You are on page 1of 11

Computers and Chemical Engineering 33 (2009) 15571567

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage:

Adaptation strategies for real-time optimization

B. Chachuat a, , B. Srinivasan b , D. Bonvin c
Department of Chemical Engineering, McMaster University, 1280 Main Street West, Hamilton, ON L8S 4L7, Canada
Dpartement de Gnie Chimique, cole Polytechnique de Montral, C.P. 6079 Succ. centre ville, Montral, QC H3C 3A7, Canada
Laboratoire dAutomatique, cole Polytechnique Fdrale de Lausanne (EPFL), Station 9, CH-1015 Lausanne, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history: Challenges in real-time process optimization mainly arise from the inability to build and adapt accurate
Received 13 October 2008 models for complex physico-chemical processes. This paper surveys different ways of using measure-
Received in revised form 23 February 2009 ments to compensate for model uncertainty in the context of process optimization. Three approaches
Accepted 28 April 2009
can be distinguished according to the quantities that are adapted: model-parameter adaptation updates
Available online 9 May 2009
the parameters of the process model and repeats the optimization, modier adaptation modies the
constraints and gradients of the optimization problem and repeats the optimization, while direct input
adaptation turns the optimization problem into a feedback control problem and implements optimality
Measurement-based optimization
Real-time optimization
via tracking of appropriate controlled variables. This paper argues in favor of modier adaptation, since it
Batch-to-batch optimization uses a model parameterization and an update criterion that are well tailored to meeting the KKT condi-
Plantmodel mismatch tions of optimality. These considerations are illustrated with the real-time optimization of a semi-batch
Model adaptation reactor system.
Model parameterization 2009 Elsevier Ltd. All rights reserved.

1. Introduction Marquardt, 2003). When measurements are available, adaptive opti-

mization, also called real-time optimization (RTO), can help adjust to
Optimization of process performance has received attention process changes and disturbances, thereby reducing conservatism
recently because, in the face of growing competition, it represents (Francois, Srinivasan, & Bonvin, 2005). It is interesting to note that
the natural choice for reducing production costs, improving prod- the above classication is similar to that found in control problems
uct quality, and meeting safety requirements and environmental with the robust and adaptive techniques.
regulations. Process optimization is typically based on a process An optimal solution has to be feasible and, of course, opti-
model that is used by a numerical procedure for computing the mal. In practice, feasibility is typically given a higher priority
optimal solution. In practical situations, however, an accurate than optimality. In the presence of model uncertainty, feasibil-
process model can rarely be found with affordable effort. Uncer- ity is usually enforced by the introduction of back-offs from the
tainty results primarily from trying to t a model of limited constraints. The availability of measurements helps reduce these
complexity to a complex process system (Bonvin, 1998; Forbes back-offs and thus improves performance (Chachuat, Marchetti,
& Marlin, 1994). The model-tting task is further complicated by & Bonvin, 2008). Generally, it is easier to measure or infer con-
the fact that process data are usually noisy and signals often do strained quantities (e.g., temperature or pressure) than estimate
not carry sufcient information for efcient process identication gradients of the cost and constrained quantities. These elements
(Yip & Marlin, 2004). Therefore, optimization using an inaccurate clearly set a priority of actions in the framework of adaptive opti-
model might result in suboptimal operation or, worse, infeasible mization.
operation when constraints are present. This paper discusses three adaptive optimization approaches
Two main classes of optimization methods are available for han- that differ in the way adaptation is performed, namely (i) model-
dling uncertainty. The main difference relates to whether or not parameter adaptation, where the measurements are used to rene
measurements are used in the calculation of the optimal strategy. the process model, and the updated model is used subsequently
In the absence of measurements, robust optimization is typically for optimization (Chen & Joseph, 1987; Marlin & Hrymak, 1997);
used, whereby conservatism is introduced to guarantee feasibil- (ii) modier adaptation, where modier terms are added to the cost
ity for the entire range of expected variations (Mnnigmann & and constraints of the optimization problem, and measurements
are used to update these terms (Forbes & Marlin, 1994; Gao & Engell,
2005; Roberts, 1979; Tatjewski, 2002); and (iii) direct input adap-
Corresponding author. Tel.: +1 905 525 9140x24703; fax: +1 905 521 1350. tation, where the inputs are adjusted via feedback control, hence
E-mail address: (B. Chachuat). not requiring on-line numerical optimization but off-line controller

0098-1354/$ see front matter 2009 Elsevier Ltd. All rights reserved.
1558 B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567

design (Francois et al., 2005; Skogestad, 2000; Srinivasan, Primus, p and  , the KKT conditions also known as the rst-order nec-
Bonvin, & Ricker, 2001). essary conditions of optimality (NCO) must hold at such optimal
These approaches are reviewed in the rst part of the paper. points (Bazaraa, Sherali, & Shetty, 1993). For Problem (3), the KKT
A critical discussion follows, which argues in favor of modier- conditions read:
adaptation methods. It is shown that the modier-adaptation
Gk ( , ) 0, k 0, k = 1, . . . , ng ,
methods do indeed share the advantages of the other methods, i.e.,
in converging to a KKT point for the plant and also in exhibiting fast   G ng

convergence since it is model-based. These issues are illustrated ( , ) + k k ( , ) = 0, (4)

