Model Predictive Control and The Optimization of Power Plant Load While Considering Lifetime Consumption
Model Predictive Control and The Optimization of Power Plant Load While Considering Lifetime Consumption
1, FEBRUARY 2002
Abstract—This paper describes a decision support system that which is similar to unit commitment but on a single plant level,
indicates to a power plant operator the effect of daily operation is to schedule the plant over a short-term horizon (one to sev-
on plant lifetime consumption and maintenance costs and rec- eral days, divided into hourly or quarter hourly intervals). This
ommends a short-term operating strategy that optimizes plant
economic performance efficiency. The recommended operating could imply satisfying demand at a minimum cost (valid for
strategy is based on the optimization of an objective function plants with multiple generators/ turbines), or, as is more likely
that includes terms for revenues from energy sales, penalties, in a deregulated market, selecting the production level given
production costs and plant aging. Plant aging is based on models predicted energy prices (i.e., prices at which produced energy
that are directly load dependent and incorporate a memory can be sold). In either case, the requirements are the same: de-
aspect, a feature that is missing from common lifetime modeling
techniques. The optimization results in a tradeoff between max- cide when to turn the plant (and individual components1 ) on and
imization of immediate profits (i.e., earnings achieved by selling choose the production levels. A key component resulting from
heat and power) and minimization of lifetime consumption. this decision is the two tuple: cost per kW and kW production.
A model predictive control (MPC) and mixed logical dynamic This information is the core component used to formulate bids
(MLD) approach are used to solve the posed optimization problem for submission into a competitive energy market. An operational
and optimize this tradeoff and support the plant operator.
optimization tool can thus be seen as a key what-if and analysis
Index Terms—Crack growth models, lifetime modeling, mixed tool for power plants operating in a competitive market.
logical dynamic (MLD) approach, model predictive control (MPC), In addition to optimising the economic performance, the
power plant optimization.
operational optimization problem must take into account
requirements such as minimum up- and downtime of plant
I. INTRODUCTION components, startup and shutdown times, and other technical
constraints such as minimum or maximum power output and
D EREGULATED energy markets and tighter environment
and security standards are increasing the pressure on plant
owners to operate power plants in more efficient ways. In this
ramping constraints.
At the core of the operational optimization problem is an
environment, decision-makers would clearly welcome tools objective function that is subjected to several constraints. This
indicating plant operation strategies able to maximize plant functional can be expressed as a cost and thus should be mini-
revenues while still meeting constraints associated with safety, mized by means of the plant controls, resulting in optimal oper-
NOx emissions, plant degradation, etc. The approach described ation of the power plant. In this case the objective function has
in this paper seeks to address the plant aging requirement in the generic form
that the influence of a given plant operating mode on the aging
Sum over time of: production costs plus aging,
process of the power plant is included in the general operational
optimization problem. or degradation costs, less revenue costs
The estimation of consumed lifetime of power plant compo- OR
nents is a well-developed field [1]. Similarly, mathematical tools
for optimising power plant operation according to energy and (1)
fuel prices, power demand, plant fuel consumption character-
istics, and fluctuating energy sale prices are also available (see where
[2], [8], [9], and [14]). However, little emphasis has been placed time-optimization horizon (order of a
on combining these two ideas and we expect such a hybrid ap- few days);
proach to have a large economic impact on a power plant’s eco- power plant control, in the case of a
nomic performance. The objective of operational optimization, combined cycle power plant as con-
sidered here, the plant power and heat
Manuscript received December 15, 2000; revised July 24, 2001.
output;
E. Gallestey and A. Stothert are with ABB Corporate Research Ltd, Baden- aging rate cost of plant components;
Dattwil, Switzerland.
M. Antoine and S. Morton are with ABB Power Automation Ltd, Baden- 1Consider a combined cycle power plant with the possibility of using pro-
Dattwil, Switzerland. duced steam for either electricity generation (through a steam turbine) or district
Publisher Item Identifier S 0885-8950(02)01070-2. heating.
cost rates;
revenue rates, which can/are time de-
pendent as power is traded on the open
market and can fluctuate.
As the notation suggests, the vectors , , and depend upon
, the vector of plant controls. Thus, we must be able to predict
both plant behavior and aging during the interval [ ].
It follows that, in order for our approach to be successful, an
accurate description of the dependence is necessary.
In fact this paper is dedicated to the analysis of the relationship
and to the presentation of methods to solve the resulting
(1).
