Professional Documents
Culture Documents
net/publication/231367023
CITATIONS READS
25 849
3 authors:
Xavier Joulia
Université de Toulouse - INP-ENSIACET
153 PUBLICATIONS 1,226 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Pascal Floquet on 13 April 2018.
This study attempts to address efficient implementation of the normal boundary intersection
(NBI) method for the tradeoff between economics and pollution prevention with appropriate
analysis methods. For the multiobjective optimization problem (MOOP), a noninferior solution
curve (or Pareto curve) is efficiently obtained by using an enhanced NBI method combined with
the SWOF (summation of weighted objective functions) method within a chemical process
simulator. Then, the Pareto curve is analyzed by the ideal compromise solution set, sensitivity
and elasticity analyses that give the decision basis between the conflicting objectives to the
decision maker. Through the combined use of the SWOF and NBI methods, the upper and lower
values of the ideal compromise set are exactly located. The total calculation time for the entire
Pareto curve can be reduced owing to evenly distributed Pareto points.
(f̂2)up ) -() (
1 1
)
f̂ + s1 + (βi)up + 1, (βi)up ) i
s1 1 s1
µ
10
,( )
1 e i e 9 (4)
(
(f̂2)down ) -s1f̂1 + s1 +
1
)
s1 i
1
(β )down - ,
s1
Figure 1. Pareto point B obtained by the NBI method and Pareto
point A obtained by the SWOF method in the normalized objective
(βi)down ) i( )
1-µ
10
+ µ, 1 e i e 9 (5)
space.
where µ corresponds to the f̂1 value at point B, as shown
in Figure 1. Here, two simple relations are introduced
and lack of information about nonconvex parts of the
such that s2 ) 1/s1 and s1 ) (µ - 1)/µ, because the two
Pareto set.7
objective functions are normalized between zero and
The NBI method is a technique motivated by a one. The modified NBI method is also efficient for
geometrical intuition to provide a better parametriza- locating more uniformly distributed Pareto points (non-
tion of the Pareto set than that provided by other inferior solution points) than the SWOF and the GP
techniques. This parametrization is better in the sense methods. Hence, reliable sensitivity (∂f2/∂f1) and elastic
that the points obtained by using the NBI method rate (∂ ln f2/∂ ln f1) can be approximated from the
produce a more even coverage of the Pareto curve and optimization results without Pareto curve fitting. For
this coverage does not miss the interesting middle part every Pareto optimal point, there exists a corresponding
of the Pareto curve. The goal programming method NBI subproblem of which it is the solution. Most NBI
constrains a parametrized value of objective functions points are guaranteed to be only locally Pareto optimal
other than the main objective function, and the SWOF points even though the Pareto curve is nonconvex.7
method uses slope values [(R - 1)/R] to obtain a set of
optimum results. In contrast, the NBI method finds The Pareto curve is analyzed by two ideal compromise
intersection points of the Pareto curve with normal points at δ1 and δ∞, for sensitivity and elasticity
directions toward the origin, emanating from any points analyses for the basis of decision-making, as mentioned
in the study of Lim et al. (1999).2 It is worth emphasiz-
of the segment PQ that connects two respective opti- ing that the Pareto points of δ1 and δ∞ will provide the
mization points (P and Q), as shown in Figure 1. In the lower and upper bounds of the ideal compromise solu-
original NBI method, an equality constraint containing tions,10 if these exist.
one parameter (β) is added to the optimization problem
2-2. Implementation of the MOOP Algorithm
(Das and Dennis, 1998).7 In the case of a biobjective
within a Chemical Process Simulator. The MOOP
problem, employing the normalization of two objective
functions (see Appendix A), approach is achieved by substituting the MOOP module
for the optimization module in the process simulator.
Figure 2 shows the MOOP algorithm using the SQP
min f̂2 optimization solver.11 The objective functions are first
x∈Ω
normalized by the SWOF method with the weighting
s.t. f̂2 ) f̂1 + (1 - 2β), 0 e β e 1.0 (2) parameters R ) 0.0 and R ) 1.0. Then, if possible, the
SWOF method at R ) 0.5 is performed to search for an
the parameter β normally is a value of f̂1 determined ideal compromise solution of δ1. In the case of β ) 0.5
by the number (n) of wanted Pareto points, as follows: in the NBI method, another ideal compromise point of
δ∞ is found. Depending on the parameters (β) of the
modified NBI method using eqs 4 and 5, Pareto points
βi ) i/n, i ) 0, 1, 2, ..., n (3)
are found to obtain the entire Pareto curve. Finally, one
can make a tradeoff between the conflicting
A problem at β0 ) 0.0 corresponds to minimization of objectives in order to understand the process charac-
f̂1 without consideration of f̂2. At βn ) 1.0, it is equivalent teristics and to determine new operating conditions
to the problem of R ) 1 in the SWOF method. When through Pareto curve analysis using ideal compromise
βn/2 ) 0.5, an ideal compromise point at norm δ∞ is solutions as well as the sensitivity and elasticity of the
obtained, i.e., f̂1 ) f̂2 ≡ µ. Pareto curve.
