http://dx.doi.org/10.5516/NET.04.2014.

720

APPLICATION OF MONITORING, DIAGNOSIS, AND

PROGNOSIS IN THERMAL PERFORMANCE ANALYSIS FOR
NUCLEAR POWER PLANTS
HYEONMIN KIM1, MAN GYUN NA2, and GYUNYOUNG HEO1*
1
Kyung Hee University, 1732, Deogyeong-daero, Giheung-gu, Yongin-si, Gyeonggi-do 446-701, Korea
2
Chosun University, 309, Pilmun-daero, Dong-gu, Gwangju 501-759, Korea
*
Corresponding author. E-mail : gheo@khu.ac.kr

Received October 30, 2014

As condition-based maintenance (CBM) has risen as a new trend, there has been an active movement to apply
information technology for effective implementation of CBM in power plants. This motivation is widespread in operations
and maintenance, including monitoring, diagnosis, prognosis, and decision-making on asset management. Thermal
efficiency analysis in nuclear power plants (NPPs) is a longstanding concern being updated with new methodologies
in an advanced IT environment. It is also a prominent way to differentiate competitiveness in terms of operations and
maintenance costs.
Although thermal performance tests implemented using industrial codes and standards can provide officially trustworthy
results, they are essentially resource-consuming and maybe even a hind-sighted technique rather than a foresighted one,
considering their periodicity. Therefore, if more accurate performance monitoring can be achieved using advanced data
analysis techniques, we can expect more optimized operations and maintenance. This paper proposes a framework and
describes associated methodologies for in-situ thermal performance analysis, which differs from conventional performance
monitoring. The methodologies are effective for monitoring, diagnosis, and prognosis in pursuit of CBM. Our enabling
techniques cover the intelligent removal of random and systematic errors, deviation detection between a best condition
and a currently measured condition, degradation diagnosis using a structured knowledge base, and prognosis for decision-
making about maintenance tasks. We also discuss how our new methods can be incorporated with existing performance
tests. We provide guidance and directions for developers and end-users interested in in-situ thermal performance
management, particularly in NPPs with large steam turbines.
KEYWORDS : Thermal Efficiency, In-situ Analysis, Condition-Based Maintenance, Turbine Cycle, Nuclear Power Plant

1. INTRODUCTION performance degradation. A plant computer system

(PCS) uses an algorithm to calculate the efficiency of a
Integrated information technology (IT) solutions for heat supply side and a power conversion side, along with
power plants are being installed not only for administra- performance indices associated with major components
tive purposes but also for plant operations and mainte- such as the reactor, steam generators, turbines, heat ex-
nance decision-making. Such a technical trend is not new changers, pumps, and so on. The traditional algorithms
or novel. Most power plants already use operator-aid embedded in PCSs calculate the performance metrics
solutions to facilitate operations and maintenance. How- with measured signals and compare them with refer-
ever, the current trend clearly differs from conventional ence values. Performance metrics thus identify increas-
systems by (1) increasing data availability via massive ing or decreasing conditions from the reference values
but low-priced databases, (2) providing convenient data in periodic performance monitoring during day-to-day
accessibility with advanced wired or wireless networks, operation. An abnormal condition observed during per-
and (3) promoting a reappraisal of the value of operator- formance monitoring or maintenance effectiveness, par-
aid solutions given the new big-data environment. ticularly before and after overhauls, needs to be checked
Given that the primary purpose of power plants is to using a special performance test conducted under well-
produce electricity at a competitive price, the most typi- ordered conditions with authorized codes and standards.
cal operator-aid solutions in nuclear power plants (NPPs) The results of performance monitoring and testing initi-
is software to monitor thermal efficiency and diagnose ate an engineering process to clarify the cause of deg-

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 737

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

radation and find appropriate maintenance activities. In The concept of in-situ thermal performance analysis
conventional thermal efficiency management, periodic we are suggesting falls between performance monitor-
monitoring has focused on the preliminary detection of ing and testing in terms of its purposes, measurements,
anomalies, and special tests have essentially determined and general procedures. The technical convenience and
whether such anomalies are problematic or not. The simplicity of conventional performance monitoring are
procedures for both performance monitoring and test- inherited in the in-situ thermal performance analysis. The
ing are provided in performance test codes (PTCs) from features of the proposed concept that distinguish it from
the American Society of Mechanical Engineers (ASME). conventional monitoring are: (1) it covers monitoring, di-
Although several PTCs are in accordance with users’ agnosis, prognosis, and decision-making for asset man-
purposes, Volumes PM, 6, and 6S are representative for agement, (2) it reduces measurement uncertainties using
evaluating overall electricity generation. PTC PM [1] physical or empirical data validation models, and (3) it
addresses performance monitoring, and PTC 6 or PTS uses a plant simulation model in the entire analysis.
6S [2, 3] are used for performance testing of NPPs with The ultimate goal of our in-situ thermal performance
large steam turbines. Table 1 shows their representative analysis is to fully support condition-based maintenance
features. (CBM). CBM, sometimes called predictive maintenance,
As Table 1 shows, performance monitoring and test- can save money by reducing time-consuming and unnec-
ing are complementary. The results analyzed by PTC 6 essary maintenance or testing activities and can also de-
or 6S are reliable and officially accepted, but the perfor- crease human errors, which is vitally important for safe-
mance test is a resource-consuming task. Conducting a ty-critical systems in NPPs. The proposed methods can
performance test for a large steam turbine cycle requires facilitate the CBM process for heat conversion cycles be-
meeting tough standards for sensor calibration, flow path cause the periodicity is much shorter than the progress of
isolation, and the stability of operational parameters [2, the usual degradation mechanism. Technical limitations
3], which is not easy in most NPPs connected to the na- in the signals coming from the sensors installed in a plant
tional grid. Despite such difficulties, performance tests and the analysis under variable plant configurations are
are usually carried out before and after a scheduled over- improved using advanced algorithms that use physical
haul to check the effectiveness of maintenance. While and empirical techniques. Because all the processes of
performance tests are necessary; however, strengthened in-situ thermal performance analysis should be managed
performance monitoring can reduce the frequency of per- on-line in real-time without an operator’s intervention, a
formance tests. high level of automation is required.

Table 1. Comparison between PTC PM and PTC 6 or 6S

PTC PM PTC 6, 6S

Quantitative identification of degradation as part of

Objective Continuous detection of performance degradation
performance monitoring

Use Trend monitoring during day-to-day operation Significant decision-making such as acceptance tests

Use of field instruments

Measurement Use of high-quality instruments
Possibly few sensors and low quality

Uncertainty of Results Not specified Specified upon satisfaction of procedures

Initialization (cycle stability, isolation),

Data acquisition, Data acquisition
Procedures Calculation of significant performance metrics, Calculation of heat balance (for PTC 6), performance
Comparison with reference metrics metrics, and corrections
Comparison with reference metrics