Related pioneering work, which introduces stress computa- Fig. 1. Steps for lifetime computation.
tions and ramping costs into generator dispatch algorithms, is
described in [10] and [11]. These references focus on managing
the effect of fatigue stress on steam turbine rotors. In [10] agation is modeled until a predefined critical crack size is at-
the intention is to compute the ramping process which gives tained. The growth of the crack is modeled using the machinery
the best economic performance under given conditions. The developed in fracture mechanics.
idea is very appealing, however, since no specific dynamic Typically, crack growth models are used during maintenance
lifetime models were developed, a somewhat unsatisfactory overhauls to assess the remaining lifetime of components with
workaround had to be applied to compute the solutions. On the detectable cracks. The results of such calculations are estima-
other hand, [11] use dynamic lifetime models in order to find tions of the number of operating hours left for the components.
the optimal ramping. Unfortunately, as the author mentions The subtle difference introduced in the MP approach is that no
in his comments, the important effect of creep on the aging measurable crack must exist and thus, the crack growth equa-
process was neglected, which to some extent limits the benefits tions are used as an abstract model giving a measure of how the
of the approach. In this paper these restrictions are dropped and component consumes its lifetime. In other words, the microc-
the lifetime models used are applicable for all components and rack propagation equations are merely a tool to indicate how
over the whole range of operation, of a modern power plant. the component ages when subjected to a particular load profile.
Moreover, our focus is on the optimal power output level (creep The process is illustrated in Fig. 1.
dependent!) and not only on computing the best shape for the Finally, note that the MP approach can be easily extended to
ramping process. include nonloading characteristics such as corrosion, erosion,
The paper is structured as follows. First, we give a brief and oxidation.
overview of component aging models and define the notion of
plant asset value. Then, the optimization approach is explained A. Crack Growth Models
and numerical results, based on simulations of a combined
Cracks within a component propagate as a result of stresses
cycle power plant, are presented. Finally, we summarize the
being applied to the component. Additionally, it has been noted
results and present some conclusions.
that crack growth rates are significantly different depending on
the nature of the applied stress. Indeed, if the stresses are con-
II. COMPONENT AGING MODEL stant then creep driven crack growth is observed. On the other
hand, if the stresses are cyclic (or periodic) fatigue crack growth
Lifetime assessment of power plant components requires a is observed [1]. These two aging mechanisms are examined
multidisciplinary approach involving knowledge of design, ma- more closely.
terial behavior and nondestructive evaluation techniques. The Fatigue Crack Growth Models: Fatigue crack growth models
literature available is vast and it is beyond the scope of this paper model how a cyclic stress causes a crack to propagate and give
to give a detailed account of these methods. However, for con- expressions for crack growth per cycle. Experimental observa-
venience of the reader an attempt will be made to provide an tion has indicated that the amplitude of the stress cycle
overview of the most relevant issues. and the maximum value of the cycle are the
For our purposes of lifetime or aging modeling, two factors driving forces. Indeed, if is too small, no crack growth oc-
are critical. First, the models must provide a direct relationship curs, while if is too large the component will break instan-
between plant load and plant aging and second, the models must taneously. This leads to a general form for fatigue crack growth
capture the operating history of the component. A survey of cur- as follows [1]:
rent literature [1], [5], indicates that fracture mechanics and in
particular the microcrack propagation (MP) approach [4], pro-
vide the required model features and gives a good theoretical
basis for the calculation of the lifetime of a component.
The main idea is as simple as it is appealing: a microcrack in a (2)
critical location or component is assumed to exist and its prop-
188 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 1, FEBRUARY 2002
where is the current cycle number and B. Aggregate Plant Lifetime Models
In order to obtain an indication of the aging of the whole plant
(3)
all individual components need to be considered. Since compo-
is a magnitude called stress intensity factor that depends upon nent aging is measured via a normalized value of the expected
the applied stress , the current crack size and a component ge- crack length to a critical crack length ( ), which is compo-
ometry dependent function . Further, is the di- nent dependent, ages of different components can be compared
mension into which the crack is propagating and and , directly. Different overall measures of plant age can be defined
and denote temperature dependent material param- using these normalized component ages, e.g., mean, or median.
eters. For example, if plant components (rotor, casing, pipes, blades,
The importance of the parameter must be stressed: it etc.) are being monitored, a sensible conservative measure is
defines the minimal cycle amplitude for a cycle to produce simply the minimum, i.e.,
damage. In applications its numerical value is such, that
fatigue becomes important only during start-ups and shutdown
procedures, i.e., when large temperature gradients are likely to An alternative and somewhat more economically intuitive ap-
appear. proach to measuring the age of a plant is to calculate the plant
Creep Crack Growth Models: As indicated previously, ap- asset value by the formula
plication of a steady load to a hot ( 350 C) cracked com-
ponent leads to crack propagation. Creep crack growth models
describe how this process occurs. AssetValue Value (6)
The general form of creep crack growth models is the empir-
ical law Here the number Value denotes the overall value (including
replacement costs) of the component , while the component
(4) lifetimes are given by (5).