through the case study of a semi-batch reactor system. k=1

k Gk ( , ) = 0, k = 1, . . . , ng .
2. Static optimization problem
where  Rng is the vector of KKT multipliers. The KKT conditions
For continuous processes operating at steady-state, optimiza- involve the quantities G1 , . . . , Gng , G1 /, . . . , Gng /, /,
tion typically consists in determining the operating point that which are denoted collectively by the vector C RnC subsequently,
optimizes process performance such as minimization of operating with nC := ng + n (ng + 1).
cost or maximization of production rate, while satisfying a num- The rst step in satisfying the NCO for a given problem consists in
ber of constraints (such as bounds on process variables or product determining the active set. This discrete decision can be taken based
specications). In mathematical terms, this optimization problem on the sign of the KKT multipliers. Once the active set represented
can be stated as follows: by Ga is known, the multipliers can be eliminated from (4), thereby
leading to two sets of conditions:
minimize : p () := (, yp )   +  
 (1)  Ga Ga
subject to : Gp () := g(, yp ) 0 a
G = 0, r  := I = 0. (5)
where  Rnand yp Rny
stand for the process input (set point)
and output vectors, respectively;  : Rn Rny R is the plant The former guarantees that the active constraints are met,
performance index; and g : Rn Rny Rng is the vector of con- whereas the latter forces the reduced gradient to 0. In real-time
straints imposed on the input and output variables. It is assumed, optimization, these two sets of conditions can be treated using quite
throughout this paper, that  and g are known and the output different techniques. This important aspect will be discussed in the
variables yp can be either measured or estimated. next section.
In contrast to continuous processes, the optimization of batch
and semi-batch processes consists in determining time-varying 3. Classication of real-time optimization schemes
control proles, u(t). This typically involves solving a dynamic opti-
mization problem, possibly with path and terminal constraints. A RTO improves process performance iteratively by adjusting
practical way of solving such problems is by parameterizing the selected optimization variables using measurement data. The goal
control proles using a nite number of parameters , e.g., a poly- of this closed-loop adaptation is to drive the operating point
nomial approximation of u(t) on nite elements. Although the towards the actual plant optimum in spite of inevitable structural
process is dynamic in nature, a static map can then be used to and parametric model mismatch. RTO methods can be classied in
describe the relationship between the process inputs  and the different ways. This paper proposes a classication based on the
outcome of the batch yp . Hence, the problem can be regarded as type and the objective of such adaptation.
a nite-dimensional static optimization problem similar to (1), The three columns in Fig. 1 differ in the way the adaptation
and the optimization approaches discussed in the following sec- is conducted. In model-parameter adaptation methods, a rst-
tions can also be used in the framework of run-to-run optimization principles model is used, the parameters of which are updated
of batch and semi-batch processes (Francois et al., 2005). The based on measurement data; in turn, the updated model is used to
approach will be illustrated in Section 5. compute the optimal inputs. The middle column corresponds to the
In practice, the mapping relating the process inputs and outputs update of so-called modiers, which capture the difference between
is typically unknown, and only an approximate model is available, the NCO predicted by the model and those of the plant; that is, a
e.g., in the form: rst-principles model is used for prediction, but its parameters are
not adapted. In the third column, nally, the inputs are adapted
y = f (, ), (2) directly via tracking of an appropriate set of controlled variables.
The rows of the classication in Fig. 1, on the other hand, cor-
with y Rny
representing the model outputs,  Rn
the model
respond to the two NCO parts discussed in Section 2, namely the
parameters, and f : Rn Rn Rny the input-output mapping.
active constraints and the reduced gradient. The former relates to
Accordingly, an approximate solution of Problem (1) is obtained
the problem of feasibility, while the latter relates to the issue of opti-
by solving the following model-based optimization problem:
mality. In practical applications, guaranteeing feasible operation is
minimize: (, ) := (, f (, )) typically of higher importance than achieving the best possible per-
 (3) formance. Interestingly, the results of a variational analysis in the
subject to : G(, ) := g(, f (, )) 0.
presence of small parametric error support the priority given to
This formulation assumes that the functions  and g are known constraint satisfaction over the gradient part of the NCO (Chachuat
and independent of the model parameters ; the dependency in  et al., 2008).
of these functions is implicit via the model outputs y.
If the objective and constraint functions in (1) and (3) are contin- 3.1. Model-parameter adaptation
uous and the feasible domains of these problems are nonempty and
bounded, (globally) optimal solution points p and  are guaran- The standard way of implementing RTO is via the so-called two-
teed to exist for (1) and (3), respectively. Note that there may also step approach (Chen & Joseph, 1987; Marlin & Hrymak, 1997), also
be multiple local optima to these problems due to nonconvexity. referred to as repeated identication and optimization in the litera-
Provided that the active constraints satisfy a regularity condition at ture. In the rst step, the values of (a subset of) the adjustable model
B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567 1559

Fig. 1. Classication of real-time optimization approaches based on adaptation strategy, feasibility and optimality.

parameters  are estimated using the process output measurements eter estimates, and the lower the required excitation. But too few
yp . This is typically done by minimizing the lack of closure in the adjustable parameters can lead to completely erroneous models,
steady-state model Eq. (2). In the second step, the updated model is and therefore to a false optimum. The question of how plant exper-
used to determine a new operating point by solving an optimization iments should be designed to contain sufcient information has
problem such as (3). been addressed, e.g., by Yip and Marlin (2003).
Model-parameter adaptation schemes can be written generi- It is well known that the interaction between the model-
cally using the following two equations (see Fig. 2): parameter adaptation and reoptimization steps must be considered
carefully for the two-step approach to achieve optimality. If the
 k = ad (yp (k ), k ), (6)
model is structurally correct and the parameters practically iden-
k+1 = opt ( k ) (7) tiable, convergence to the plant optimum can be achieved in
one iteration. However, in the presence of plant/model mis-
where ad : Rny Rn
is the map describing model- match, whether the scheme converges, or to which point it
parameter adaptation; and opt : Rn Rn , the map describing converges, becomes anybodys guess. This is due to the fact that
the optimization step. Note that the handles for correction, i.e., the the adaptation objective might be unrelated to the cost and
adjustable model parameters , are a subset of the model parame- constraints in the optimization problem; hence, minimizing the
ters. A possible adaptation strategy is to minimize the weighted sum mean-square error in y may not help in our quest for feasibility and
of squared errors between the measured and predicted outputs, optimality.
Model exibility is a critical issue with model-parameter adapta-
 k = argminyp (k ) f(k , ).
 tion methods. As far as feasibility is concerned, the selected model
parameterization should permit sufcient exibility for the mea-
In practice, the results of both the identication and optimiza-
sured and predicted constraint values to match. Since the constraint
tion steps clearly need to be diagnosed before implementation; for
expressions g are independent of , this latter requirement is auto-
example, Miletic and Marlin (1998) have proposed to implement
matically enforced via output matching, i.e., y = yp . In the case
only signicant changes from current operation in the sense of a
of ny independent outputs, a necessary condition for feasibility
statistical test.
is therefore that the number of adjustable parameters be greater
A key, yet difcult, decision in model-parameter adaptation is
than the number of outputs, n ny . In terms of optimality, the
to select the parameters to be adapted. These parameters should
selected model parameterization should permit sufcient exi-
be identiable, represent actual changes in the process, and con-
bility for component-wise matching of C and Cp . To achieve this,
tribute to approach the process optimum; also, model adequacy
matching of y and yp alone is no longer sufcient, and it must
proves to be a useful criterion to select candidate parameters for
be complemented with matching of the output derivatives, i.e.,
adaptation (Forbes, Marlin, & MacGregor, 1994). Clearly, the smaller
y/ = yp /. Provided that all the outputs and output deriva-
the subset of parameters, the better the condence in the param-
tives are independent, a model parameterization with at least
ny (n + 1) degrees of freedom is then necessary to achieve optimal-
ity. With such a parameterization, it would also become possible to
modify the identication map ad in (6), e.g., as the minimization
of the weighted sum of squared errors between Cp and C, thereby
explicitly enforcing KKT matching.
Convergence under plantmodel mismatch has been addressed
by several authors (Biegler, Grossmann, & Westerberg, 1985; Forbes
et al., 1994). It has been shown that optimal operation is reached if
model adaptation leads to matched KKT conditions for the model
Fig. 2. Optimization via model-parameter adaptation: two-step approach. and the plant.
1560 B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567

