-line systems per-
European power plants, we use as reference a large forming all the calculations every few minutes and
coal fired generation asset both in the presentation of presenting the results on operators’ screens at the
theoretical material and in practical results. DCS or auxiliary computer displays.
It is important to bear in mind the system changes
that have taken place in the energy market. The cur- 3. Performance monitoring – the problems of
rent methodology focuses on controlling factors con- conventional application
nected with the efficiency of the energy production
process (and thus only takes into account the cost of The performance calculation methodology, though
fuel). The market nowadays is more complex and the no doubt necessary and effective when properly im-
real costs (and losses) are largely dependent on other plemented, also has a series of drawbacks. It is ap-
variable costs, such as environmental credits/costs parent that after so many years (and many platform
(emissions and the future influence of emissions trad- revisions to calculate results) it is possible to evalu-
ing), maintenance factors (maintenance and repair ate the results more critically and to attempt a more
costs, and costs connected with the operation of a in-depth analysis.
deregulated energy trade market (e.g. filling in con- The basic problems with the current performance
tracts, possibility of spot transactions, etc.). What is monitoring application:
postulated is modifying the method in the direction
• Reference values – most deviations and losses
of Market Performance Control, as it affects all the
are calculated and monitored against reference
above issues. That, however, lies outside the scope
values – usually the nominal values given by the
of this article.
OEM manufacturer. For devices often with a
10–20 year life cycle and upgraded on numer-
2. Current Performance Monitoring ous occasions these nominal values do not con-
stitute a real reflection of the actual as found pa-
The typical performance monitoring methodology rameters.
is presented in numerous conference materials. In
short, it is based on calculating the unit chemical en- • Correction curves for defining the control-
ergy usage rate (ASME Power Test Codes) and as- lable losses (measured losses) – the influence
signing the measured losses deviations of the unit of operational parameter deviation (tempera-
chemical energy usage rate from the expected value ture, pressure etc.) from the assumed values
(nominal, or resulting from the last design or war- (achievable, design, theoretical. . . ) is assigned
ranty measurements), following from operating the largely using the so-called manufacturer’s cor-
unit at parameters other than the nominal parameters rection curves. Leaving aside the accuracy of
[1, 2]. The basic parameters whose influence over the these curves and the problems commonly en-
unit heat rate is usually taken into consideration are countered in obtaining this data, the basis of

— 64 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Figure 1: Examples of typical operator’s graphics showing the results of on-line performance monitoring calculations for process
operators – Europe (left), US (right)

this theory is to define the influence of these pa- ters, the automatic control systems perform a
rameters (xi )(gradient) as a unit heat rate (qb ) − shift of the unit status into a different operat-
∂qb /∂xi but the manufacturer’s data does not ing point (also modifying the other parameters).
necessarily correspond to the real, dynamic op- Because of this, deviations assigned using cor-
eration of a maintained unit. On the other hand rection curves cease to have any practical sig-
there seems to be a serious theoretical prob- nificance. For example, even if at a given mo-
lem with assigning the deviation for the given ment we assign deviations of a unit heat rate
control value. In the case of building a cor- for a series of main parameters (and we obtain
rection curve, it is assumed that a clear assign- a negative deviation for one of them, resulting
ment of the influence of a given value on the from the difference between the current and the
unit heat rate will be possible (qb ). In other nominal value – referential), then in the case of
words, variables such as pressure, temperature, canceling this difference (bringing the param-
etc are treated as independent variables (this re- eter to the nominal-reference value – and thus
sults, among others, from the method of assign- reducing the deviations), all the other parame-
ing correction curves through balance calcula- ters will remain unchanged (!), and we will ob-
tions and the change of an individual parameter tain a entirely different system of parameters,
in simulation calculations) which finally leads and their differences from reference values, and
to obtaining a dependence ∂qb /∂xi = f (xi ) . as a result completely different values of non-
In actual practice, a strong relationship exists measured deviations.
between these parameters (they are interrelated
– e.g. in the form of a turbine equation) – in • Separating startup from normal perfor-
brief, during normal operation it is not possible mance – the procedures of calculating startup
to change one parameter without modifying a losses are commonly used (and are typical)
number of others. Additionally, assigning rela- – but they allow assignment of higher and
tionships between these parameters is not only predicted losses during startup – and not a
dependent on the thermodynamic dependencies “smooth” shift to losses during normal opera-
(balance) but is also influenced by the operation tion modes.
of the automatic control system controlling the • Applying statistical balance models for as-
unit. In other words – in practice when chang- signing losses during load following unit op-
ing one of the main unit operational parame- eration – models used in performance monitor-