The function AssetValue is easy to compute and provides a
where is referred to as the crack tip parameter and and weighted average of the current component value. Its drawback
are temperature dependent material parameters. The crack tip is that one component could exceed its critical age, while the
parameter consists of a small scale creep term, and a steady plant asset value would remain nonzero. This function should
state creep term, , i.e., thus be combined with a minimum type function as described
earlier.
The small-scale creep term is given in [1], while the
III. OPTIMIZING PLANT LIFETIME CONSUMPTION
steady state creep term is described by a term of the form
As mentioned in the introduction, our primary goal is the min-
imization of an objective function through appropriate choice
Here, denotes a geometry dependent function and and of plant controls—in this case the plant load level. Recall that
are the so-called Norton parameters. The remaining variables the objective function includes terms for, revenues, production
are defined as previously. costs and plant aging. Clearly, the new element of this contri-
Component Lifetime: The current crack size is bution with respect to earlier works on plant optimization is the
computed adding the contributions of both fatigue and creep. explicit handling of the aging terms. This problem is difficult
The lifetime of a component can thus be defined as follows: since it leads to an optimization with mixed discrete time (fa-
tigue) and continuous time (creep) evolutions, see (2) and (4).
(5) Relief comes from the fact that fatigue and creep can be
treated independently. Indeed, as a rule (see Section II), the
Notice that the lifetime has an initial value of one and de- discrete term (fatigue) becomes important only if severe
creases to zero. thermal shocks are present. Since this happens only during the
Lifetime Model Validation: At this point, the reader will start-up and shutdown procedures, it follows that the lifetime
have noted that these models have an empirical nature and thus, consumed during these procedures can be considered as an
that they contain many parameters and cross dependencies. additional cost associated with start-ups and shutdowns. On
However, the situation is not as hopeless as one might think the other hand, aging due to creep is present in all operating
at a first sight. Indeed, relatively straightforward procedures modes and, at least without further simplifications, can not be
can be used to adjust the parameters, so that material experi- treated in a standard static optimization framework.
mental data on rupture strength, creep and fatigue damage is We propose two methodologies to handle this problem. The
satisfied by the models. Furthermore, the computation of the first has a more classical structure and addresses the problem
dependence “component stress and temperature versus plant of how, in the context of lifetime consumption, to quantify
load”, although clearly difficult and laborious, can be achieved the impact of a given operation strategy on the economic
off-line by solving the corresponding (partial) differential performance of the plant. The second idea is novel and in our
equations. Due to their component specific character, details of opinion promising. It poses the optimization problem in terms
these “calibration” issues are omitted. of tracking a given plant lifetime trajectory. In both cases the
GALLESTEY et al.: MODEL PREDICTIVE CONTROL AND THE OPTIMIZATION 189
result is a tool for decision-makers that economically quantifies The plant value is directly related to its consumed lifetime.
the effect of plant operation on plant age. As indicated earlier, Thus, the ability to produce good estimates of plant lifetime
the result of the optimization is a cost per kW (including aging consumption offers new opportunities for plant operation. For
costs) and a kW production level, characteristics essential for example, the plant owner may want to operate his plant so that
actions such as bidding into a pool. its lifetime tracks a predefined trajectory. Other applications of
these ideas are possible, for example, it can be used to fulfil or
A. Approach I match given overhaul schedules.
In this subsection we present a natural idea, namely, that of In order to tackle this problem we propose the following pro-
minimizing a “cost minus revenue” functional, while satisfying cedure.
constraints related to plant dynamics and load satisfaction. Note 1) First, a nominal “desired” power plant lifetime
that these constraints are not essential for our exposition and
for the sake of brevity are omitted below. In other words, (1)
becomes is defined by taking into account the high-level plant
owner strategy.
2) Second, a nominal “optimal” load
such that
tradeoff between pure economic optimization and power plant [13] G. Anders, R. Entriken, and P. Nitu, “Risk assessment and financial man-
degradation. agement,” in IEEE Power Eng. Soc. Winter Meeting 1999 Tutorial.
[14] A. Wood and B. Wollenberg, Power Generation Operation and Con-
Further investigations are required, for example, to validate trol. New York: Wiley, 1996.
the lifetime models and/or to determine the influence of the
weights within the objective function presented in Section III-B.
