You are on page 1of 13

Desalination 184 (2005) 259–271

Global optimal design of reverse osmosis networks for


seawater desalination: modeling and algorithm

Marian G. Marcovecchio, Pı́o A. Aguirre, Nicolás J. Scenna*


INGAR–Instituto de Desarrollo y Diseño, Avellaneda 3657, S 3000 GLN Santa Fe, Argentina
Tel. þ54 342 4534451; Fax þ54 342 4553439; email: mariangm@ceride.gov.ar, mussati@ceride.gov.ar,
paguir@ceride.gov.ar, nscenna@ceride.gov.ar

Received February 23, 2005; accepted March 20, 2005

Abstract
A novel global optimization algorithm to solve nonconvex problem is used to find the global optimal design
of reverse osmosis networks for seawater desalination. The objective is to determine the optimal process design
and operating conditions for a given water production. The networks were designed by using hollow fiber
reverse osmosis modules. The Kimura–Sourirajan model was used for describing transport phenomena of solute
and water transport through the membrane. The concentration polarization phenomenon has been taken into
account. It was mathematically described using the film theory. The objective function to be minimized is the
cost, which includes capital investment (membrane cost, pumping and energy recovery system, intake and pre-
treatment systems, etc.) and operation and maintenance costs (membrane replacement, chemical treatment,
spares, required and recovered energy, etc.). The proposed algorithm is deterministic and attains finite con-
vergence to the global optimum. It is iterative and a main problem is solved each iteration. The main problem
has convex constraints and a nonconvex objective function. The main problem solution indicates either a better
solution for the original problem, or a region which can be discarded. Therefore, the feasible region to improve
the objective function is reduced each iteration. The algorithm finishes when the whole region has been analysed
and discarded. A bound reduction technique is performed in order to accelerate the convergence speed. The
algorithm shows a good performance and efficient execution time. Different cases are solved in order to show
the methodology and computational performance.

1. Introduction
The scarcity of natural fresh water supplies
*Corresponding author. is an important problem in several countries.

Presented at the Conference on Desalination and the Environment, Santa Margherita, Italy, 22–26 May 2005.
European Desalination Society.
0011-9164/05/$– See front matter Ó 2005 Elsevier B.V. All rights reserved
doi:10.1016/j.desal.2005.03.056
260 M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271

The natural sources are diminishing, but not Two-stage SWRO desalting plant
RO Permeators-1º Stage
the demand. This problem is solved by pro- SWIP
HPP
Qf,1 NM1
ducing fresh water from seawater. Several Qp,1 NM1

systems are available for this propose. Qf,1 NM1

The MSF process has been traditionally


Qp,1 NM1+ Qp,2 NM2
used to desalt seawater. However, the RO pro- Qb,1 NM1 RO Permeators-2º Stage

cess is becoming most popular. During the last


years hundreds of reverse osmosis sweater desa- Qf,2 NM2 Qp,2 NM2

linations plants have been built. That is because


RO systems are able to produce fresh water Qb,2 NM2

with lower cost. RO has low consumption of


energy when it is compared to thermal separa- Fig. 1. Two-stages SWRO desalting plant.
tions. Also, in the last decades, significant
improvements in the quality membrane have DuPont’s B10 hollow fiber reverse osmosis
been made. Now, RO systems are able to pro- modules were used in the formulation.
duce a permeated stream with lower salt con- A detailed cost function is formulated,
centration. This fact has straight consequences including all the most important factors
over the total cost. affecting the price of the fresh water.
An important effort has been made in obtain The problem has been formulated in such
accuracy models for the transport phenomena way that the design with only one reverse osmo-
across the membrane [1–6]. However, there are sis stage is included. Fig. 2 represents this case.
less studies addressing the problem of designing The cost function has been also formulated to
reverse osmosis networks including reliable provide the exact cost in this last case.
models to predict the transport across the mem- The RO system is made of five major
brane [7–9]. Efficient design has been found parts: The seawater intake and pre-treatment
employing different resolution methods. Never- system (SWIP), the high pressure pumps
theless, global optimization has not been used (HPP), the first and the second stage of
to solve this kind of problems. reverse osmosis, and the energy recovery sys-
The main scopes of this work are two: to tem (ERS).
formulate a detailed optimization problem The seawater feed passes through the
for the design of reverse osmosis networks SWIP where is chemically treated. Then, the
including an accuracy model for the transport
phenomena across the membrane and a com-
plete cost function; and to solve this problem Single-stage SWRO desalting plant
for global optimization. RO Permeators
HPP
Different cases varying the feed concentra- SWIP Qf NM
tion are solved in order to study their impact Qp NM