Convergence result 1 Let the model-parameter adaptation (6)

be such that the quantities C predicted by the model match the
plant measurements Cp . Then, upon convergence, the model-parameter
adaptation scheme (6)(7)reaches a KKT point for the plant.

3.2. Modier adaptation

In order to overcome plantmodel mismatch, several variants

of the two-step approach have been presented in the literature. Fig. 3. Optimization via modier adaptation: matching the KKT conditions.

Generically, they consist in modifying the cost and constraints of the

optimization problem in such a way that the KKT conditions of the The handles for correction are now the modier terms  instead
model and the plant can match. Letting k denote the actual oper- of the parameters  used in the context of model-parameter adap-
ating point, the optimization problem with modiers to be solved tation. The simplest adaptation strategy is to calculate the modiers
during the next RTO execution is given by by making the full correction (9)(11) at each RTO execution, i.e.,
T k = Cp (k ) C(k , ). However, this might lead to excessive cor-
minimize: m (, ) := (, ) + ( k )
 (8) rections when operating far away from the optimal operating point
subject to: Gm (, ) := G(, ) + G + G ( k ) 0, and might make modier adaptation very sensitive to measure-
ment noise. A better strategy consists in ltering the modiers, e.g.,
where G Rng is the constraint-value modier, G Rn ng with a rst-order exponential lter (Marchetti, Chachuat, & Bonvin,
the constraint-gradient modier, and Rn the cost-gradient in press):
modier; these modiers will be denoted collectively by  =
G1 T Gng T T k = (I K)k1 + K[Cp (k ) C(k , )], (14)
[G1 , . . . , Gng ,  ,..., , ] RnC subsequently.
where K RnC nC is a (nonsingular) gain matrix.
The constraint-value modier represents the difference between The experimental quantities Cp required for adaptation are
the experimental and predicted constraints, directly related to the KKT conditions. In practice, the output
measurements yp can be used to estimate these quantities. The
Gi = [Gp,i () Gi (, )]= , i = 1, . . . , ng . (9)
k experimental constraint values Gp () are readily estimated as
Adapting only G leads to the so-called constraint-adaptation g(, yp ). This way, feasibility is enforced, though no guarantee can
scheme (Chachuat et al., 2008; Forbes & Marlin, 1994). Such a be given regarding the correct determination of the active set. On
scheme is rather straightforward and corresponds to common the other hand, estimating the experimental cost and constraint
industrial practice (Marlin & Hrymak, 1997). gradients, (p /)() and (Gp /)(), is more involved and is
The cost-gradient modier represents the difference between the one of the major bottlenecks in the implementation of modier-
experimental and predicted values of the cost gradient, adaptation schemes.
  Observe the one-to-one correspondence in the adaptation law
p  (14) between the number of components in (Cp C) and the num-
= () (, ) . (10)
  ber of adjustable parameters in . Since the Jacobian of (Cp C)
with respect to the modiers is the gain matrix K (a full-rank
The pertinent idea of adding a gradient modier to the cost matrix), identiability is automatically veried, and so are the
function of the optimization problem dates back to the work of KKT-matching conditions upon convergence. These nice theoretical
Roberts (1979). Note that it was originally proposed in the frame- properties are formalized next.
work of two-step methods to better integrate the model update Convergence result 2 Let the cost and constraint functions be
and optimization subproblems, which has led to the so-called parameterized as in Problem (8). Let also the information on the val-
ISOPE approach (see Brdys & Tatjewski, 2005 for an overview). The ues of Cp be available and used to adapt the modiers . Then, upon
idea of modifying the cost gradient without the need to estimate convergence, the modier-adaptation scheme (12) and (13) reaches a
the model parameters was introduced by Tatjewski (2002). KKT point of the plant.
The constraint-gradient modier, nally, represents the differ-
ence between the experimental and predicted values of the 3.3. Direct input adaptation
constraint gradients,
Gp,i Gi This last class of methods provides a way of avoiding repeated
= () (, ) , i = 1, . . . , ng . (11) numerical optimization by transforming the optimization prob-
=k lem into a feedback control problem that calculates (set points
The idea of adding such a rst-order modier term to the process- for) the input variables. This is motivated by the fact that practi-
dependent constraints, in addition to the constraint bias G , was tioners like to use feedback control of selected variables as a way
proposed recently by Gao and Engell (2005). This modication to counteract plantmodel mismatch and plant disturbances, due
permits the matching, not only of the constraint values, but also to its simplicity and reliability compared to on-line optimization.
the constraint gradients. The challenge is to nd functions of the measured variables which,
when held constant by adjusting the input variables, enforce opti-
The update laws in modier-adaptation schemes can be written mal plant performance (Morari, Stephanopoulos, & Arkun, 1980;
generically as (see Fig. 3): Skogestad, 2000). In other words, the goal is to choose appropriate
controlled variables in order to achieve a similar steady-state per-
k = ad (Cp (k ) C(k , )), (12) formance as would be realized by a (ctitious) on-line optimizing
  controller; these variables are denoted by Zp subsequently, and the
k+1 = opt k (13)
corresponding references by Zref .
where ad : RnC RnC stands for the modier-adaptation map; In self-optimizing control (Skogestad, 2000), the idea is to use a
and opt : RnC Rn , the optimization map. process model to select linear combinations of the output variables,
B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567 1561