However, from the example presented, the authors conclude that
Eduardo Gallestey was born in Havana, Cuba, in 1967. He received the M.Sc
the approach has the ability and the potential to predict the degree (honors) in mechanics and mathematics from the Moscow State Univer-
power plant aging process and its associated asset depreciation sity, Moscow, Russia, in 1991, and the Ph.D. degree in control theory from the
and to thus generate better loading strategies for daily plant op- Mathematics Department, University of Bremen, Bremen, Germany.
He was with the Engineering Department, Australian National University,
eration. This should lead to optimal long- and short-term power Canberra, Australia, and since February 2000, he has been with the ABB Re-
plant economic performance. search Center, Baden, Switzerland. His research interests include mathematical
modeling, model-based control, and robust (nonlinear) control techniques.
ACKNOWLEDGMENT
The authors would like to thank the reviewers for their helpful
Alec Stothert received the B.Sc. degree in engineering (cum laude), the M.Sc.
comments and for indicating the references [10] and [11]. degree in electrical engineering, and the Ph.D. degree from the University of the
Witwatersrand, Johannesburg, South Africa, in 1991, 1993, and 1996, respec-
REFERENCES tively.
After a postdoctoral position with the University of the Witwatersrand fo-
[1] R. Viswanathan, Damage Mechanisms and Life Assessment of High- cusing on modeling of competitive energy markets, he joined the Control and
Temperature Components. New York: ASM, 1989. Automation Group, ABB Corporate Research Center, Baden, Switzerland. His
[2] R. Baldick, “The generalized unit commitment problem,” IEEE Trans. interests include distributed control, model-based control, and optimization.
Power Syst., vol. 10, Feb. 1995. Dr. Stothert is a member of the IEEE Control Systems Society.
[3] E. Camacho and C. Bordons, Model Predictive Control. Berlin, Ger-
many: Springer-Verlag, 2000.
[4] M. P. Miller and D. C. McDowell et al., “A Life Prediction Model for
Thermomechanical Fatigue Based on Microcrack Propagation,” Amer.
Marc Antoine was born in Antwerp, Belgium, in 1959. He received the M.Sc
Soc. Testing Materials, Philadelphia, PA, 1993.
degree in mechanical engineering from the University of Brussels, Brussels,
[5] C. Cangan et al., “Recent developments in the thermomechanical fatigue
Belgium, in 1982.
lifetime prediction of superalloys,” e-JOM, vol. 51, no. 4, April 1999.
He then joined the Virginia Polytechnic Institute, Blacksburg, as a Research
[6] A. Bemporad and M. Morari, “Control of systems integrating log dy-
Assistant with the Department of Aerospace and Ocean Engineering. In 1985, he
namics and constraints,” Tech. Rep. AUT98-04, Automat. Contr. Lab.,
joined the Brown Boveri Control Systems Division, Baden, Switzerland. Since
Tech. Univ. Zürich, Zürich, Switzerland, vol. 35, 1999.
January 1999, he has been Product Manager with ABB Power Automation,
[7] F. Allgöwer and A. Zheng, “Nonlinear model predictive control,” in Ap-
Baden, Switzerland, where he is responsible for development of plant moni-
plications of Nonlinear Predicticve Control, F. Allgöwer and A. Zheng,
toring and optimization systems.
Eds. Basel, Switzerland: Birkhäuser, 2000, pt. III.
[8] F. Schweppe, M. Caramanis, R. Tabors, and R. Bohn, Spot Pricing of
Electricity. Dodrecht, The Netherlands: Kluwer, 1988, ch. 7.
[9] R. Green, “Competition in generation: The economic foundations,”
Proc. IEEE, vol. 88, pp. 128–139, Feb. 2000. Steve Morton was born in Sheffield, U.K., in 1963. He received the B.Sc degree
[10] C. Wang and S. Shahidehpour, “Optimal generation scheduling with (honors) in applied physics from Portsmouth University, Portsmouth, U.K., in
ramping costs,” IEEE Trans. Power Syst., vol. 10, pp. 60–67, Feb. 1995. 1987.
[11] J. Delson, “Thermal stress computation for steam-elecrtic generator dis- He is a Chartered Physicist, and since 1987, he has worked in the research
patch,” IEEE Trans. Power Syst., vol. 9, pp. 120–127, Feb. 1994. of thermal and electromagnetic weapon systems in the U.K. and Switzerland,
[12] M. Pereira, M. McCoy, and H. Merrill, “Managing risk in the new power and also worked on modeling of thermal power plants. He has been with ABB
business,” IEEE Computer Applicat. Power, vol. 13, pp. 18–24, Apr. Power Automation, Baden, Switzerland, since 1999, where his primary interest
2000. is in the area of power plant management and optimization.