over the design and the cost. Qf NM

2. Mathematical model
Qb NM
A framework including two reverse osmo- ERS
sis stages is modeled. Fig. 1 represents the
network modeled in the present work. Fig. 2. Single-stage SWRO desalting plant.
M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271 261

pressure is raised by the high pressure pumps. 2.1. Reverse osmosis model
The high pressure stream enters to the per- Table 1 contains the equations describing
meators of the first stage of reverse osmosis. the reverse osmosis modeling.
The totality or part of the rejected stream is The Kimura–Sourirajan model [10] was
introduced in the permeators of the second used to describe the transport phenomena of
stage of reverse osmosis. The permeated solute and water through each membrane.
streams from both stages are blended, requir- This model is the most used for this propose,
ing a maximum total salt concentration. The because it is able to provide an accuracy pre-
brine flow rate remainder of the first stage (if diction of the flow of fresh water and salt
it does exist) and the rejected stream from the through the membrane at standard condi-
second stage pass through the energy recov- tions. Two membrane parameters are
ery system, and then they are discarded. required: the pure water permeability con-
Therefore, part of the power for the stant (A) and the solute permeability constant
pumps is supplied by the energy recovery (B). Both parameters are known for the
system, and the rest is provided by an exter- DuPont’s B10 hollow fiber permeator.
nal source.

Table 1
Equations describing the reverse osmosis modeling
 
b p iRTb ðCsm Csp Þ
(1) Jsw ¼ 3600A Ps  Ps  106 Ms 101325 s = 1,2
3600Bð Csm Csp Þb
(2) JsS ¼ 106
s = 1,2
ðJsw þJsS Þ
(3) Vsw ¼ p s = 1,2
JsS 106
(4) Csp ¼ VsW p
s = 1,2
(5) Qps ¼ Vsw Am s = 1,2
(6) Qfs ¼ Qbs þ Qps s = 1,2
(7) Qfs Csf ¼ Qbs Csb þ Qps Csp s = 1,2
(8) Shs ¼ ksD2ro s = 1,2
2r U S b
(9) Res ¼ obs s = 1,2
b
(10) Scs ¼ b D s = 1,2
(11) Shs ¼ 2:725ðRes Þ1=3
ðScs Þ
1=3
s = 1,2
Csm Csp Vsw
(12) Csb Csp
¼ exp 3600k s
s = 1,2
Qfs
(13) Ussi ¼ 36002R iL
s = 1,2
Qfs Vsw Am
(14) Usso ¼ 36002Ro L s = 1,2
Ussi Usso
(15) UsS ¼    s = 1,2
log Ussi Usso
p 16p r Vsw L2
(16) Ps ¼ 1 þ 12 3600r4o101325 s = 1,2
i
2
b 150ð1"Þ  2 b
UsS 1:75ð1"Þ ð
b
UsS Þ
(17) Ps ¼ Pfs  12 "3 dp2 101325
ðRo  Ri Þ  12 "3 dp 101325
ð Ro  Ri Þ s = 1,2
262 M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271

Eqs. 1 to 17 describe the process happen- superficial velocity is approximated as the