— 65 —
 Journal of Power Technologies 91 (2) (2011) 63–76

ing are based on a strictly static approach and case for given projects seems impractical and
in order to obtain good results they require a devoid of purpose. It may however be mean-
good thermal status (or quasi-static) isolation of ingful to evaluate losses (or the possibility of
the unit operation – in the simplest approach performance improvement) in relation to aver-
this requires a momentary stabilization of unit age values (achieved throughout long-term per-
power and its principle parameters. In the con- formance) or the best practices observed during
ditions of the present (ISO or deregulated mar- performance.
ket) situation this is simply impossible – in fact
the entire regime of condensation unit opera- • Which performance losses are the most sig-
tion actively participating in the power market nificant ones and which can actually be im-
is based on operation during dynamic (ramping proved? In industrial practice the real evalua-
or transitional) states. Using this methodology tion of deviation is important (heat rate) (cost
for temporary (dynamic) performance monitor- of deviation); using a correction curve to as-
ing might seem questionable in light of the typ- signing this value seems far from perfect. In
ical approach for obtaining good global results the past the advantages of the present method
(characteristics, optimization), where it is com- were advertised as, for example, the possibil-
mon to use diverse static processing of perfor- ity of calculating how much it costs to operate
mance data, which averages the results (consid- a unit with parameters different than the nom-
ering the normal distribution of calculation er- inal ones. Even if by nominal parameters we
rors and influence of dynamic states) and can- mean optimal parameters (the best achievable),
cels momentary errors. then by making direct use of the current method
and correction curves we will obtain an ideal-
4. Postulated change ized (and thus fictional and practically nonexis-
tent) solution. Theoretical assumptions of this
This postulation focuses principally on two ba- method assume that the influence of each of the
sic problems: of reference values and the possibil- parameters can be treated individually and that
ity of evaluating performance deviations on the ba- it is possible to change each one of them to the
sis of reference curves. Putting theoretical consid- reference value without changing the other pa-
erations into practical questions, these two problems rameters. This is not possible in normal unit
boil down to the following questions (and attempts performance. It seems purposeful to search for
to derive answers!): a method which would at the same time iden-
• What are the currently achievable unit pa- tify the losses (deviations) that can be reduced
rameters? Units constructed in the 70s and and the real influence of process parameters on
80s have been subject to various upgrades and the effectives of unit operation (taking into con-
maintenance repairs. A large part of the util- sideration real unit characteristics and thus, the
ity generation assets burn fuel which signifi- reaction of automatic control systems).
cantly differs from the original design values.
The basic equipment (boiler, turbine) has been 4.1. Extending the possibility of performance moni-
repaired/rebuilt and modernized. In practice toring through a deep statistical analysis of pro-
this often means that we are faced with a com- cess data
pletely different site than the one originally en- When equipping the power units with a digital au-
visioned by the design team and performance tomatic control system, the heat rate calculations are
monitoring analysis involves large deviations in practice conducted on-line, so we have a large
(both positive and losses) which the process op- volume of calculation data whose statistical analysis
erator is unable to eliminate. These deviations may turn out to be a valuable tool for correct reason-
result from a long-term change in unit opera- ing. Of course, we assume that in the case of per-
tion parameters. Assigning deviations in this formance analysis conducted on-line we are aware