738 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

This paper explains the technological transition monitoring is less accurate than performance testing. In
in thermal performance management, specifically the other words, the detection of performance variation is
conceptual framework for in-situ thermal performance important in performance monitoring, whereas the abso-
analysis highlighting the contribution of CBM. We also lute values of performance indices are significant in per-
comprehensively explain several enabling technologies formance testing.
needed to implement the suggested framework. Even The in-situ thermal performance analysis we describe
though we focus on large scale steam turbine cycles in aims to achieve the accuracy and coverage of perfor-
NPPs, we expect our ideas to be shared with other power mance testing with the periodicity of performance moni-
cycles. toring. Therefore, the conventional codes and standards
cannot be directly applied to in-situ thermal performance
analysis. Nonetheless, those conventional codes and
2. FRAMEWORK FOR IN-SITU THERMAL PERFOR standards remain the technical basis for in-situ analysis.
MANCE ANALYSIS Aging vs. Degradation
The meaning of aging and degradation, which both
The framework for thermal performance management represent the decrease of an intended functionality, will
in NPPs has already been established by industrial codes sometimes be shared and sometimes distinguished in this
and standards such as ASME PTCs, Heat Exchange Insti- paper. Aging is defined as the natural decrease of thermal
tute standards, and so on. These codes were developed to performance over time. Most components go through
provide guidelines for performance monitoring and test- aging due to various causes such as mechanical, electri-
ing. Performance tests are a series of calculations used to cal, and chemical stressors. Aging is, therefore, both in-
determine heat balance and find the performance indices evitable and irreversible. On the other hand, degradation
at cycle or component levels. To guarantee their accuracy, is defined as a performance decrease caused by factors
tests must be conducted under strictly controlled condi- that can be removed or corrected to restore the degraded
tions with qualified sensors. Depending on the purpose component to its original functionality. Aging caused by
of the tests or upon a user’s request, a performance test degradation is not recovered, so a kind of hysteresis can
can be followed by a full-scope procedure [2] or a sim- occur in the functionality.
plified procedure [3]. On the other hand, the procedure Controllable vs. Uncontrollable Loss
for performance monitoring [1] focuses on surveillance Regardless of theoretical or practical controllability,
of trends in a few significant performance indices and is controllable loss is defined as loss due to degradation,
much less stringent than performance testing. The present and uncontrollable loss is caused by aging, which re-
codes and standards are not relevant to in-situ thermal quires replacement or cleaning of a component to return
performance analysis aiming to effectively support CBM. it to a normal condition. The sources of controllable loss
This chapter describes the in-situ thermal performance are incorrect settings of operating parameters categorized
analysis and how it compensates for the weaknesses of as (1) less than optimized operating conditions, (2) unde-
conventional performance monitoring and testing. sirable site conditions, and (3) a faulty flowpath causing
an internal or external leak. For example, a controllable
2.1 Terminology loss could be derated due to a poor condenser vacuum.
This section defines terms related to in-situ thermal Usually when a condenser vacuum deviates from its de-
performance analysis to clarify their technical distinc- sign condition, plant efficiency decreases. In this case,
tiveness. Most of these terms are common, but we refine operators can use additional vacuum pumps or coolant
their definitions here to distinguish our use of them from circulation pumps to regulate the condenser vacuum. If
their conventional uses. operators fail to control the pumps, the controllable loss
Performance Testing vs. Performance Monitoring falls under less than optimized operating parameters. If
Performance monitoring aims to detect variations in such degradation is caused purely by the temperature of
indices for cycles or components by periodic calculations, the ultimate heat sink, it could be considered an undesir-
and performance tests are conducted in highly controlled able site condition. The poor condenser vacuum might
conditions to calculate accurate and quantitative per- also be caused by abnormal venting or emergency drains,
formance indices. The codes and standards for periodic which would make it a flowpath fault.
surveillance are limited to calculating significant per- Achievable vs. Actual Performance
formance indices and monitoring their trends compared Fig. 1 shows a schematic diagram of plant efficiency
with reference indices. under various operating conditions [4]. Fig. 1 also ex-
The main benefit of performance monitoring is ex- plains plant efficiency in terms of the combination of un-
pected to be its short periodicity. Periodic checking controllable and controllable losses. The upper curve rep-
permits observation of malfunctions in a timely man- resents the achievable efficiency of a power plant without
ner because frequent performance testing is impractical aging or degradation. Usually, the design heat balances
because of its pre-requisites. Nonetheless, performance provided by a vendor or the heat balance analyzed from

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 739

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

Fig. 1. Comparison of Achievable and Actual Performance

an acceptance test is considered as an achievable effi- of controllable loss, less than optimized operating condi-
ciency. The lower curve represents the actual efficiency tions and flowpath faults are eliminated by test pre-requi-
of a plant with degradation and aging. A single point of sites. Undesirable site conditions are compensated using
the lower curve can be captured by performance testing correction curves. The corrected efficiency can now be
or monitoring. To move from a point on the actual effi- compared with the rated efficiency to observe whether
ciency curve to a point on the achievable efficiency curve the uncontrollable loss is acceptable. It should be noted
with the same operating condition, uncontrollable loss that corrected is valid only when the differences between
must be eliminated by replacing or maintaining compo- Conditionref and Conditionmea are within certain narrow
nents. Plant efficiency can also depend on the operating bands [4]. Beyond those limits, correction is not officially
parameters. guaranteed. Because plant conditions during daily opera-
The horizontal axis in Fig. 1 represents two operating tion are likely to violate those narrow limits, the compari-
conditions: a reference (or design) operating condition, son between rated and corrected is irrelevant for in-situ
Conditionref, and a measured operating condition that de- thermal performance analysis.
viates from the reference condition, Conditionmea. It is as- Even if the comparison between a rated and corrected
sumed that plant efficiency is maximized at the reference efficiency is recommended in a performance test, an-
condition. In other operating conditions, plant efficiency other approach is possible to detect uncontrollable loss.
will be changed. This kind of performance change is not Expected is defined as the point as indicating a plant’s
related to aging but to degradation, that is, controllable condition with controllable loss only. In other words,
loss. the deviation between an expected and a measured ef-
Corrected vs. Expected ficiency also presents only uncontrollable loss, which is
The four circles in Fig. 1 are significant in under- the same consequence as the conventional approach. The
standing the principles of conventional performance feasibility of this idea depends on how accurately we can
analysis and the proposed in-situ analysis. Rated is the simulate expected efficiency using component design pa-
best conceivable efficiency without any controllable or rameters and operational parameters. Because a flexible
uncontrollable losses. In practice, it is the guaranteed de- expected value can be produced as long as a simulation
sign efficiency provided by vendors or the efficiency of model is qualified and supported, this approach is prefer-
an acceptance test. Measured is the efficiency calculated able for in-situ analysis.
using a current plant condition. It includes controllable
and uncontrollable losses and so differs from the rated 2.2 Framework
point. We need a process to eliminate controllable loss The framework for the in-situ thermal performance
so we can identify the pure uncontrollable loss caused analysis follows a series of supervision loops suggested
by aging; this is corrected. The ASME PTCs for perfor- by Isermann [5]. They are composed of fault detection,
mance tests explain the conversion from measured effi- diagnosis, evaluation, and decision. Fig. 2 characterizes
ciency to corrected efficiency. Correction materials are the framework for in-situ thermal performance manage-
usually provided by a vendor. Among the three sources ment.