Here act is the map describing the selection of controlled vari-

ables, including active set detection (switching logic). Note that
only the plant outputs yp (k ) are needed to dene this map in self-
optimizing control, whereas all KKT quantities Cp (k ) are required
for tracking the NCO components in (15). con : Rn Rn is the
map describing the square multivariable controller. A simple choice
for this map is in the form of a discrete integral-type controller,
Fig. 4. Optimization via direct input adaptation: tracking appropriate controlled
variables using feedback control.
k+1 = k + K(Zref Zp (k )), (18)
where K Rn nis a gain matrix. Note that, unlike optimization-
the tracking of which results in good performance: based schemes, the required regularization is provided naturally
via appropriate controller tuning.
Zp = Myp , A crucial question in practice regards the design of the multi-
variable controller con . Since the constrained quantities can often
where M is an appropriately chosen matrix. In particular, differ- be estimated reliably from the output measurements yp , adjusting
ent methodologies have been investigated for the choice of M, the inputs to track the active constraints is rather straightforward
which are based on (near) invariance in the presence of uncertainty (Bonvin & Srinivasan, 2003; Srinivasan et al., 2001; Stephanopoulos
(Govatsmark & Skogestad, 2005). & Arkun, 1980). On the other hand, gradient estimation is more
However, tracking linear combinations of the output variables, involved and, furthermore, may not lead to signicant performance
although very convenient, hardly provides any guarantee that a improvement. Hence, an interesting option is to devise a two-
plant optimum is reached upon convergence. To circumvent this time-scale adaptation strategy, wherein adaptation of the reduced
difculty, a rather natural idea is to choose the controlled variables gradient is implemented at a much slower rate than that of the
Zp as the NCO components, active constraints (Francois et al., 2005).

T Similar to modier adaptation, direct input adaptation also pos-
Zp = Gap r p , (15) sesses nice theoretical properties; for example, when the controller
tracks the NCO components (15), the following convergence result
with the corresponding set points Zref = 0, thereby enforcing the
plant NCO. Two classes of approaches fall within this category:
Convergence result 3 Let the information on the values of Cp be
extremum-seeking control (Ariyur & Krstic, 2003; Guay & Zhang,
available and used to select the active set and track the variables Zp to
2003), and NCO tracking (Francois et al., 2005; Srinivasan, Biegler,
Zref = 0. Then, upon convergence, the input-adaptation scheme (16)
& Bonvin, 2008).
and (17)reaches a KKT point of the plant.
Perhaps the main difculty with direct adaptation methods lies
in the fact that the tracked variables do change with the active
4. Comparison of approaches
set. Also, detection of such a change needs to be dealt with sep-
arately from the tracking problem. One possibility is to use a
In this section, we take a critical look at the three aforemen-
switching strategy, whereby all constraints and KKT multipliers
tioned classes of adaptive optimization schemes. Table 1 compares
are monitored: a constraint is included in the active set when it
these approaches in terms of various criteria. It is interesting to
is binding, and deactivated when the associated KKT multiplier
see that modier adaptation can be positioned between model-
becomes negative (Woodward, Perrier, & Srinivasan, 2007). Another
parameter adaptation and direct input adaptation; several features
approach consists in precomputing regions that correspond to
are shared between the rst and second columns, while other
different sets of active constraints based on a (steady-state) pro-
features are shared between the second and third columns. This
cess model (Manum, Narasimhan, & Skogestad, 2008). Since these
comparison shall lead us to argue in favor of modier-adaptation
switching approaches render the scheme more complex, it is often
methods, in the sense that they provide a parameterization and an
assumed, in the interest of simplicity, that the active constraints do
update criterion that are well tailored to KKT matching.
not change. Note that such an assumption is typically veried for
The methods differ greatly in the adjustable parameters and the
small parametric variations in the case of a strict minimum and is
real-time information (either measured or estimated) that is used
observed in many practical situations.
to make the adaptation. Model-parameter adaptation considers the
Input-adaptation schemes obey the following equations (see
parameterization  and the output measurements yp , which makes
Fig. 4):
KKT matching difcult to achieve in general. In contrast, modier
Zp (k ) = act (yp (k )), or Zp (k ) = act (Cp (k )), (16) adaptation resolves the challenging task of selecting the adjustable
parameters by introducing the modiers  as handles and assum-
k+1 = con (Zp (k )). (17) ing that the KKT-based quantities Cp are available in real-time. Since

Table 1
Comparison of various adaptation strategies for real-time optimization.

Model-parameter adaptation Modier adaptation Direct input adaptation

Adjustable parameters [n ] [nC ]

Real-time information a yp [ny ] Cp [nC ] Cp [nC ] b
Adaptation criterion for adjustable parameters minyp y  = Cp C

Update of RTO inputs Optimization Optimization Control
Regularization External lter External lter Controller tuning
Active set detection Optimizer Optimizer Switching logic
Requirement for feasibility Model exibility, Ga = Gap b
Requirement for optimality Model exibility, C = Cp b
Either measured or estimated.
For the case of a controller tracking the NCO components (15).
1562 B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567

the number of components in Cp equals the number of handles, a 5. Case study

square adaptation problem is obtained, and KKT matching is read-
ily enforced upon convergence. In direct input adaptation, there The considerations discussed in the previous sections of the
are no adjustable parameters as the handles are the input variables paper are now illustrated for the run-to-run optimization of a semi-
 themselves; for the case of NCO tracking, convergence to a KKT batch reactor process. This problem is recast as a static optimization
point again requires that all Cp components be available. problem via parameterization of the time-varying control prole,
Major differences also exist in terms of the adaptation crite- as discussed in Section 2.
rion for the adjustable parameters. Model-parameter adaptation The reaction system is the acetoacetylation of pyrrole with
proceeds by solving an optimization problem. Sufcient model ex- diketene and consists of 4 reactions:
ibility (see Section 3.1) is needed to bring y exactly to yp , which
requires at least as many adjustable parameters as there are mea- A + BC
surements, n ny , in the case of independent outputs. Moreover, k2
for the parameter estimates to be accurate, the available measure-
ments must contain sufciently rich information. Note that one BE
could consider the alternative criterion minCp C to facilitate KKT C + BF,