— 66 —
 Journal of Power Technologies 91 (2) (2011) 63–76

of measurement-calculation problems (and ways of sionality reduction in a data set while retaining those
solving them) such as: credibility of measurement characteristics of the data set that contribute most to
devices (it is assumed that the basic measurement its variance, by keeping lower-order principal com-
tools (especially flow measurements) operate cor- ponents and ignoring higher-order ones. Such low-
rectly and are of an appropriate measurement class, order components often contain the "most important"
the influence of delayed chemical analysis of fuel aspects of the data.
(when there are no on-line analyzers), process data This transformation retains all the most important
and results are appropriately processed to eliminate information concerning the original variables. In par-
measurement errors and filtering the non-stationary ticular, the first component shows the direction of
statuses of units, etc. After solving these problems, the biggest dispersal of analyzed variables. Referring
we finally possess a large, credible base of calcula- to this specific analysis of performance data through
tion data and unit heat rate for various unit operat- PCA, it is possible to:
ing states. Some similarity to the proposed approach
may be found in [3–5]. • Assign new variables (fictional) – each is a com-
bination of the basic and process parameters,
4.1.1. Statistical analysis not correlated.
The primary objective of this analysis was to cre- • Assigning the first component of PCA – and
ate histograms and assign mean values of process through its analysis to identify parameters with
parameters and comparing them with process val- the highest changeability.
ues. This will allow one to check how real the per-
formance parameters are (mean values and the most 4.1.3. Analysis using the linear regression model
common ones) against reference values (nominal). The principle objective of the analysis is to build
Thanks to current automatic control systems it is now an empirical (linear at this point) model of unit heat
possible to archive data from a practically unlim- rate in the form of a function f : Rd → R linearly
ited period of performance, and create an unlimited dependent on main technological parameters and de-
database. When assigning the basic reference op- fined by the formula f (x) =< x, w > −γ , directly
erating parameters it is postulated to aggregate the applicable to unit heat rate approximation in the form
data in unit efficiency data (steam flow, power). The as
calculation examples present only the results for an Xm m
X
arbitrarily accepted power range 120–160 MW (low qb = a j x1, j + ... an, j xn, j .
power) and 160–200 MW (high power) (correspond- m− j m− j

ing to the typical performance regimes). In the devel- A model of this kind can make it possible to as-
oped form of course, it is possible to obtain a func- sign the direct influence of a given parameter over
tion of any parameter depending on the efficiency changeability qb (and more precisely, on its empir-
(power) of a unit. ical model). With such an approach we may obtain
a similar (linear) function to that obtaining in most
4.1.2. PCA corrective curves.
PCA – (Principal Component Analysis) is a A linear regression model constitutes the simplest
method which makes use of linear transformation to, empirical approximations of unit heat rate from the
change input variables into new variables (‘princi- basic process parameters. It is possible to construct
ple components’) which are uncorrelated. PCA is an such a model with sufficient accuracy, the next step
orthogonal linear transformation that transforms the is to assign the correlation of basic parameters which
data to a new coordinate system such that the great- directly leads to assigning the influence of these pa-
est variance by any projection of the data comes to lie rameters onto the unit heat rate. Of course, the lin-
on the first coordinate (called the first principal com- ear regression model can then be modified (nonlin-
ponent), the second greatest variance on the second ear models, neural networks, fuzzy networks, etc.)
coordinate, and so on. PCA can be used for dimen- to improve mapping accuracy.