740 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

Fig. 2. Framework for In-situ Thermal Performance Analysis

2.2.1 Monitoring module timation, we use a professional toolbox to simulate a heat

The monitoring module calculates the measured and balance [6-12]. Even though the expected performance
expected performances and compares them to identify can be calculated by simulation toolboxes, it should be
the presence of uncontrollable loss. Uncontrollable loss noted that the accuracy of the expected performance de-
is detected when deviation between the measured and ex- pends strongly on engineers’ expertise in developing heat
pected performances becomes unacceptable. balance models. The heat balance model is also impor-
Implementing this module raises two technical issues. tant in the diagnosis and prognosis modules, which will
The first is how to accurately calculate an expected per- be explained in the next sections.
formance. The expected performance is determined by The second issue is how to accurately calculate a
calculating the heat balance for a clean and new power measured performance using uncalibrated sensors in-
system considering all types of controllable losses. Clean stalled in the field. Because the proposed framework uses
and new means all components are in the condition guar- sensors installed in a power plant, signal validation is re-
anteed by vendors or measured in an acceptance test. quired. Furthermore, because all the processes should be
Coping with controllable losses requires automatic data performed on-line, signal validation should be carried
acquisition of parameters corresponding to controllable out without an operator’s intervention. This step is called
losses. For instance, the following list belongs to the pa- data reconciliation in Fig. 2. The purpose of data recon-
rameters representing controllable losses in NPPs: steam ciliation is to detect signal anomalies and replace them
conditions at steam generator outlets; the ultimate heat with appropriate values, which is different from the man-
sink’s condition; makeup water condition; and flowpath agement of systematic and random errors suggested by
valve arrangement. In summary, the inputs needed to cal- ASME PTCs [13]. We define signal anomaly as a condi-
culate an expected performance are new and clean com- tion that is physically impossible such that a heat balance
ponents and the parameter values representing control- calculation fails or provides physically unreasonable
lable losses. The output is the expected heat balance and results if the raw data are used without any correction.
performance indices under those inputs. For credible es- Signal anomalies can be caused by systematic errors or

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 741

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

random errors. Pressure measurements at a low-pressure thermal efficiency management is usually knowledge-
turbine are likely to be anomalies caused by fluctuating based and represents the relationship between intrinsic
vacuum conditions, for example. Such anomalies result causes and superficial consequences. In other words, a
in meaningless performance indices such as greater than kind of supervised learning process is a pre-requisite.
100% turbine or pump efficiency, which should make The knowledge base can be accumulated by simulation
operators doubt the reliability of the whole solution. An- or historical records. However, no matter how we accu-
other specific issue in NPPs is the estimation of feedwater mulate historical knowledge, the knowledge base itself
flow rate [14-16]. The most significant parameter in cal- could be incomplete for unforeseen causes and conse-
culating turbine cycle performance is feedwater or con- quences. The closed loop of the Rankine cycle in NPPs
densate flow rate, and it is likely to be overestimated due makes it more difficult to find root causes because the
to fouling of the flow sensing element. Signal anomalies effect of the fault in a component propagates to other
must thus be distinguished from conditions caused by the components, so that any component could show abnor-
system’s aging or degradation. The role of data reconcili- mal performance indices [17, 18], which can easily lead
ation should therefore be minimal so it does not elimi- operators to a wrong decision. Furthermore, the simul-
nate the effects of aging or degradation. Limiting data taneous occurrence of multiple causes makes it difficult
reconciliation can be achieved using the periodic update to diagnose reliably [19]. The second challenge is the
scheme for plant models, which will be described in the possibility of aleatory uncertainty, in other words, signal
next chapter. accuracy. Even if signals are qualified through data rec-
After the expected and measured performance indices onciliation, their accuracy is still doubtful for diagnostic
are produced, the detection of uncontrollable loss is the purposes because data reconciliation does not estimate
final step of the monitoring module. Monitoring of the a true value but merely replaces damaged signals with
residual, that is, the deviation between the performance an approximate value. Therefore, we have to indicate the
indices of the measured and expected heat balances, can resolution limits of the installed sensors. The sensors for
provide symptoms associated with uncontrollable loss as thermal efficiency management provide only thermal-
promptly as possible. This step is called early warning hydraulic properties such as pressure, temperature, flow
in this framework. The early warning can bring various rate, level, and so on. These sensors might not observe
uncertainties because it is the final calculation merged by material-related degradation or be able to isolate a spe-
all the parameters’ uncertainty, so it can be misinterpret- cific root cause.
ed by operators. Therefore, statistical methods should be Several diagnosis methodologies, from an expert
used to cope with the uncertain factors for trustworthy system to advanced artificial intelligence, have been pro-
decision-making. posed, but the aforementioned limitations still hold for
industrial applications. The methodology proposed by
2.2.2 Diagnosis module ASME is a type of expert system such as cycle interrela-
The diagnosis module begins working when the mon- tions or a logic tree [1]. Cotton [20] suggested a diagnosis
itoring module warns that unexpected residuals in the matrix for large steam turbines. Both methodologies are
performance indices can be observed in the early warn- qualitative approaches unrelated to the turbine cycles in
ing annunciator. The residual represents change from a NPPs. Research and outcomes for performance diagnosis
reference condition due to uncontrollable loss. The pur- in process industries are summarized in Table 2. In ad-
pose of the diagnosis module is to address and determine dition to the methodologies in Table 21, auto-associative
the severity of the root cause of uncontrollable loss. kernel regression; auto-associative neural networks; an
In this case, two challenges emerge: One is from the auto-regressive multivariate state estimation technique;
viewpoint of epistemic uncertainty. The diagnosis for neural expert systems; neuro-fuzzy systems; and genetic

Table 2. Diagnosis Methodologies and Their Applications

Multivariate
Expert system Fuzzy PCA, ICA ANN SVM
Regression

[19] , [23], [24], [25],

[22] [27], [28] [18] [29], [30], [31], [32] [21], [33], [34]
[26]
(ANN: Artificial Neural Networks, ICA: Independent Component Analysis, PCA: Principal Component Analysis, SVM: Support
Vector Machine)

1
The numbers in Table 2 are reference numbers.

742 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

algorithms have also been suggested [21]. Most of the still estimates approximated states of uncontrollable loss
methodologies are classification tools, which means they by observing external environmental conditions. Further-
are unable to address unforeseen situations. Their accura- more, it is not easy to figure out the relationship between
cy and robustness differ depending on their pre-process- many stressors and uncontrollable loss. Effect-based
ing methods to reduce dimensions, transfer domains, or prognosis characterizes the lifetime of a specific unit or
eliminate noise. Some of them focus on the degradation system operating in its specific condition [35]. Weibull
mode of a specific component, and others try to find the analysis, the proportional hazards model, physics of fail-
component causing overall performance degradation. ure models, regression analysis, a Markov chain-based
Several authors have attempted to infer degradation model, LEAP-Frog, the particle filter model, the general
causes and their severity using quantitative and qualita- path model, and the shock model have all been suggested
tive methods [17-19, 22]. However, no single method has as prognosis methodologies [36]. However, no accepted
fully overcome the limitations and showed satisfactory standard guarantees the performance of any of them.
performance in field tests. From these experiences, we The purpose of the prognosis module is to analyze
decided that for the thermal analysis diagnosis module: the remaining useful lifetime below a certain probabil-
1) the knowledge base should be updated periodically to istic failure limit. As long as the failure limit is reason-
cover unforeseen faults. A clustering tool is therefore nec- ably converted to an allowable limit in terms of thermal
essary in order to distinguish a new case from the learned efficiency, the prognostics in reliability engineering can
cases; 2) multiple algorithms, including expert judgment be applied.
to determine root causes, should be provided in parallel;
3) system walkdown and inspection using advanced sen- 2.2.4 Decision-making module
sors should be used for more confident diagnoses. The purpose of the decision-making module is to
provide cost–benefit analysis for maintenance tasks by
2.2.3 Prognosis module integrating information from the diagnosis and progno-
The prognosis module forecasts the pattern of uncon- sis modules. Given the results of the diagnosis and prog-
trollable loss for future operating conditions. The pattern nosis modules, the choice may be to promptly stop for
is ultimately used to predict accumulated electric loss maintenance or to continue operation until a scheduled
during a certain interval to compare the cost–benefit for overhaul with or without a change in operating param-
specific maintenance tasks to improve asset management. eters to compensate for electric loss (Fig. 3). The diag-
Three categories of methodologies have been suggest- nosis module aims at providing significant causes for
ed in reliability engineering depending on the informa- aging or degradation and inducing relevant maintenance
tion available for degradation modeling: 1) failure-data tasks, which can affect the cost of the first choice. Mean-
based prognosis, 2) stress-based prognosis, and 3) effect- while, the prognosis module provides the expected risk
based prognosis [35]. Because the failure-data based ap- and proposes continuous operation if the risk is less than
proach uses only generic data from industrial experience, the benefit of maintenance. At this time, some operating
it cannot benefit from condition monitoring. The stress- parameters can be changed to alleviate electric loss.
based approach makes a better prognosis model, but it The options that can appear in the decision-making