matching. However, with the available n degrees of freedom, it may
be even more difcult to bring the nC components of C exactly to Cp . with A: pyrrole; B: diketene; C: 2-acetoacetyl pyrrole; D:
In modier adaptation, one deals with a square adaptation problem, dehydroacetic acid; E: oligomers; F: undesired by-product. A rst-
wherein the handles  are essentially decoupled and can be read- principles model for the semi-batch reactor is as follows:
ily computed from Cp and C. In direct input adaptation, nally, the
adaptation corresponds to a square multivariable control problem. cA = k1 cA cB cA (19)
A difculty therein is that a natural pairing between the manipu-
lated variables  and the controlled variables Zp may not exist, thus F in
complicating the task of designing the RTO controller. The foregoing cB = k1 cA cB 2k2 cB2 k3 cB k4 cB cC + (c cB ) (20)
considerations lead us to argue that modier-adaptation schemes
possess the adequate parameterization and use the adequate update F
cC = k1 cA cB k4 cB cC cC (21)
criterion for real-time optimization. V
It should also be noted that direct input adaptation differs from F
the other two adaptation schemes in that an optimization problem cD = k2 cB2 cD (22)
is not solved at each iteration, thus removing much of the on-line
computational burden and complexity. On the other hand, it can V = F, (23)
be argued that the use of an on-line optimizer helps determine the
active set and alleviates the complicated task of devising a multi- where cA , cB , cC and cD stand for the concentrations of species A, B,
variable controller at the RTO level. Also, faster convergence may be C and D, respectively; V, the reactor volume; F, the inlet ow rate
achieved with an on-line optimizer, although some amount of l- of species B; and cBin , the concentration of B in the feed. The initial
tering (regularization) is needed to reduce sensitivity to noise and values for the state variables are reported in Table 2. See Ruppen,
guarantee stability in the presence of plantmodel mismatch. Bonvin, and Rippin (1998) for a detailed description of the process
In terms of constraint satisfaction, the ability of model- and model.
parameter adaptation to yield a feasible operating point is directly Throughout this case study, the full reaction mechanism is con-
linked to model exibility. With modier adaptation, on the other sidered for the simulated plant, with the corresponding kinetic
hand, feasibility is automatically guaranteed upon convergence parameter values given in Table 2. On the other hand, only the rst
as long as the active process constraints Gap can be measured two reactions (A + B C and 2B D) are taken into account in
or estimated. With input adaptation, nally, feasibility is trivially the plant model, i.e., the remaining two reaction are considered to
established when either the active set remain unchanged with be unknown side reactions (k3 = k4 = 0). This way, the model has
uncertainty or active set changes can be detected; otherwise, track- structural mismatch.
ing an inappropriate set of constraint may lead to infeasibility.
Last but not least, the need for experimental gradient informa-
5.1. Optimization Problem A
tion is an important element of comparison. Modier adaptation
and direct input adaptation typically make use of experimental gra-
The rst problem consists in determining the feed prole of
dients, while model-parameter adaptation does not. In the former
species B that maximizes the number of moles of C at nal time,
schemes, gradient information is key to achieving feasible and opti-
while keeping the concentrations of B and D at terminal time below
mal operation despite model mismatch and process disturbances.
specied threshold values:
Several techniques for estimating the plant gradients have been
proposed, which differ in terms of their relying on a model or
maxF(t) J A := cC (tf )V (tf ) (24)
not, as well as their use of steady-state vs. transient measurement
data; these techniques, which greatly vary in accuracy too, include:
(a) nite differences (Roberts, 1979), (b) black-box identication
(Bamberger & Isermann, 1978, Golden & Ydstie, 1989; Mansour & Table 2
Parameters values and initial conditions.
Ellis, 2003; Zhang & Forbes, 2006), (c) dither-signal superposition
(Ariyur & Krstic, 2003; Guay & Zhang, 2003, (d) linearization (neigh- Parameter Value Concentration Value
boring extremals) (Gros, Srinivasan, & Bonvin, 2009), (e) Broyden k1 0.053 L mol1 min1 cA (0) 0.72 mol L1
methods (Gao & Engell, 2005; Mansour & Ellis, 2003), and (f) use k2 0.128 L mol1 min1 cB (0) 0.05 mol L1
of multiple parallel units (Srinivasan, 2007; Woodward, Perrier, & k3 0.028 min1 cC (0) 0.08 mol L1
Srinivasan, 2009). Which approach to use is clearly problem depen- k4 0.001 L mol1 min1 cD (0) 0.01 mol L1
cBin 5 mol L1 V (0) 1L
dent and lies beyond the scope of this paper.
B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567 1563

Fig. 5. Optimal operation in Problem A for two different control parameterizations. Dash-dotted line: CVP with 100 constant-length control stages. Solid line: CVP with 3
variable-length control stages.

the measured and predicted concentrations of species B, C and D,

s.t. Model (19 23)

cB (tf ) cBmax   c (t ; )
cD (tf ) cD (k1 , k2 ) = argmin 1 i f , (25)
 cp,i (tf )
i {B,C,D} =k
0 F(t) F max ,

with cp,i (tf ) denoting the concentration of species i measured at the

with the nal time tf = 250 min; the maximal inlet ow rate F max = end of the batch. To avoid parameter corrections that would be too
2 103 L min1 ; and the maximal concentrations of species B and aggressive and reduce sensitivity to measurement noise, the tted
D at nal time cBmax = 0.025 mol L1 and cD max = 0.15mol L1 .
kinetic coefcients are ltered via a simple exponential lter,
The optimal control and state trajectories, as obtained from
two different control parameterizations, are shown in Fig. 5.      
k1,k k1,k1 k1
The piecewise-constant parameterization over 100 constant-length = (1  ) + 
k2,k k2,k1 k2
stages shows that the optimal feed prole consists of 3 parts or
arcs: (i) a rst arc where F = F max ; (ii) a second arc where the
feed level is intermediate between 0 and F max ; and (iii) a last arc with the lter parameter  (0, 1]. In turn, the ltered kinetic
where the feed rate is equal to 0. Moreover, both terminal con- coefcients are used to update the control variables k+1 =
straints are active in the optimal solution. Also shown in Fig. 5 are (tm,k+1 Fs,k+1 ts,k+1 )T by solving the optimization problem (24).
the results for a control parameterization with 3 variable-length The results for this adaptation scheme, with  = 1.0 and 0.5,
stages. With this parameterization, the input (decision) variables are shown in the upper row of Fig. 6. Observe rst that the ter-
are  := (tm Fs ts )T with tm , the switching time between the rst and minal constraint on cB is slightly violated when no ltering is
second arcs; Fs , the (constant) feed rate along the second arc; and ts , used (  = 1), but this violation can be avoided through lter-
the switching time between the second and third arcs. The values ing (  = 0.5). In terms of performance, some improvement is
of the objective function for both parameterizations are very close, obtained the converged cost value being around Jpar A 0.43 mol