ing in one membrane of each stage of reverse log mean average of both velocities, Eq. 15.
osmosis.
The water and salt flux trough the mem-
2.2. Relationship between the two reverse
brane according to the Kimura–Sourirajan
osmosis stages
model are given as Eqs. 1 and 2. The osmotic
pressure was calculated using the van’t Hoff Table 2 contains the equations relating
equation [5]. both stages.
The permeation velocity is given by Eq. 3. As it is shown in Fig. 1, the permeated
Eq. 4 gives the permeate concentration in streams are blended. Then, the sum of both
ppm, and Eq. 5 the permeate flow rate. Eqs. permeated streams must be equal to the
6 and 7 give the mass and salt balances, desired total production, Eq. 18.
respectively. Eq. 19 requires that the blended produc-
The concentration polarization is another tion satisfy the maximum allowed salt
important phenomenon to be taken into concentration.
account. This phenomenon gets down the Eqs. 20 and 21 establish that the feed
fresh water flow, and must be considered. pressure and the feed salt concentration of
The film theory has been widely used to the second stage be the pressure and the con-
describe the concentration polarization and centration of the brine flow rate exiting at the
it is also used here. first stage, respectively.
For this propose, it is necessary to esti- Eq. 22 constraints the total flow rate
mate the mass transfer coefficient. It was cal- entering to the second stage to be less or
culated using the relation between the equal to the brine flow rate exiting at the
Sherwood, Reynolds and Schmidt numbers, first stage.
Eqs. 8, 9, 10 and 11. The values for the The total seawater feed enters to the first
coefficient and exponent of Reynold’s num- stage, Eq. 23.
ber were taken from [2]. According to the The variable Pout is the pressure used to
film theory, Eq. 12 models the concentration calculate the capital and operating costs of
polarization. the energy recovery system. And Eqs. 24
In order to estimate an average of the drop and 25 constraint this variable to the brine
pressure in the fiber bore and an average of flow rate pressure exiting the first and the
the drop pressure on the shell side of the fiber second stages.
bundle, the methodology provided in [2] is
used. It is based on the Hagen–Poiseuille
equation for the drop pressure in the fiber Table 2
bore and the Ergun equation for the drop Equations relating the two reverse osmosis stages
pressure on the shell side of the fiber bundle,
(18) prodc ¼ Qp1 NM1 þ Qp2 NM2
Eqs. 16 and 17.
(19) Cmax prodc  NM1 Qp1 C1p þ NM2 Qp2 C2p
It is necessary to estimate the superficial (20)
b
Pf2 ¼ 2  P1  Pf1
velocity in the radial direction. It is carried (21) f
C2 ¼ C1 b

out by calculating the superficial velocity at (22) NM1 Qb1  NM2 Qf2
the inner radius of the fiber bundle, Eq. 13 (23) Qfeed ¼ NM1 Qf1
b
and the superficial velocity at the outer radius (24) Pout  2  P2  Pf2
f
(25) Pout  P2
of the fiber bundle, Eq. 14. Finally, the
M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271 263

The optimization procedure will decide the exactly cost in both cases: one or two reverse
best option between one or two stages of reverse osmosis stage.
osmosis. In case that one stage of reverse osmo- Table 3 shows the cost equations.
sis is chosen, the second stage will be eliminated. Capital and Operating costs are calculated.
This option will be indicated through the vari- Eqs. 26 to 30 refer to capital cost.
able NM2: number of membranes in the second The cost functions reported in [7] were
stage. If NM2 = 0, just the first osmosis reverse adopted to calculate the capital cost for the
stage is necessary for optimality. Else, SWIP, Eq. 26 and the HPP, Eq. 27. The cost
(NM2 > 0) both stages are necessary. for the ERS is similar to the HHP one, since
In the last case, the flow rate entering to it consists of a reverse running centrifugal
the second stage can be the totality or just pump, Eq. 28.
part of the brine flow rate exiting the first Eq. 29 establishes the capital cost for the
stage, as it is indicated at Fig. 1. This fraction permeators.
will be determined by the optimization proce- Then, the plant equipment cost is:
dure, according to the best cost condition. cceq = ccswip þ cchpp þ ccers þ ccprm.
The remaining brine flow rate will pass The cost of civil work is estimated as a
through the energy recovery system. 10% of cceq [11]. The indirect cost is esti-
mated in the same way [9]. Eq. 30 gives the
total capital cost.
2.3. Cost functions Eqs. 31 to 37 refer to annual operating cost.
Each one of the five major parts of the The capital charge cost is given by Eq. 31 [7].
network has been taken into account for the The permeator replacement cost is calcu-
cost function. The cost functions have been lated by taking a RO membrane replacement
formulated in such way that they give the of 20% per year, Eq. 32.