Fig. 3. Diagnosis and Prognosis Modules for Decision-making

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 743

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

module are expressed in Fig. 4, where the vertical axis cases, it should be noted that the slope of Cases B and C
represents a loss of electric output. Case A, correspond- can have uncertainty.
ing to the first choice, stops the plant for immediate The decision-making module provides information
maintenance followed by a plant restart. When a plant for end users to determine the best decision by compar-
shuts down, the lost electricity that should be considered ing the expected costs, including the uncertainties of both
is the full capacity. However, uncertainty in calculating diagnosis and prognosis, as explained above. However,
the expected cost for maintenance remains because the the calculation presented in Fig. 4 was solely conducted
maintenance strategy determined by the diagnosis mod- from the viewpoint of power plants’ thermal efficiency.
ule might be incorrect. The effects of reliability problems that could occur in
Cases B and C correspond to the second choice. Case the materials when degraded performance is maintained
B depicts continued operation with changed controllable have been excluded from consideration. Therefore, mak-
loss, which can temporarily reduce lost electricity, but a ing the final conclusion after review with experts in other
close examination is required to analyze whether such areas is desirable.
change accelerates aging or degradation, such as, for in-
stance, structural reliability. Case C shows a pattern of
continuous operation without specific maintenance tasks. 3. ENABLING TECHNIQUES FOR ADVANCED
Cases B and C save the costs associated with mainte- THERMAL EFFICIENCY MANAGEMENT
nance because maintenance tasks are not executed, but
the saved costs must be less than the expected loss of Many methodologies for performance monitoring
electricity. To estimate the cost for power loss in both can be borrowed from performance testing. Therefore,
this chapter describes unique methodologies for imple-
menting the framework of on-line thermal efficiency
management discussed in Chapter II.

3.1 Data Reconciliation

Data reconciliation is a special branch of data valida-
tion that replaces in advance any data that could cause
noticeable problems in calculating the heat balance or
performance indices. The range check is an example of
data validation. In this method, a range in which certain
data are meaningful is determined in advance, and any
data that exceed that value are replaced by predetermined
values. Ranges can be determined according to physical
properties or operational characteristics. However, be-
cause the data are replaced by predetermined alternative
values regardless of the value of the surrounding varia-
bles or the operational state of the power plant, mass and
energy conservation might still be violated in calculat-
ing the heat balance. Data reconciliation considers that
Fig. 4. Comparison of Options in the Decision-making Module shortcoming.

Table 3. Comparison of Physical and Empirical Models

Physical Model Empirical Model

● White-box model ● Black-box model

●Calculation based on the first principles of physics, thermo- ●Prediction based on the previous dataset accumulated during
hydraulics, material science, etc., so more strict in terms of mass
sound operation
and energy conservation

● Possible to predict unforeseen states ● Impossible to be implemented before plant operation

● A model is hard to develop and limits system customization ● Models are easy to generate with high customization flexibility

744 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

Data reconciliation is performed before calculating guaranteed to accurately conserve mass and energy, they
a measured/expected heat balance using a data snapshot. are included as data reconciliation in this study because
Fig. 5 illustrates the information flow in the monitoring they flexibly reflect power plant operating conditions.
module. Once a set of raw signals is prepared, the ran- Whereas physical models use mass and energy conser-
dom errors in those signals are alleviated by averaging. vation equations and various first principles to create an
Systematic errors need to be handled differently. First, expected heat balance, empirical models use only the
the expected values are obtained from models that show dataset previously accumulated while a plant was oper-
the normal states of the power plant. Those models are ating without uncontrollable loss [37-39]. For example,
largely divided into physical models based on first prin- various linear/non-linear regression analyses, kernel re-
ciples and empirical models based on prior data from gressions, and artificial neural networks have been suc-
the power plant. To benefit from diverse algorithms and cessfully applied in various fields. If the final value made
compensate for the weakness of other techniques, we through these complicated stages shows a difference
suggest data fusion from two or more data reconciliation larger than the set point from the measured value, the
techniques. Fig. 5 explains how to merge the results from measured value will be replaced by the final value.
a physical model and an empirical model to generate final One physical model we used was a simulation model
datasets. Table 3 summarizes the pros and cons of both that can be also applied to calculate expected heat bal-
models [1]. ances. Data reconciliation based on physical models was
Although empirical models cannot be called data first proposed a long time ago [40]. When we applied a
reconciliation in the strict sense because they cannot be conventional technique to the Rankine cycles of power