thus indicating that this latter parameterization yields an adequate , yet an important optimality gap persists. In particular, the
approximation. leftmost plot indicates that the adapted control prole remain
Solving the discretized optimal control problem using the plant signicantly different from the optimal prole, especially the con-
model (i.e., with k3 = k4 = 0) gives the same optimal solution struc- trol value Fs . The difculty in this case is linked to the fact that
ture (3 arcs, 2 active terminal constraints). Open-loop application of the model parameters  do not permit sufcient exibility for
this solution, although feasible, yields a much worse performance the model to match the experimental measurements; in particu-
A 0.3874 mol compared to J A 0.5079 mol.
Jol opt lar, we have n = 2 < ny = 3 here. A way to circumvent this lack
The use of various real-time optimization techniques to reduce of exibility would be, e.g., to add a ctitious reaction to the
this large optimality gap is investigated next. Both terminal con- plant model and estimate the additional kinetic coefcient so that
straints being active in the optimal solution, only constraint n = ny = 3. Another way to guarantee feasibility upon conver-
satisfaction schemes are considered, i.e., without cost and con- gence would be to not take the measurement of cC into account
straint gradient correction. It is assumed that measurements for in the estimation problem (25) so as to force the condition Ga = Gap
the reactor volume and the concentrations of species B, C and D (see Table 1).
are available at the end of the batch. Moreover, a 5% zero-mean Modier adaptation The optimization problem (24) is modi-
Gaussian multiplicative noise is added to the concentration mea- ed by introducing the constraint-value modiers B and D , while
surements, in response to which a 10% back-off is dened for the the cost and constraint gradients remain uncorrected. In the k th
terminal constraints in the optimization problem. RTO execution, the updated control variables k are determined by
Model-parameter adaptation The adjustable model parame- solving the optimization problem:
ters correspond to the kinetic coefcients of the rst two reactions,
 := (k1 , k2 ). At the end of the k th RTO execution, their values are max cC (tf )V (tf ) (26)
updated by minimizing a weighted sum of squared errors between tm ,Fs ,ts
1564 B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567

Fig. 6. Application of real-time optimization to Problem A. Upper plots: model-parameter adaptation. Middle plots: modier adaptation. Lower plots: direct input adaptation.

s.t. Model (19 23)

adapted and thus do not match those of the plant at the converged
cB (tf ) + B,k1 cBmax operating point.
It should be noted that implementation of the full modier
cD (tf ) + D,k1 cD scheme has been found to be benecial only in the case where
0 Fs F max . reliable gradient estimates are available. In particular, it could be
shown that adapting the gradient part is detrimental to the opti-
Next, constraint-modier adaptation proceeds through a rst-order mization objective with a 10% multiplicative white noise in the
exponential lter, gradient estimates.
   Direct input adaptation Both terminal constraints are binding
B,k B,k1 cp,B (tf ) cB (tf )
= (1 ) + , according to the nominal model-based solution of Problem (24).
D,k D,k1 cp,D (tf ) cD (tf ) Two degrees of freedom are thus necessary to track the active
constraints. In terms of performance, it is found that the occur-
with the lter parameter (0, 1]. rence of the switching time tm has a very limited inuence; for
The results for modier adaptation, with the initial constraint- instance, forcing tm = 0 in the parameterization (thereby leaving
modier values B,0 = B,0 = 0, are shown in the middle row of only 2 arcs), would result in a performance loss lower than 0.1%.
Fig. 6 for the lter parameters = 1.0 and 0.5. Most of the optimal- For simplicity, this switching time is xed at its nominal value,
ity gap is recovered upon application of this simple scheme, with tm = 4.71 min, subsequently. There results a 2 2 control prob-
the cost value on the order of Jmod 0.49 mol. The remaining loss lem, with the manipulated variables (Fs , ts ), the controlled variables
is incurred by (i) the back-offs imposed on the terminal constraints, (cp,B (tf ) cBmax , cp,D (tf ) cD
max ), and the set-points (0,0). It is found,

and (ii) the fact that the cost and constraint gradients have not been using the relative gain array (RGA) technique (Ogunnaike & Ray,
B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567 1565

Fig. 7. Optimal operation in Problem B for two different control parameterizations. Dash-dotted line: CVP with 100 constant-length control stages. Solid line: CVP with 2
variable-length control stages.

1994), that two independent control loops can be devised by pair- B 0.2135 mol, i.e., not too far from the
the open-loop solution is Jol
ing cp,B (tf ) cBmax with ts , and cp,D (tf ) cD
max with F . Simple discrete
s B
plant optimum Jopt 0.2179 mol.
integral controllers are then used for each loop, Various real-time optimization techniques are investigated next
       to reduce this small optimality gap. Because the solution is uncon-
ts,k ts,k1 ts cp,B (tf ) cBmax
= max , strained, the emphasis is on gradient-correction schemes. The
Fs,k Fs,k1 Fs cp,D (tf ) cD
=k1 same measurements as in Problem A are considered. In addi-
tion, estimates of the cost derivatives with respect to the control
with the controller parameters ts [L min mol1 ] and Fs parameters Fs and ts are calculated numerically for each batch and
[L2 min1 mol1 ]. adding a 10% zero-mean Gaussian multiplicative noise to these
The results of direct input adaptation are given in the lower values.
row of Fig. 6, for two different controller settings, starting from the Model-parameter adaptation The values of the adjustable
nominal values of the control variables. Observe that the perfor- model parameters k1 and k2 are updated in a similar way as pre-
mance and adaptation speed of direct input adaptation are similar viously in Problem A. The results for this adaptation scheme, with
to those of modier adaptation. Moreover, on-line complexity is  = 1.0 and 0.5, are shown in the upper row of Fig. 8. Quite unex-
greatly reduced in this approach compared to the other schemes. pectedly, model-parameter adaptation in this case results in a loss
of performance (JparB 0.198 mol) compared to the open-loop nom-