Table 3
Description of the cost equations
(26) ccswip  996ðQfeed 24Þ0:8
 
(27) cchpp ¼ Q450
feed
393; 000 þ 10; 710Pf1 1:01325
(28) ccers ¼ ðQfeed450
prodcÞ
ð393; 000 þ 10; 710  Pout  1:01325Þ
(29) ccprm = (NM1 þ NM2)cmodulo
(30) TCC = 1.2 (ccswip þ cchpp þ ccers þ ccprm)
(31) coc = 0.08 TCC
(32) corp = (NM1 þ NM2)0.2 cmodulo
 
P Qfeed 24 Pf1 1:01325Qfeed 24
(33) coe ¼ 0:03fc swip
effswip þ effhpp  effers Pout  1:01325ð Qfeed  prodc Þ24
(34) cos = prodc 24365fc0.033
(35) coch = Qfeed24365fc0.018
(36) coom = prodc24365fc0.126
(37) AOC = coc þ corp þ coe þ cos þ coch þ coom
TCC factorþAOC
(38) cost ¼ prodc24365
264 M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271

The energy cost is calculated by using the ybk ¼ xi1 ðkÞ  xi2 ðkÞ k ¼ 1; . . . ; 20
cost function in [7] and the power cost
reported in [11], Eq. 33. where the vector x  represents the original vari-
The spares cost is calculated using the ables and the variables added for the reformula-
value reported in [11], Eq. 34. tion. The functions: fj(xj) are univariate concave.
The chemical treatment cost is estimated using There are 13 univariate concave terms and 20
the cost per m3 of feed reported in [9], Eq. 35. bilinear terms in the reformulation.
The operation and maintenance costs are Now, the main problem is defined as follows
calculated using the cost per m3 of produc-
tion reported in [11], Eq. 36. X
13  2
Then, the annual operating cost is given by MP : min Fse ¼ fj ðxj Þ  ycj
j¼1
Eq. 37.
Based on a discount rate of i = 8% and a X
20  2
plant life of n = 25 years, the capital recovery þ xi1 ðkÞ xi2 ðkÞ  ybk
k¼1
factor (crf) is: crf = [(i þ 1)n  1]/[i(i þ 1)n].
Then, the fresh water cost per m3 is given by Eq. 38.
s.t.

3. Optimization problem cost  costUP  "


3.1. Algorithm
Eqs. 39 to 84
The problem P to be solved is:
 
P: minimize cost satisfying Eqs. 1 to 38, while
ycj  fj xLO
all the variables have appropriated bounds.  j   
First, a new problem: RP is generated, by fj xUP
j  fj xLO
j
 
adding variables in order to simplify the þ   xj  xLO
j
terms involved in the constraints. Only two xUP
j  xLO
j
kinds of nonconvexity terms are allowed for j ¼ 1; . . . ; 13
RP: univariate concave terms and bilinear
terms. The problem RP must be completely ybk  xLO LO LO LO
i1 ðkÞ  xi2 ðkÞ þ xi1 ðkÞ  xi2 ðkÞ  xi1 ðkÞ  xi2 ðkÞ
equivalent to the problem P. Table 4 shows
this reformulation for the constraints k ¼ 1; . . . ; 20
involved in problem P.
ybk  xUP UP UP UP
i1 ðkÞ  xi2 ðkÞ þ xi1 ðkÞ  xi2 ðkÞ  xi1 ðkÞ  xi2 ðkÞ
Therefore, the problem RP is
RP: minimize cost satisfying Eqs. 39 to 103. k ¼ 1; . . . ; 20
RP is equivalent to the problem P, that is:
each feasible solution to RP has a correspond- ybk  xUP LO UP LO
i1 ðkÞ  xi2 ðkÞ þ xi1 ðkÞ  xi2 ðkÞ  xi1 ðkÞ  xi2 ðkÞ
ing feasible solution to P and vice versa. k ¼ 1; . . . ; 20
The objective function and constraints in
problem RP are convex, except for those con- ybk  xLO UP LO UP
i1 ðkÞ  xi2 ðkÞ þ xi1 ðkÞ  xi2 ðkÞ  xi1 ðkÞ  xi2 ðkÞ
straints defining the variables ycj and ybk.
k ¼ 1; . . . ; 20
These constraints have the form:
  The nonconvex constraints in RP were
ycj ¼ fj xj j ¼ 1; . . . ; 13
changed by linear sub-estimations. For
M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271 265