Fig. 5. Information Flow in the Monitoring Module

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 745

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

plants, we found the following difficulties: (i) Huge ma- ters because they do not have fouling problems. However,
trix for a full scope plant model, (ii) Nonlinear Jacobian ultrasonic flow meters are difficult to install in existing
matrix caused by enthalpy calculation, and (iii) Lack of NPPs.
redundant sensors. These difficulties can be practically Fouling of the Venturi meter is the most significant
resolved by iterative simulations, as published and ap- contributor to the derating of the power level, up to 3% of
plied to power plants [41]. full power in some cases [42]. The most common way of
The results from physical and empirical models can resolving this problem is to inspect and clean the Venturi
be different. Those cases require an appropriate tech- meters every fuel cycle, but corrosion deposits near the
nique to determine a final dataset. Sometimes in practical orifice can reappear in as little as one month. Therefore,
field applications, it is also necessary to manually over- several on-line monitoring techniques have been present-
ride certain values, which can be part of data reconcili- ed to address the problem [14-16].
ation. In this paper, we introduce on-line monitoring tech-
The model that provides alternative values in data niques for the feedwater flow rate. There are generally
reconciliation is not the clean and new model mentioned two approaches to estimating the feedwater flow rate.
in Fig 1. Because data reconciliation must be meaningful One is to use an analytical and mechanistic model and
for models reflecting performance degradation to provide the other is to use data-based modeling that depends only
alternative values, it should be able to reflect the charac- on other measured values. Data-based modeling, such as
teristics of the past most similar to current power plant artificial intelligence, is preferred because it can model
conditions. Because alternative values eventually come complicated processes that are difficult to describe us-
from power plant conditions that are simulated by physi- ing analytical and mechanistic models. Therefore, data-
cal or empirical models, periodic updates of models are based models have been widely attempted, including the
essential to conform to the purpose of data reconciliation. following three.
Therefore, if data reconciliation replaces many variables The first method is a fuzzy inference system (FIS)
simultaneously, model reliability becomes uncertain, and [14]. The FIS is constructed from a collection of fuzzy
a process to update model should be started. if–then rules. The FIS combines linguistic and numerical
An important variable in managing NPPs’ thermal ef- information (mainly input–output data pairs). Linguistic
ficiency is main feedwater flow rates. Data reconciliation information can be directly incorporated, but the numeri-
methods for main feedwater flow rates have long been cal information must be incorporated by training the FIS
studied. Here, we consider the importance of measuring to match the target input–output data pairs. The FIS is
main feedwater flow rates as an example of data recon- trained using both the genetic algorithm and the least-
ciliation. squares method.
In most pressurized water reactors (PWRs), ther- Second, Na et al. [14] presented a support vector re-
mal reactor power is estimated using secondary system gression (SVR) to estimate the feedwater flow rate. Sup-
calorimetric calculations that rely on measurements of port vector machines (SVMs) have generally been ap-
the feedwater flow rate. But it is essential to actually plied to function classification problems. However, with
measure the feedwater flow rate in PWRs. Venturi flow the introduction of Vapnik’s -insensitive loss function,
meters measure the feedwater flow rate in most PWRs. SVMs have been extended and widely used to solve non-
However over time, corrosion product builds up near linear regression estimation problems. The SVR maps
the meter orifice and induces measurement drift. These the input data into a high dimensional feature space in
fouling phenomena increase the pressure drop across the which the linear regression is then carried out.
meter and cause an overestimation of the feedwater flow Third, a group method of data handling (GMDH)
rate. Whenever the calorimetric calculation is carried out model [16] was developed for soft-sensing the feedwater
during an operating fuel cycle, the thermal reactor power flow rate of a PWR. The GMDH algorithm can generally
must be reduced to match the false feedwater flow rate find interrelations in data to improve prediction accuracy
measured by the Venturi meter, which causes NPPs to and select the optimal structure for the model or network.
operate at lower-than-planned power levels. The GMDH algorithm uses a data structure similar to
According to the former 10CFR50 Appendix K for that of multiple regression models. The acquired data are
the emergency core cooling system (ECCS) evaluation usually divided into three subsets: a training data set, a
model, the original thermal power margin required to checking data set, and a test data set. The GMDH model
evaluate an ECCS was 2%, irrespective of the demon- cross-validates itself to prevent over-fitting and main-
strated instrument accuracy. A revision to 10CFR50 Ap- tains model regularization. The measured feedwater val-
pendix K was made to allow a margin equal to the actual ues are monitored using a sequential probability ratio test
instrument accuracy. The revision of the 10CFR50 Ap- (SPRT).
pendix K encourages the use of advanced feedwater flow Those three algorithms were confirmed using real
instruments in PWRs. Currently, ultrasonic flow meters plant startup data from the Hanbit Nuclear Power Plant
are considered a competitive alternative to Venturi me- Unit 3. The data consisted of signals measured from the

746 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

primary and secondary systems of the NPP, focusing on The data-based models have several sources of un-
the steam generator. Table 4 summarizes and compares certainty in their predicted values: the selection of the
the performance results of the three models [16]. The training data; model structure (including complexity);
GMDH model is not better than the FIS and SVR models and noise in the input and output variables. Statistical un-
for the training and verification data, but outperforms the certainty analysis generates many bootstrap samples for
other two models in the test data, which is most impor- training and checking data sets initially and retrains the
tant. model parameters on each bootstrap sample. After repeti-
In Fig. 6, the feedwater flow rate was assumed to be tive sampling and training, the values predicted by the
artificially degraded after the first 20 hr. The predicted data-based models provide a distribution. That distribu-
feedwater flow rate was similar to the actual feedwater tion is then used to calculate the prediction intervals.
flow rate. In addition, the figure shows the diagnostic re- To calculate the prediction intervals, 100 GMDH
sults of the existing Venturi meter using the SPRT [16]. models were developed from 100 bootstrap sample sets.
The SPRT detected sensor drift 12.7 hr after the sensor Fig. 7 shows the actual feedwater flow rate and the pre-
drift had started. It is possible to monitor the over-meas- diction intervals of the GMDH model. The prediction
urements of the feedwater flow rate during the entire op- intervals are shown better in the partly enlarged graph.
erating fuel cycle. The actual feedwater flow rate was covered by the up-

Fig. 6. Soft-sensing and Monitoring of Feedwater Flow rate Fig. 7. Actual Feedwater Flow rate and Prediction Intervals for
with Artificial Drift the Test Data

Table 4. Performance of the Data-based Models

Model Data type RMS error (%) Relative maximum error (%) Number of data points

Training data 0.1832 2.6978 1000

GMDH Verification data 0.1408 2.6978 1800
Test data 0. 0581 0.4098 201
Training data 0.1018 1.2092 1000
FIS Verification data 0.0838 1.2092 1800
Test data 0.3181 3.7025 201
Training data 0.2105 1.4278 1000
SVR Verification data 0.1896 1.4278 1800
Test data 0.2085 1.2497 201

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 747

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

per and lower bounds determined by the statistical un- 3.3 Diagnosis Tables
certainty method. At only one data point among 201 test Although diagnosis methods for thermal efficiency
data points did the actual feedwater flow rate exceed the management have been extensively proposed, each of
predicted interval, indicating 99.5% coverage. Therefore, them had application limitations in terms of accuracy,
we expect that the data-based model can be applied suc- credibility, or practicality. Even though ASME PTCs rec-
cessfully to validate and monitor existing feedwater flow ommend diagnosis procedures, plant engineers struggle
meters. to develop customized methods that can take their specif-
ic plant conditions into account. In Chapter II, we men-
3.2 Uncontrollable Loss Detection tioned the unavoidable limitations of our on-line diag-
Even though data reconciliation and uncontrollable nosis module. Although plant engineers want to be able
loss detection look completely different, they are exactly to diagnose unforeseen degradation modes, the current
the same in that both of them detect deviations between systems do not effectively cover even the malfunctions
a measured value and an expected (or reference) value. commonly known and experienced.
In uncontrollable loss detection, measured values (where In this paper, we call the knowledge-base containing
we are) correspond to performance indices obtained the causes and consequences of thermal efficiency degra-
through heat balance calculations using measured data dation a diagnosis table. The consequence is observed by
and coordinated data; expected values (where we should monitoring the key parameters in the diagnosis table. We
be) are performance indices calculated from clean and assume a root cause can be determined by observing the
new models. All the performance indices observed in this unique combination of changes (increasing, decreasing,
step correspond to the key parameters in the diagnosis or staying the same) in key parameters. As an example,
table that will be explained in the next section. Table 5 summarizes the changes of key parameters for
The process of uncontrollable loss detection focuses diagnosing aging in high pressure turbines in a Rank-
on time-series anomalies for an individual parameter. In ine cycle [18]. The maintenance required is determined
other words, considering the statistical uncertainty con- through this table by checking the area changes of each
tained in individual parameters, this process checks for turbine stage. As depicted, the area of the first stage of
any statistically significant deviation between the meas- the HP turbine might decrease when the throttle flow and
ured performance index and the expected performance first stage pressure decrease, which thus decreases the ef-
index. Well-known examples of this method are sequen- ficiency of the HP turbine. Thus, we can suspect deposit
tial tests [46] and statistical quality control (SQC) charts and peening at the first stage of the HP turbine.
[44]. SQC charts such as the Shewhart chart or Cumu- In the same way, it is possible to diagnose perfor-
lative-Sum chart can track various behaviors in a single mance degradation in heat exchangers. Table 6 and 7
parameter and expose the hidden characteristics of plant show the diagnosis tables for a condenser and a feedwa-
processes by pattern types. As explained in the next sec- ter heater [22].
tion, methodologies are necessary to identify increases or The diagnosis tables presented in Tables 5, 6, and 7
decreases in variables. SPRT and SQC can also be used are not in their complete forms because they will likely
for this purpose. vary a little depending on power plant structures, and they
Commercial products for uncontrollable loss detec- can be expanded to other instruments with similar meth-
tion are currently being sold [45-47]. In addition to study ods. In most cases, diagnosis tables can be analyzed us-
papers, those products can identify operation principles ing the professional simulation toolbox mentioned in sec-
or technical methodologies from multiple patents [48- tion II.1. The professional simulation toolbox provides
52]. the degrees of changes in key parameters according to