5.2. Optimization Problem B B

inal solution Jol . The reason for this behavior is twofold. Firstly,
the selected parameter identication criterion does not adequately
The second problem differs from Problem A in that large feeds match the optimization objective; and secondly, the kinetic coef-
of B are now penalized in the objective function: cients k1 and k2 do not allow sufcient exibility for the model to
 tf match the experimental data; matching the components of C and
maxF(t) J B := cC (tf )V (tf ) F(t)2 dt (27) Cp would require ny (n + 1) = 12 degrees of freedom, whereas only
0 n = 2 are available.
s.t. Model (19 23) Modier adaptation The optimization problem (27) is modi-
cB (tf ) cBmax ed by adding a gradient-modier term to the cost function; on
cD (tf ) cDmax the other hand, the constraints are left uncorrected for simplicity,
since the optimal solution is known to be unconstrained. The con-
0 F(t) F max ,
trol variables k applied during the k th RTO execution are obtained
by solving the optimization problem:
with the penalty coefcient taken as = 2500 L2 min1 mol1 . The
parameters and initial conditions are the same as those given in max cC (tf )V (tf ) ts Fs2 + k1 ( k1 ) (28)
Fs ,ts
Table 2, and so are the constraint bounds.
The optimal control and state trajectories, as obtained from
two different control parameterizations, are shown in Fig. 7. s.t. Model (19 20)
The piecewise-constant parameterization over 100 constant-length cB (tf ) cBmax
stages shows that the optimal feed prole now consists of one single max
cD (tf ) cD
arc, and both terminal constraints are inactive. Also shown in Fig. 7
are the results of a control parameterization with 2 variable-length 0 Fs F max ,
stages, for which the inputs  := (Fs ts )T consist of a constant ow where  R2 denotes the vector of cost-gradient modiers. Adap-
rate Fs followed by a zero ow rate, the switching between the two tation of these modiers proceeds as follows:
occurring at time ts . The values of the objective function for both  
parameterizations are very close, thus indicating that this 2-stage p 
k = (1 )k1 + ,
CVP yields an accurate approximation.  
Solving this optimal control problem using the plant model (i.e.,
with k3 = k4 = 0) results in the same structure for the optimal solu- with the lter parameter (0, 1] and  := cC (tf )V (tf ) ts Fs2
tion (2 arcs, all terminal constraints inactive); the performance of denoting the cost function.
1566 B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567

Fig. 8. Application of real-time optimization to Problem B. Upper plots: model-parameter adaptation. Middle plots: modier adaptation. Lower plots: direct input adaptation.
The actual plant optimum is represented by the symbol
on the right plots.

The results for modier adaptation, with the initial modier val- are used,
ues 0 = (0 0)T , are shown in the middle row of Fig. 8 for the
lter parameters = 0.5 and 0.2. This scheme converges quickly
to the actual plant optimum. Since no terminal constraint is active, ts,k ts,k1 ts p /ts
the residual optimality gap results from the effect of measurement = ,
Fs,k Fs,k1 Fs p /Fs
noise only. The experimental gradient estimates being corrupted =k1

with large amounts of noise, the use of a small value of the lter
parameter produces better results in this case.
Direct input adaptation The optimal solution of Problem (27) with the controller parameters ts [min2 mol1 ] and Fs
being unconstrained, the available degrees of freedom can be used [L2 min2 mol1 ].
to force the cost gradient to zero. The resulting 2 2 control prob- The results of direct input adaptation are given in the lower row
lem has the manipulated variables (Fs , ts ), the controlled variables of Fig. 8, for two different controller settings. To limit the amount of
(p /Fs , p /ts ) and the set-points (0,0). The control map is interaction between the control loops, the controller Fs p /Fs
found to be scale-independent and diagonally dominant via RGA is tuned to proceed at a faster rate than the controller ts p /ts .
analysis, which leads to the natural pairing of p /Fs with Fs and The convergence of this direct input-adaptation scheme appears to
of p /ts with ts ; yet, some amount of interaction is also detected be rather slow, even though the iterates eventually reach a neigh-
between the two control loops. Simple discrete integral controllers borhood of the actual plant optimum.
B. Chachuat et al. / Computers and Chemical Engineering 33 (2009) 15571567 1567