650 Ind. Eng. Chem. Res., Vol. 40, No. 2, 2001
Figure 3. Comparison of the three MOOP methods at even parameter spread in the small biobjective problem.
x∑
Curve. The test example is n
1
(sj - ∆si)2
[ ]
η) (7)
n
x21
f1(x) ) + + + +x22 x23 x24 x25 i)1
min x3 3
(6) where sj denotes the mean arc length. Thus, as the
x f2(x) ) 3x1 + 2x2 - + 0.001(x4 - x5) equidistribution index approaches zero, the points ob-
3
tained are more uniformly distributed for the same
s.t. x1 + 2x2 - x3 - 0.5x4 + x5 - 2.0 ) 0.0 number of parameters (R, , and β for the SWOF, GP,
and NBI methods, respectively). Using an even spread
4x1 - 2x2 + 0.8x3 + 0.6x4 + 0.5x25 ) 0.0 of parameters (with 21 points at 0.05 step size, 0.0 e R,
, or β e 1.0), three Pareto curves are shown in Figure
x21 + x22 + x23 + x24 + x25 - 10.0 e 0.0 3. Note that Pareto points obtained by the SWOF
method in this example are clustered near the minimum
point of f1, and the middle part for the ideal compromise
First of all, the three methods such as SWOF, GP and zone is not well captured. In contrast, the GP method
the original NBI are compared by their numerical provides poor information in the steep regions of the
results for this biobjective problem. For the solution Pareto curve. Employing the NBI method with even
quality, a positive equidistribution index (η) of Pareto spread of parameters, more uniformly distributed Pare-
Ind. Eng. Chem. Res., Vol. 40, No. 2, 2001 651
( )
are obtained directly from optimization results without rA ) S (10)
PA
Pareto curve fitting, are reliable because of the evenly PK 1 + KAPA + KAK
distributed Pareto points. PK
3-2. A Biobjective Problem for a Nonconvex
Pareto Curve. A nonconvex biobjective problem pre- The notation is explained in the nomenclature, and the
sented by Li et al. (1999)12 is described as constants such as C, K, KA, and KAK are calculated from
[ ]
the relation equations with the reaction temperature,
f1 ) 1 - x1x2 which were developed by Perona and Thodos (1957).13
min 1 (8) The catalyst surface area (S) is assumed to be equal to
x f2 ) 3(2x1 + x2) 2500 m2.
In Figure 8, the MEK process flowsheet2,8,14 is shown.
s.t. 0.0 e x1, x2 e 1.0 The SBA feed containing 0.1 mol % water is mixed at a
fixed feedrate (45.8128 kmol/h at 15 °C and 1 atm) with
It is easy to verify that this example problem is a recycle stream from the purification column. The
nonconvex because the Hessian matrix of the objective combined liquid stream is fed to a hydrogen scrubber
function f1 is negative definite. Applying the SWOF where the SBA scrubs residual MEK from the hydrogen
Figure 5. Sensitivity and elasticity analyses of the Pareto curve in the small biobjective problem.
652 Ind. Eng. Chem. Res., Vol. 40, No. 2, 2001
Figure 6. Comparison of the three MOOP methods in the case of the nonconvex Pareto curve.
(
fU
)
L
2 - f 2
γ) ) (56.525 - 52.138)/(71.446 -
fU L
1 - f 1
57.397) ) 0.31225 (11)
Figure 10. Profiles of four independent variables with respect to the pollution function values. (a) Dehydration column reflux ratio (R1).
(b) Purification column reflux ratio (R2). (c) Ratio of liquid distillate (RLD). (d) Reaction temperature.
4. Conclusions
Within a chemical process simulator, a new MOOP
algorithm (normalization, the combined use of SWOF
and NBI, and finally Pareto curve analysis) is employed
in terms of economics and potential environmental
impact through the MEK process. The modified NBI
method gives evenly distributed Pareto points and
reliable optimization results even in a nonconvex feature
of the Pareto curve.
Decision information for making a tradeoff between
profitability and pollution is obtained by performing
sensitivity and elasticity analyses of the Pareto curve.
Figure 11. Sensitivity and elasticity analyses of the Pareto curve. It is worth noting that the sensitivity of the Pareto curve
tion function (R ) 0.0), and traditional operating condi- is a marginal cost, that is, the cost change over a unit
tions of the minimum cost-benefit function (R ) 1.0). change of environmental impact. The elasticity of the
Ind. Eng. Chem. Res., Vol. 40, No. 2, 2001 655
Pareto curve means the ratio of cost change rates to a The point (f L1 , f L2 ) can be defined in the objective
unit rate of the environmental impact change. space as the utopia point. The utopia point (or shadow
The ideal compromise solution set between δ1 and δ∞ minimum) will generally be infeasible because it will
on the Pareto curve is thought of as a mild condition to lie outside the attainable objectives (Λ, inside the
simultaneously satisfy two objectives. The proposed noninferior solution curve) given by
MOOP algorithm has exactly identified its upper and
lower values. Λ ) {(f1, f2) |x ∈ Ω}
Acknowledgment
Also, the following normalization can be employed:
We thank the Ministry of Foreign Affairs of France
and the Embassy of France in Korea, which support our (f1 - f L1 ) (f2 - f L2 )
research in the framework of France-Korea collabora- f̂1 ) , f̂ 2 )
tion. fU1 - f 1
L
fU2 - f 2
L