Table 4
Reformulation for the constraints involved in problem P
 
iRTb ðCm Cp Þ s = 1,2
(39) Jw ¼ 3600AðP  Þ ¼ 3600A Pb  Pp  106 Ms 101;325
3600BðCsm Csp Þb s = 1,2
(40) JsS ¼ 106
w ðJsw þJsS Þ s = 1,2
(41) Vs ¼  p
(42) yb1s p ¼ JsS 106 s = 1,2
(43) Qps ¼ Vsw Am s = 1,2
(44) Qfs ¼ Qbs þ Qps s = 1,2
(45) yb2s ¼ yb3s þ yb4s s = 1,2
(46) Shs ¼ ksD2ro s = 1,2
2r U S b s = 1,2
(47) Res ¼ obs
b
(48) Scs ¼ b D s = 1,2
(49) Shs  2:725ðRes Þ1=3 ðScs Þ1=3  0 s = 1,2
(50) Shs þ 2:725 ycs1 ðScs Þ1=3  0 s = 1,2
(51) w1s ¼ Csm  Csp s = 1,2
(52) w2s ¼ Csb  Csp s = 1,2
(53) w3s  ln w1s þ yc2s  0 s = 1,2
 
(54) w3s þ yc3s  ln w2s  0 s = 1,2
(55) 3600yb5s ¼ Vsw s = 1,2
Qfs s = 1,2
(56) Ussi ¼ 36002R iL
Qfs Vsw Am s = 1,2
(57) Usso ¼ 36002R
  oL
(58) w4s  ln Ussi þ yc4s  0 s = 1,2
 
(59) w4s þ yc5s  ln Usso  0 s = 1,2
(60) yb6s ¼ Ussi  Usso s = 1,2
p 16p ro Vsw L2 s = 1,2
(61) Ps ¼ 1 þ 12  3600r4 101;325
i
2
b 150ð1"Þ2 b UsS 1:75ð1"Þb ðUsS Þ s = 1,2
(62) Ps  Pfs þ 12  "3 dp2 101;325
ðRo  Ri Þ þ 12 "3 dp 101;325
ðRo  Ri Þ  0
b 150ð1"Þ2 b UsS 1:75ð1"Þb yc6s s = 1,2
(63) Ps þ Pfs  12 "3 dp2 101;325
ðRo  Ri Þ þ 12 "3 dp 101;325
ð Ro  Ri Þ  0
(64) prodc ¼ yb71 þ yb72
(65) Cmax prodc  yb81 þ yb82
b
(66) Pf2 ¼ 2  P1  Pf1
f b
(67) C2 ¼ C1
9 10
(68) yb  yb
(69) Qfeed ¼ yb71 þ yb9
b
(70) Pout  2  P2  Pf2
(71) Pout  Pf2
(72) ccswip  996  240:8 yc7
11
393;000
(73) cchpp ¼ Qfeed 450 þ 10;7101:01325yb
450
 
(74) ccers ¼ ðQfeed prodcÞ393;000
450 þ 10;710
450 1:01325yb12  prodc  Pout  1:01325
(75) ccprm = (NM1 þ NM2)cprm
266 M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271

(76) TCC = 1.2(ccswip þ cchpp þ ccers þ ccprm)


(77) coc = 0.08TCC
(78) corp = (NM1 þ NM2)0.2cprm
P Qfeed 24 11   
(79) coe ¼ 0:03fc swip
effswip þ yb 1:0132524
effhpp  effers 1:01325 prodcPout  yb12 24
(80) cos = prodc24365fc0.033
(81) coch = Qfeed24365fc0.018
(82) coom = prodc24365 fc 0.126
(83) AOC = coc þ corp þ coe þ cos þ coch þ coom
TCC factorþAOC
(84) cost ¼ prodc24365
(85) yb1s ¼ Csp  Vsw s = 1,2
(86) yb2s ¼ Qfs  Csf s = 1,2
(87) yb3s ¼ Qps  Csp s = 1,2
(88) yb4s ¼ Qbs  Csb s = 1,2
(89) yb5s ¼ ks  w3s s = 1,2
(90) yb6s ¼ UsS  w4s s = 1,2
(91) yb7s ¼ Qps  NMs s = 1,2
(92) yb8s ¼ Csp  yb7s s = 1,2
(93) yb9 ¼ NM1  Qb1
(94) yb10 ¼ NM2  Qf2
(95) yb11 ¼ Qfeed  Pf1
(96) yb11 ¼ Qfeed  Pout
1=3 s = 1,2
(97) yc1s ¼ ðRe
 sÞ 
(98) yc2s ¼ ln w2s s = 1,2
 