— 67 —
 Journal of Power Technologies 91 (2) (2011) 63–76

5. Calculation examples is off. In boiler B, the optimization system was

practically in continuous operation, resulting in
The conducted tests use calculation data of the unit
better O2 control.
heat rate of two of the Polish power plants. In each
case, analyses were conducted for two similar units
PCA analysis seeks to identify process parameters
– 225 MW of identical construction with twin auto-
with the highest changeability by transforming them
matic control systems. The results (appropriately av-
into a set of independent (uncorrelated) parameters.
eraged and aggregated in appropriate ranges of unit
See below for the first principle component of the
power) of the current performance calculation cov-
two units. The participation of a given process pa-
ering an approximately 12-month period were used
rameter in the principle component is presented by
for the purposes of analyzing the data. To simplify
the bars on the chart – the number of the variable on
the analysis results, the article lists three histogram
the x axis corresponds to main steam pressure (1),
values for two load ranges – low (120–160 MW) and
main steam temperature (2), reheat steam tempera-
high (160–200 MW). The table below presents data
ture (3), pressure drop in the superheater (4), feed-
analysis results for both unit – histograms and princi-
water temperature (5), condenser pressure (6)).
ple statistical measures for appropriate power ranges.
PCA enables swift identification of parameters
The statistical analysis leads to a series of conclu-
leading to the biggest changes in unit heat rate. In
sions:
this specific example, the calculation for unit A, it is
• Even the most similar power units have dif- the temperature of reheat steam for low loads (vari-
ferent performance characteristics and different able no. 3). For unit B, the parameters which have
histograms of the basic parameters. the highest changeability are pressure drop in the
superheater and pressure in the condenser (change-
• In many cases the current process values (ob- ability caused by seasonality), thus one can assume
tained during operation) are significantly differ- that there are no significant performance problems
ent than the nominal values (often assumed as for this unit. In the next step, an approximation of
referential) and can vary significantly depend- heat rate (qb ) was performed using linear regression.
ing on the unit power range (boiler efficiency) – This analysis method is simply an empirical (based
for example, the reheat steam temperature for on historical data) model approximation (prediction)
unit A at low loads has an average value of of the unit heat rate. Although the results obtained
526 ◦ C and very high changeability (standard indicate the possibility of only a very rough estima-
deviation) – case a in the figure in Table 2. tion of qb , model correlation is much higher than the
• Observing the changeability of a certain param- correlation of a single variable. The correlations of
eter (standard deviation) allows one to draw basic parameters allow one to evaluate the influence
conclusions on the level of tuning of the auto- of a given parameter on qb .
matic control system – unit A during low load The results confirm the previous conclusions from
operation has a significant problem with achiev- the statistical analysis. As long as the correlations of
ing the design level of reheat steam temperature process variables for high powers are at a high level
(case a in the figure in Table 2). (none of the process parameters are responsible for
substantial changes in qb ), at a low level of power
• Observing O2 concentration in the boiler (fig- we can see the visible dominant influence of reheat
ures in Table 3) below, we may compare the per- steam temperature (variable no.3), which largely re-
formance of units A and B (and therefore the in- duces unit performance. One may be tempted to test
fluence on boiler efficiency and unit heat rate). the qualitative influence of this parameter. Linear re-
In boiler A we find a characteristic twin peaks gression leads directly to obtaining the dependence
histogram (case b Table 3) – with these boil- (linear) of influence of a given parameter to a change
ers optimization systems are implemented and of correlation of unit heat rate. This values was then
higher O2 values are achieved when the system compared against the data from the correction curve

— 68 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Unit A unit heat rate [kJ/kWh] average I (low load) left – 8850 II (high load) right – 8630
Median I – 8840 II –8638 Std, Dev. I –397 II –174

Unit B unit heat rate [kJ/kWh] average I –8756 II – 8107

Median I – 8787 II – 8425 Std, Dev. I – 353 II – 214
Table 1: Results of statistical analysis of the principle performance parameters – heat rate; analysis from plan #1

— 69 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Unit A twt [◦ C] reheat steam temperature average I –525.2 II – 536.45

Median I – 536.7 II –538 Std, Dev. I – 9.55 II –4

Unit B twt [◦ C] reheat steam temperature average I – 536.8 II – 539.6

Median I – 538.2 II – 539.64 Std, Dev. I – 5.11 II –1.89
Table 2: Results of statistical analysis of the principle performance parameters – reheat steam temperature – analysis from plan #1

— 70 —
 Journal of Power Technologies 91 (2) (2011) 63–76

O2 [%] Concentration average I – 5.61 II – 4.9

Median I – 5.6 II – 4.7 Deviation I – 0.66 II – 0.77

O2 [%] Concentration average I – 5.14 II – 4.58

Median I – 5.14 II – 4.56 Deviation I – 0.45 II – 0.42
Table 3: Results of statistical analysis of the principle performance parameters – O2 concentration- analysis from plan #1

— 71 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Unit A Unit B

Figure 2: Analysis of results using the PCA method – analysis from plan #1

— 72 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Real Data Low Load Correction Curve Real Data High Load
200

150

100

Heat Rate, Btu/kWh

50

-50

Unit A -100

-150

-200
505 510 515 520 525 530 535 540 545 550
Reheat S team Temperature, °C

Figure 4: Heat Rate correction curve qb for reheat steam tem-

perature and the corresponding values from statistical tests –
analysis from plan #1

(manufacturer’s data), obtaining the final dependen-

cies shown in the figure below.
The blue line in the diagram shows the factory cor-
rection curve, the yellow line shows the appropriate
values from linear regression for high power and the
magenta line shows the same for low power. Com-
paring the data in the diagram shows that in a series
of cases it is highly questionable to use correction
curves. This is especially the case with evaluation of
the operator’s work (and the possible merit system),
which does not take into account statistical analysis
and can lead to significant inaccuracies.