Table 5. Diagnosis Table for Steam Turbines

Key Parameter
Throttle Flow PT P1st … HP EFF.
Cause

Area at 1st Stage Increased (SPE) ↑ - ↑ ↓

︙
Area at 1 Stage Decreased
st
↓ - ↓ ↓
(Deposits / peening)
(PT: Pressure at Throttle Valve Inlet, P1st: First Stage Pressure, HP EFF.: Efficiency of High Pressure Turbine, SPE: Solid Particle Erosion)

748 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

Table 6. Diagnosis Table for Condenser

Key Parameter
TTD ∆T … Tsubcooling
Cause

Pin Hole ↑ - ↑

Fouling ↑ - -

(∆T: Tin – Tout, TTD: Terminal Temperature Difference, Tsubcooling: Difference between Saturated Temperature at Condenser Pressure
and Temperature on Hotwell side of Condenser)

Table 7. Diagnosis Table for Feedwater Heater

Key Parameter
TTD DCA ∆Ttube … PShell
Cause

High Drain Level ↑ ↓ ↓ ↑

Low Shell Pressure ↓ ↓ ↓ ↓

Tube Leakage ↓ ↑ ↑ ↑

(DCA: Drain Cooler Approach, ∆Ttube: Ttube_out – Ttube_in, Pshell: Shell side Pressure)

the severity of each cause. Therefore, if the degrees of is not always the case. Therefore, the effects of multiple
change in key parameters are known, the severity of each causes should be added to the diagnosis table as sepa-
cause can be predicted by inverse calculation. rate phenomena. The second issue is that ad hoc sensors
Fig. 8 shows the procedure for using the diagnosis are unlikely to be used because of the characteristics of
tables. The SPRT or SQC used for uncontrollable loss in-situ analysis. Therefore, in some cases, performance
detection enables judgment about the increases or de- degradation modes cannot be identified with sufficient
creases in key parameters presented in Tables 5, 6, and resolution using the signals collected from the limited
7. If these observations match each other, the causes of
degradation and actions to be taken can be identified. If
they do not, we continue monitoring the key parameters.
If the types of parameters registered in a diagnosis table
are not enough to determine a specific cause, we can add
more parameters to create another unique combination.
The parameters to be added could be measured signals or
calculated parameters.
Three issues might come up in using the diagnosis
tables. The first one is diagnoses when multiple causes
occur simultaneously. Although some performance deg-
radation factors are unlikely to be left unattended, such
as serious leakages, performance degradation factors that
progress slowly, such as tube fouling, are quite likely to
cause other performance degradation factors. Although
the effects of certain causes occurring simultaneously
can be predicted based on linear combinations, which Fig. 8. Flowchart of Diagnosis

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 749

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

number of sensors installed. In such cases, providing abnormal conditions were detected by uncontrollable
several suspected causes simultaneously using weighting loss detection, and the present point is determined as a
or credibility is expected to improve decision-making as point t=ti between t=tC and t=tM, which is the point at
the second-best plan. Finally, because the diagnosis table which the next overhaul begins. Here, because the linear
is based on known degradation phenomena, it is not clas- model has been regarded as a prognosis model, future
sification through supervised learning. Therefore, new key parameter values can be predicted using the key pa-
degradation phenomena cannot be accurately judged. In rameter values observed between t=t0 and tC.
this respect, clustering through unsupervised learning is At t=ti for a key parameter, k, k, i estimated from the
necessary. In the case of clustering, the fact that combi- linear regression is
nations of key parameters are neither normal, nor con-
sequences of any cause in the diagnosis table, should be
(1)
judged and made clear to the user. Although many studies
on clustering have been conducted, a few power plant ap-
plication can be used in thermal efficiency analysis [53]. where t is time,
 is an intercept,
3.4 Prognosis Using Linear Model  is the slope of the line.
We introduced the concept of risk in forecasting the The confidence interval given by the specified confidence
trend of uncontrollable loss by the next period. In engi- level is
neering, risk is defined as the multiplication of likelihood
and the consequence of a certain event. For thermal effi-
ciency management, consequence is regarded as the loss (2)
of electric output or decreased efficiency, and likelihood
represents the uncertainty level of a prognosis result.
The targets of prognosis can reasonably be shared
with the key parameters defined in the diagnosis tables. where s is sample variance,
The patterns of the key parameters, their increasing or n is sample size. This is the number of data points in
decreasing status, can be stretched to the next scheduled the period from t1 when degradation began to tn when
overhaul in the prognosis module to calculate the accu- degradation was detected,
mulated loss of electric output using expected heat bal- m is min or max. If m is min, choose ‘–’, and if m is
ances. max, choose ‘+’.
In this study, we used linear regression models to de- t(1-a) is the critical value of t-distribution at confidence
velop theories because they are the most intuitive to use. level (1-a).
Linear regression can also determine whether parameters Estimated values of electric losses can be simulated
have increased, decreased, or stayed the same. Fig. 9 using the estimated values of key parameters obtained
shows our overall prognosis scheme based on linear re- through the prognosis. Equations (3) and (4) provide the
gression, and the equations for simple linear regression electric losses expected to occur from the beginning of
are shown below. Prognosis is predicting the degree of degradation to the overhaul as a range considering the
performance degradation should the equipment be con- uncertainty of the prognosis.
tinuously operated until the next overhaul or thereaf-
ter without correcting performance degradation factors.
Therefore, in Fig 9, t=t0 is regarded as the point at which (3)

(4)

where Ploss is electric loss [MWd],

f is the day-averaged power loss under key param-
eters using heat balance simulation [MW],
d is unit time [day].
The outputs of the expected electric loss, Ploss have
the intervals ( ) considering the uncertainty of the
degradation estimation. This result will be provided for
Fig. 9. Overall Scheme of Prognosis Based on Linear cost–benefit analysis for final decision-making, as de-
Regression scribed in the next section.