(99) yc3s ¼ ln w1s s = 1,2
 
(100) yc4s ¼ ln Usso s = 1,2
 
(101) yc5s ¼ ln Ussi s = 1,2
 2
(102) yc6s ¼  UsS s = 1,2
(103) yc7 ¼ ðQfeed Þ0:8

each univariate concave function, the sub- However, the objective function in MP: Fse
estimation is the straight line joining the is nonconvex.
extreme points. For the bilinear term, the Let suppose that the problem MP is solved
sub-estimations are the Mc Cormick’s linear with a local solver. Three cases are possible:
constraints. 1) An optimal point with objective: Fse = 0 is
The constant costUP is an upper bound obtained. The point is a feasible point to the
for the objective. It can be any high problem P, and cost  costUP  ", so it is a
enough number. And " is the error allowed better solution.
for the global optimality. An acceptable 2) An infeasibility is obtained. Based on the fact
value for " is the 1–5% of the optimal that the constraints defining MP are convex, the
objective value. infeasibility of MP can be insured. Since the fea-
All the constraints of the problem MP are sible region in MP is an over-estimation of the
convex, so they define a convex set. feasible region in RP, each feasible point to RP is
M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271 267
n  
a feasible point to MP. Therefore, it can be con-  Tbk ¼ x; yc; yb; cost : ybk > ybk þ xi1 ðkÞ
cluded that there is not solution to RP having    o
objective value less or equal to (costUP  "). The xi2 ðkÞ  xi2 ðkÞ þ xi2 ðkÞ xi1 ðkÞ  xi1 ðkÞ
same conclusion is obtained for the P problem,
since RP and P are equivalent. for each k 2 K þ :
3) An optimal point with objective: Fse > 0 n  
is obtained. The point is infeasible to the  Tbk ¼ x; yc; yb; cost : ybk < ybk þ
problem RP. However, since the objective    o
function in MP is nonconvex, this point xi1 ðkÞ xi2 ðkÞ  xi2 ðkÞ þ xi2 ðkÞ xi1 ðkÞ  xi1 ðkÞ
could be a local optimum. Therefore, it is for each k 2 K  :
not possible to insure that the problem RP
is infeasible in the whole region (requiring
If a point p is such that:
that cost  costUP   ").   T T  T 
Let be p ¼ x ; yc ; yb ; cost the (local) p 2 ð j2J Tcj Þ \ k2K þ Tck \ k2K  Tck ,

optimum
 to MP with 2Fse > 0. then p does not satisfy (*). So, the point
 j0 must be infeasible to problem MP.
If fj0 ðxj0 Þ  yc for some j0, it 
> 0,  is pos-
 Therefore, p cannot be a feasible point to
sible to show that must be: fj0 ðxj0 Þ  ycj0 > 0.
problem P. T 
On the other hand, if Then, the region D ¼ Tc
 2 T  T  j2J j
k0
xi1 ðk0 Þ xi2 ðk0 Þ  yb
> 0 for some k0, both \ k2K þ Tck \ k2K  Tck can be dis-
 

signs are possible for xi1 ðk0 Þ xi2 ðk0 Þ  ybk0 . carded, since there is not solution to the pro-
blem P there.
It is possible to show that being p either a
Therefore, the region to be considered is
local optimum or a KKT point to MP, the
Dc. However, it would be better to continue
following inequality must be truth:
the search considering the region R defined
by the points in Dc such that
rFse ðp Þt  ðp  p Þ  0
ycj ¼ fj ðxj Þðj 2 J Þ, and ybk ¼ xi1 ðkÞ xi2 ðkÞ ,
8p feasible solution to MP ðÞ respectively ðk 2 K þ and k 2 K  Þ.
The algorithm employed here, continues
Let be: the search in a region R0 which contains
 2 R, instead of the proper R. That approach
J¼ 1  j  13 : fj ðxj Þ  ycj > 0 is used in order to simplify the following
steps.
n  
 o Several regions are generated in order to
K þ ¼ 1  k  20 : xi1 ðkÞ xi2 ðkÞ  ybk > 0 describe R0 . The number of such regions is
bounded by 60 in the present problem.
n   o