• As regards calculation, the basic performance

problem is the continually insufficient heating
level of reheat steam at low powers caused by
either poor tuning of the reheat steam control
system or (more likely) construction faults (re-
building of heated surfaces or change of fuel).

• In this case it is not possible to expect the pro-

cess operator to be able to operate the unit at
Figure 3: Linear regression – correlations of the basic parame- close to the nominal steam temperature (535 ◦ C)
ters with regression models of heat rate qb- analysis from plan
#1 • In real operation of this unit, reheat steam tem-
perature fluctuations cause much larger devia-
tions in the heat rate (qb ) than are shown by the
correction curve.

Similar type analyses were conducted for plan #2 –

also twin 225 MW units with identical control sys-
tems.

— 73 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Unit C, Main steam temp. [◦ C], Low load – left, High load – right

Unit D, Main steam temp. [◦ C], Low load – left, High load – right
Figure 5: Results of statistical analysis of the principle performance parameters – feedwater temperature – analysis from plan #2

— 74 —
 Journal of Power Technologies 91 (2) (2011) 63–76

When analyzing histograms we find that unit C ex- [4] G. Prasad, E. Swidenbank, B. W. Hogg, A novel perfor-
periences problems with main steam control (below mance monitoring strategy for economical thermal power
nominal average temperature level with problems of plant operation, Energy Conversion, IEEE Transactions on
14 (1999) 802 – 809.
set point quality (high deviation) at low loads (case [5] V. V. Kantubhukta, M. Abdelrahman, A feasibility study
A) and high loads (case B). In unit D, at high loads on using neural networks in performance analysis of coal-
these problems do not exist (case D) although at low fired power plants, Dept. of Electr. & Comput. Eng., Ten-
loads average temperature level is lower than nomi- nessee Technol. Univ.,Cookeville, TN, USA; System The-
ory, 2004. Proceedings of the Thirty-Sixth Southeastern
nal and is accompanied by with poor control quality
Symposium on ; Publication Date: 2004.
(case C)
When analyzing feedwater temperature, problems
may be observed with feedwater train operation in
unit D (cases 3 and 4 – frequent lowering of feedwa-
ter temperature), contrary to proper operation as seen
in unit C (cases 1 and 2). Results from histogram
analysis were confirmed with PCA and linear regres-
sion (with correlation) approaches.

6. Conclusion

The calculations of the performance monitoring

method used to date, as used in their present scope,
appear to have exceeded their limits. Modern per-
formance control, having at its disposal tools in the
form of constant efficiency calculations, automatic
control systems integrated with archive systems and
data analysis, is able to deliver a far more detailed
and precise analysis of reasons for reductions in the
quality of efficiency parameters. A particular cause
for concern is the continued application of correc-
tion curves. What is postulated here is a modifica-
tion through applying statistical analysis in a large
scope and, as a result, systems of automatic rea-
soning. On the other hand, performance monitor-
ing methods must be used alongside evaluation of
other variable costs (emissions, energy trading, re-
pairs, etc.) for a full market evaluation.

References
[1] Performance Monitoring System, User Manual, Emerson
Process Management, revision VIII, 2005.
[2] B. Owsianka, Kompleksowy system bilansowania i nad-
zoru eksploatacji elektrowni i elektrociepłowni metoda
TKE (in Polish), Energetyka 3 (2002) 147.
[3] S. Munukutla, P. Sistla, A novel approach to real-time per-
formance monitoring of a coal-fired power plant, Electric
Utility Deregulation and Restructuring and Power Tech-
nologies. Proceedings. DRPT 2000. International Confer-
ence , 4-7 April 2000 273 – 277.

— 75 —
 Journal of Power Technologies 91 (2) (2011) 63–76

Unit C, Feedwater temp. [◦ C], Low load – left, High load – right

Unit D, Feedwater temp. [◦ C], Low load – left, High load – right
Figure 6: Results of statistical analysis of the principle performance parameters – main steam temperature – analysis from plan #2

— 76 —

View publication stats