6. Conclusions Gao, W., & Engell, S. (2005). Iterative set-point optimization of batch chromatogra-
phy. Computers and Chemical Engineering, 29, 14011409.
Golden, M. P., & Ydstie, B. E. (1989). Adaptive extremum control using approximate
This paper provides a classication of real-time optimization process models. AIChE Journal, 35(7), 11571169.
schemes and analyzes their ability to enforce feasible and optimal Govatsmark, M. S., & Skogestad, S. (2005). Selection of controlled variables and robust
process operation by using on-line measurements. The similarities setpoints. Industrial and Engineering Chemistry Research, 44(7), 22072217.
Gros, S., Srinivasan, B., & Bonvin, D. (2009). Optimizing control based on output
and differences between the various schemes have been high- feedback. Computers and Chemical Engineering, 33(1), 191198.
lighted, and it has been argued that modier-adaptation schemes Guay, M., & Zhang, T. (2003). Adaptive extremum seeking control of nonlinear
use a model parameterization and an update criterion that are tai- dynamic systems with parametric uncertainty. Automatica, 39, 12831294.
Maarleveld, A., & Rijnsdorp, J. E. (1970). Constraint control on distillation columns.
lored to the matching of KKT conditions. These considerations have
Automatica, 6, 5158.
been illustrated through the run-to-run optimization of a semi- Mansour, M., & Ellis, J. E. (2003). Comparison of methods for estimating real process
batch reactor system in the presence of structural model mismatch. derivatives in on-line optimisation. Applied Mathematical Modelling, 27, 275291.
Manum, H., Narasimhan, S., & Skogestad, S. (2008). Explicit MPC with output feed-
The combination of specic features of the various approaches
back using self-optimizing control. In Proc 17th World Congress of IFAC Seoul,
certainly holds much promise towards improved performance of Korea, (pp. 69566961).
adaptive optimization. For example, the combination of model- Marchetti, A., Chachuat, B., & Bonvin, D. (in press). Modifer-adaptation methodol-
parameter adaptation (which ensures fast convergence for the rst ogy for real-time optimization. Industrial and Engineering Chemistry Research,
few iterations and can detect changes in the active set) with direct Marlin, T. E., & Hrymak, A. N. (1997). Real-time operations optimization of continuous
input adaptation (which provides the necessary gradients in the processes. In AIChE Symp Ser - CPC-V, Vol. 93 (pp. 156164).
neighborhood of the plant optimum) has been demonstrated by Miletic, I. P., & Marlin, T. E. (1998). On-line statistical results analysis in real-
time operations optimisation. Industrial and Engineering Chemistry Research, 37,
Srinivasan and Bonvin (2003). Another interesting combination 36703684.
would be to use modier adaptation at one time scale and perform Mnnigmann, M., & Marquardt, W. (2003). Steady-state process optimization with
model-parameter adaptation at a slower rate, thus giving rise to a guaranteed robust stability and feasibility. AIChE Journal, 49(12), 31103126.
Morari, M., Stephanopoulos, G., & Arkun, Y. (1980). Studies in the synthesis of control
two-time-scale adaptation strategy. structures for chemical processes. Part I. AIChE Journal, 26(2), 220232.
Ogunnaike, B. A., & Ray, W. H. (1994). Process dynamics, modeling and control. New
York: Oxford University Press.
Roberts, P. D. (1979). An algorithm for steady-state system optimization and param-
eter estimation. International Journal of Systems Science, 10, 719734.
The authors would like to acknowledge useful discussions on Ruppen, D., Bonvin, D., & Rippin, D. W. T. (1998). Implementation of adaptive optimal
operation for a semi-batch reactor system. Computers and Chemical Engineering,
several topics of this paper with Alejandro Marchetti of EPFL and 22, 185189.
Moncef Chioua of cole Polytechnique, Montral. Skogestad, S. (2000). Self-optimizing control: The missing link between steady-
state optimization and control. Computers and Chemical Engineering, 24,
References Srinivasan, B. (2007). Real-time optimization of dynamic systems using multiple
units. International Journal of Robust and Nonlinear Control, 17, 11831193.
Ariyur, K. B., & Krstic, M. (2003). Real-time optimization by extremum seeking. New Srinivasan, B., Biegler, L. T., & Bonvin, D. (2008). Tracking the necessary conditions
York: John Wiley & Sons. of optimality with changing set of active constraints using a barrier-penalty
Bamberger, W., & Isermann, R. (1978). Adaptive on-line steady-state optimization of function. Computers and Chemical Engineering, 32(3), 572579.
slow dynamic processes. Automatica, 14, 223230. Srinivasan, B., & Bonvin, D. (2002). Interplay between identication and optimiza-
Bazaraa, M. S., Sherali, H. D., & Shetty, C. M. (1993). Nonlinear programming: Theory tion in run-to-run optimization schemes. In Proc American Control Conference
and algorithms (2nd Edition). New York: John Wiley & Sons. Anchorage, AK, (pp. 21742179).
Biegler, L. T., Grossmann, I. E., & Westerberg, A. W. (1985). A note on approximation Srinivasan, B., & Bonvin, D. (2003). Convergence analysis of iterative identication
techniques used for process optimisation. Computers and Chemical Engineering, and optimization schemes. In Proc American Control Conference Denver, CO, (pp.
9(2), 201206. 19561961).
Bonvin, D. (1998). Optimal operation of batch reactorsA personal view. Journal of Srinivasan, B., Primus, C. J., Bonvin, D., & Ricker, N. L. (2001). Run-to-run optimization
Process Control, 8(56), 355368. via constraint control. Control Engineering Practice, 9(8), 911919.
Bonvin, D., & Srinivasan, B. (2003). Optimal operation of batch processes via the Stephanopoulos, G., & Arkun, Y. (1980). Studies in the synthesis of control structures
tracking of active constraints. ISA T, 42(1), 123134. for chemical processes. Part IV. AIChE Journal, 26(6), 975991.
Brdys, M., & Tatjewski, P. (2005). Iterative algorithms for multilayer optimizing control. Tatjewski, P. (2002). Iterative optimizing set-point controlThe basic principle
London, UK: Imperial College Press. redesigned. In Proc 15th World Congress of IFAC Barcelona, Spain.
Chachuat, B., Marchetti, A., & Bonvin, D. (2008). Process optimization via constraints Woodward, L., Perrier, M., & Srinivasan, B. (2007). Multi-unit optimization with gra-
adaptation. Journal of Process Control, 18(3/4), 244257. dient projection on active constraints. In Proc 8th Int Symp DYCOPS. Vol. 1. Cancn
Chen, C. Y., & Joseph, B. (1987). On-line optimization using a two-phase approach: An Mexico, (pp. 129134).
application study. Industrial and Engineering Chemistry Research, 26, 19241930. Woodward, L., Perrier, M., & Srinivasan, B. (2009). Improved performance in the
Forbes, J. F., & Marlin, T. E. (1994). Model accuracy for economic optimizing con- multi-unit optimization method with non-identical units. Journal of Process Con-
trollers: The bias update case. Industrial and Engineering Chemistry Research, 33, trol, 19(2), 205215.
19191929. Yip, W. S., & Marlin, T. E. (2003). Designing plant experiments for real-time opti-
Forbes, J. F., Marlin, T. E., & MacGregor, J. F. (1994). Model adequacy requirements mization systems. Control Engineering Practice, 11, 837845.
for optimizing plant operations. Computers and Chemical Engineering, 18(6), Yip, W. S., & Marlin, T. E. (2004). The effect of model delity on real-time optimization
497510. performance. Computers and Chemical Engineering, 28, 267280.
Francois, G., Srinivasan, B., & Bonvin, D. (2005). Use of measurements for enforc- Zhang, Y., & Forbes, J. F. (2006). Performance analysis of perturbation-based meth-
ing the necessary conditions of optimality in the presence of constraints and ods for real-time optimisation. Canadian Journal of Chemical Engineering, 84,
uncertainty. Journal of Process Control, 15(6), 701712. 209218.