750 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014

 KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

4. CONCLUSIONS AND RECOMMENDATIONS power plants. Nonetheless, if new methodologies are not
proceduralized, they cannot advance. Therefore, we em-
This paper has suggested a framework and enabling phasize that concrete administrative support is required
techniques for in-situ thermal efficiency management, until the methodologies settle down.
which is based on the industrial codes and standards of
conventional performance tests, but requires additional REFERENCES_______________________________
methods to implement as in-situ analysis which is an on- [ 1 ] ASME PTC PM-1993, “Performance Monitoring
line, automated, and intelligent procedure. Specifically, Guidelines for Steam Power Plants”, American Society of
we described, the detection of performance degradation Mechanical Engineers, 1993.
[ 2 ] ASME PTC 6-2004, “Performance Monitoring Guidelines
and techniques for diagnosis, prognosis, and decision- for Steam Turbine”, American Society of Mechanical
making modules helpful for operator policymaking. Engineers, 2004.
The framework and methods proposed here are al- [ 3 ] ASME PTC 6S-1988, “Procedures for Routine
ready being applied in many Korean NPPs and fossil- Performance Test of Steam Turbines”, American Society
fueled power plants. The algorithm of methodologies of Mechanical Engineers, 1988.
discussed is generically applicable, but some cases need [ 4 ] R. Gay, “Power Plant Performance Monitoring”, 1st ed.,
to be adjusted according to the characteristics of individ- R-Squared Publishing, 2004.
ual plants. Although diverse methods were mentioned in [ 5 ] R. Isermann, “Supervision, Fault-detection and Fault-
Diagnosis Methods - An Introduction,” Control
this paper, the most important factor for the accuracy of
Engineering Practice, Vol. 5, Issue 5, pp. 639–652, 1997.
efficiency management is including a sufficient number [ 6 ] SCIENTECH, “PEPSE”, Retrieved on July 24, 2014,
of high-quality sensors. Many enabling techniques are http://scientech.cwfc.com/software /spokes /06_PEPSE.
eventually required to enhance the quality of the data htm
collected from sensors. Therefore, to effectively manage [ 7 ] GE Measurement & Control, “GateCycle Software”,
thermal efficiency with limited resources, enhancing the Retrieved on July 24, 2014, http://www.ge-mcs.com/en/
quality of the sensors and installing redundant sensors at bently-nevada-software/320-performance/1831-siweb-
important positions is recommended first. Fundamental pl655.html
technologies are the next most important factor, includ- [ 8 ] GP strategies, “Virtual Plant”, Retrieved on July 24, 2014,
http://www.etapro.com /platform /virtualPlant.aspx
ing models that can calculate the heat balance of power [ 9 ] Thermoflow, INC., “Thermoflow” Retrieved on July 24,
plants according to operating conditions. Although tur- 2014, http://www.thermoflow.com /index.html
bine cycle vendors currently provide heat balance dia- [ 10 ] ProSim Software & Services in Process Simulation,
grams for major loads on utilities, there are difficulties “ProSim”, Retrieved on July 24, 2014, http://www.prosim.
in managing utilities’ thermal efficiency because the heat net/fr/index.php
balance simulation models used to prepare the heat bal- [ 11 ] Engineered Software, INC., “CONVAL”, Retrieved on
ance diagrams are not disclosed. Although heat balance July 24, 2014, http://eng-software. com/ products/conval/
models per se should be manufacturers’ intellectual prop- [ 12 ] Sim Tech, “IPSE pro”, Retrieved on July 24, 2014, http://
www.ipsepro.com /IPSEpro /english /IPSEpro.php
erty, if the models could be provided through the tool-
[ 13 ] ASME PTC 19.1-2005, Performance Monitoring
boxes, it would be very helpful in effectively managing Guidelines for Test Uncertainty, American Society of
utilities’ thermal efficiency while protecting vendors’ in- Mechanical Engineers, 2005.
tellectual property rights. [ 14 ] Man Gyun Na, Sun Ho Shin, and Dong Won Jung, “Design
Dissemination of CBM stems from the increase of of a Software Sensor for Feedwater Flow Measurement
available data caused by IT development. Because mass Using a Fuzzy Inference System,” Nuclear Technology,
storage devices were not installed until a few years ago, Vol. 150, No. 3, pp. 293-302, June 2005.
operator-aid solutions did not spread. But blind adop- [ 15 ] Man Gyun Na, In Joon Hwang, and Yoon Joon Lee,
tion of such solutions can waste money and have other “Inferential Sensing and Monitoring for Feedwater
Flowrate in Pressurized Water Reactors,” IEEE Trans.
negative effects. The most important goal is not to find Nucl. Sci., Vol. 53, No. 4, pp. 2335-2342, Aug. 2006
a suitable efficiency management solution but to foster [ 16 ] Dong Hyuk Lim, Sung Han Lee, and Man Gyun Na,
thermal efficiency experts to distinguish pros and cons in “Smart Soft-Sensing for the Feedwater Flowrate at PWRs
developing thermal management solutions for their par- Using a GMDH Algorithm,” IEEE Trans. Nucl. Sci., Vol.
ticular plants. 57, No. 1, pp. 340-347, Feb. 2010.
Finally, this paper recommends preparing administra- [ 17 ] G. Heo, S. Chang, “Algebraic Approach for the Diagnosis
tive backup so in-situ efficiency management can settle of Turbine Cycles in Nuclear Power Plants,” Nuclear
down into daily operation. The concept proposed in this Engineering and Design, Vol. 235, Issue 14, pp. 1457-
1467, June 2005.
paper has not yet been officially recognized internation- [ 18 ] C. Jee, S. Jang, I. Lee, G. Heo, “Thermal Efficiency
ally, and its accuracy can vary greatly according to the Degradation Diagnosis Method Using Regression Model,”
proficiency of its development. Furthermore, as men- ICI 2011(ISOFIC, CSEPC, ISSNP 2011), Daejeon, Korea,
tioned earlier, sensor quality is a fundamental limitation August 21-25, 2011.
of in-situ analysis. However, those problems should de- [ 19 ] H. Kim, Y.K. Kang, S.Y. Song, G. Heo, “Diagnosis of
crease as in-situ analysis becomes more widely used in Feedwater Heater Performance Degradation using Fuzzy

NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014 751

KIM et al., Application of Monitoring, Diagnosis, and Prognosis in Thermal Performance Analysis for Nuclear Power Plants

Approach”, Korean Nuclear Society, 2014 Applications”, U.S. Department of Energy, PNNL-21515,
[ 20 ] K. C. Cotton, “Evaluating and Improving Steam Turbine July 2012.
Performance,” 2nd ed., Cotton Fact Inc., 1998, pp.332. [ 37 ] G. Heo, “Condition Monitoring using Empirical Models:
[ 21 ] J. Ma, J. Jiang, “Applications of fault detection and Technical Review and Prospects for Nuclear Applications,”
diagnosis methods in nuclear power plants: A review” Nuclear Engineering and Technology, Vol. 40, No. 1, pp.
Progress in Nuclear Energy, Vol. 53, pp. 255-266, 2011 49-68, Feb. 2008.
[ 22 ] S. Kim and G. Heo, “Strategy of Risk-Informed Inspection [ 38 ] S. An, G. Heo, and S. Chang, “Detection of Process
for Secondary Systems in NPPs”, Korean Nuclear Society, Anomalies using an Improved Statistics Learning
Vol. 1, pp. 991, 2012 Framework,” Expert Systems with Applications, Vol. 38,
[ 23 ] H. Eliasi, H. Davilu, M.B. Menhaj, “Adaptive fuzzy model Issue 3, pp. 1356-1363, 2011.
based predictive control of nuclear steam gnerators”, [ 39 ] R. Ahmed, G. Heo, and S. An, “Intelligent Condition-Based
Nuclear Engineering and Design, Vol. 237, pp. 668-676, Maintenance Using Plant Health Index,”, 8th International
2007. Topical Meeting on Nuclear Plant Instrumentation, Control
[ 24 ] K. Salahshoor, M. Kordestani, M.S. Khoshro, “Fault and Human Machine Interface Technologies, San Diego,
detection of an industrial steam turbine using fusion California, USA, July 22-26, 2012.
of SVM(support vector machine) and ANFIS(adaptive [ 40 ] R. Mah, G. Stanley, and D. Downing, “Reconciliation and
neuro-fuzzy inference system) classifiers”, Energy, Vol. Rectification of Process Flow and Inventory Data,” Ind.
35, pp. 5472-5482, 2010. & EC Proc. Des. Dev., Vol. 15, Issue 1, pp. 175-183, Jan.
[ 25 ] W. Yang, K.Y. Lee, S.T. Junker, “Fuzzy Fault Diagnosis 1976.
and Accommodation System for Hybrid Fuel-Cell/Gas- [ 41 ] S. Park and G. Heo, “Simulation Based Data Reconciliation
Turbine Power Plant”, IEEE transactions on energy for Monitoring Power Plant Efficiency,” Transactions of
conversion, Vol. 25, No. 4, 2010. the Korean Nuclear Society, Jeju, Korea, Oct 21-22, 2010.
[ 26 ] L. Su and Z. Zhao, “Performance Diagnosis of Power [ 42 ] A.M.C. Chan and A.K. Ahluwalia, “Feedwater Flow
Plant Boiler Based on Fuzzy Comprehensive Evaluation”, Measurement in U.S. Nuclear Power Generation Stations”,
Electrical and Control Engineering, pp. 3076-3079, 2011. EPRI TR-101388, Electric Power Research Institute, Nov.
[ 27 ] M. Daneshvar and B.F. Rad, “Data Driven Approach for 1992.
Fault Detection and Diagnosis of Boiler System in Coal [ 43 ] A. Wald and J. Wolfowitz, “Optimum Character of
Fired Power Plant using Principal Component Analysis”, the Sequential Probability Ratio Test”, The Annals of
International Review of Automatic Control, Vol. 3, No. 2, Mathematical Statistics, Vol. 19, No. 3, pp. 326-339, 1948
pp. 198-208, 2010 [ 44 ] M. Camargo, W. Filho, A. Dullius, J. Maciel, M. Pinto,
[ 28 ] A. Ajami and M. Daneshvar, “Data driven approach for “Statistical Quality Control: A Case Study Research,” in
fault detection and diagnosis of turbine in thermal power Proc. of IEEE International Conference on Management
plant using Independent Component Analysis (ICA)”, of Innovation and Technology, Bangkok, Thailand, Sept.
Electrical Power and Energy Systems, Vol. 43, pp. 728- 21-24, 2008, pp.746-750.
735, 2012. [ 45 ] GE Intelligent Platforms, “Smart Signal”, Retrieved on
[ 29 ] R.B. Joly, S.O.T. Ogaji, R. Singh, S.D. Probert, “Gas- Sep 10, 2014, http://www.ge-ip.com/products/proficy-
turbine diagnostics using artificial neural-networks for smartsignal/p3704
a high bypass ratio military turbofan engine”, Applied [ 46 ] Instep, “PRISM”, Retrieved on Sep 10, 2014, http://www.
Energy, Vol. 78, pp. 397-418, 2004. instepsoftware.com/instep-software-products/prism-
[ 30 ] M. Fast, M. Assadi, S. De, “Development and multi-utility predictive-asset-monitoring
of an ANN model for an industrial gas turbine”, Applied [ 47 ] BNF, “Plant Health Index”, Retrieved on Sep 10, 2014,
Energy, Vol. 86, pp. 9-17, 2009. http://www.bnftech.com/eng/product/product03.php
[ 31 ] P. Deshpande, N. Warke, P. Khandare, V. Deshpande, [ 48 ] D.D. Benedetto, R. Berry, J. Gargus, G. Millinger,
“Thermal Power Plant Analysis Using Artificial Neural “Methods and systems for operating an automated system
Network”, NUiCONE, 2012. using a process definition model”, U.S. Patent, 7,822,492,
[ 32 ] S. Simani and C. Fantuzzi, “Fault Diagnosis in Power Oct 19, 2007
Plant using Neural Networks”, Information Sciences, Vol. [ 49 ] D.H. Miller, F. Mercer, J. Popelas, “Service and diagnostic
127, pp. 125-136, 2000. logic scan apparatus and method”, U.S. Patent, 7,836,347,
[ 33 ] K. Salahshoor, M. Kordestani, M. S. Khoshoro, “Fault Oct 17, 2007
detection and diagnosis of an industrial steam turbine [ 50 ] Y.H. Xu, M.A. Lobuono, R.A. Rucinski, J. Simons,
using fusion of SVM (support vector machine) and ANFIS “Methods and systems for operating a sequence of event
(adaptive neuro-fuzzy inference system) classifiers”, recorder”, U.S. Patent, 8,489,554, July 16, 2013
Energy, Vol. 35, pp. 5472-5482, 2010 [ 51 ] H.J. Seo, “Method Measuring of Healthy Indicia of Plant
[ 34 ] K.Y. Chen, L.S. Chen, M.C. Chen, C.L. Lee, “Using SVM Reflected Status of Sub-component and Storage Media
based method for equipment fault detection in a thermal Thereof”, Korea Patent, 1014337240000, Aug 19, 2014
power plant”, Computer in Industry, Vol. 62, pp. 42-50, [ 52 ] S.Y. Kim “Data Collecting Method for Detection and On-
2011 time warning System of Industrial Process”, Korea Patent,
[ 35 ] J. Hines and A. Usynin, “Current Computational Trends 1010967930000, Dec 14, 2011
in Equipment Prognostics,” International Journal of [ 53 ] M.A. Jianping, J. Jin, “Semi-Supervised Classification for
Computational Intelligence Systems, Vol. 1, No. 1, pp. Fault Diagnosis in Nuclear Power Plants”, International
94–102, Jan. 2008. Symposium on Future I&C for Nuclear Power Plants/
[ 36 ] J.B. Coble, P. Ramuhalli, L.J. Bond, J.W. Hines, B.R. International Symposium on Symbiotic Nuclear Power
Upadhyaya, “ Prognostics and Health Management in Systems 2014 (ISOFIC/ISSNP 2014), pp.24-25, Aug 24-
Nuclear Power Plants: A Review of Technologies and 28, 2014

752 NUCLEAR ENGINEERING AND TECHNOLOGY, VOL.46 NO.6 DECEMBER 2014