K ¼ 1  k  20 : xi1 ðkÞ xi2 ðkÞ k
 yb <0 Once, the regions have been generated,
the procedure is repeated in each new sub-region.
n    Bound reduction techniques were applied
 Tcj ¼ x; yc; yb; cost : ycj > ycj þ fj0 xj in order to accelerate the convergence.
 o Also, the variables were scaled to avoid
xj  xj for each j 2 J : precision errors.
268 M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271

3.2. Basic steps of the algorithm Table 5


Step 1: Initialization: Let be " = 0.001 The most important parameters of the NLP problem
and costUP = 1.5. Let be P = {Ro}: the i, number de ions for ionized solutes 2
hyperrectangle defined by the bounds in RP. R, ideal gas constant 8315
Set iter = 1 and Ri = Ro. (Nm/kg mole  K)
Step 2: Solution of the problem MP: Solve Ms, molecular weight of solute 58.8
the MP problem in region Ri with a local T, seawater temperature ( C) 25
Pb, average brine density (kg/m3) 1060
solver. p, average pure water 1000
If MP is infeasible, set P = P\{Ri}. Go to density (kg/m3)
Step 4.  
p
 , permeate viscosity (kg/m s) 0.9  103

Otherwise, let p ¼ x ; yc ; yb ; cost be b, brine viscosity (kg/m s) 1.09  103
the optimum. D, diffusion coefficient (m2/s) 1  109
Pswip, SWIP outlet pressure (bar) 5
If Fse = 0, update: pOPT ¼ p , costUP =
effslip, intake pump efficiency 0.74
cost* and go to Step 5, else go to Step 3. effhpp, high pressure pumps 0.74
Step 3: Generating R0 : Generate new sub- efficiency
regions describing R0 and place them on the effers, energy recovery system 0.80
list P. Set P = P\{Ri}. Go to Step 4. efficiency
Step 4: Stop criterion: If P = Ø, stop: the fc, load factor 0.90
cprm, permeator cost (estimation) 10 Am
current best solution is the global optimal. Other-
wise, set iter = iter þ 1 and choose Ri from P.
Return to Step 2.
Several cases were solved, varying the sea-
4. Solutions water salt concentration. The total produc-
The algorithm was implemented in GAMS tion required is 2000 m3/h, and the
[12] and GAMS/CONOPT solver was used to maximum allowed salt concentration is
solve the NLP problems. 570 ppm.
Table 5 contains the most important para- In general, the algorithm takes between
meters of the problem and Table 6 contains 709 and 3402 iterations and 194 and 957
the DuPont’s B10 hollow fiber parameters. CPUs to finish.

Table 6
DuPont’s B10 hollow fiber parameters

A, pure water permeability constant (kg/m2satm) 4.294  1010101325


B, salt permeability constant (m/s) 4.0109
Ri, inner radius of the fiber bundle (m) 1.27102
Ro, outer radius of the fiber bundle (m) 5.334102
L, length of fiber bundle (m) 0.75
Am, membrane area (m2) 152
ri, inner fiber radius (m) 2.1105
ro, outer fiber radius (m) 5.0105
", void fraction 0.4
dp, specific surface diameter (m) 1.2104
Operating pressure range, bar 55.2–68.8
Maximum brine flow rate (m3/h) 0.917
M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271 269

From 1 to 5 local optimums were found were obtained using the model and the algo-
before the global solution in each case. rithm proposed in this work.
Table 7 show the main variable values of Also, the optimal operating conditions for
the global optimal solutions for different sea- each case were determined.
water salt concentrations. The model includes all the most important
aspects of the reverse osmosis process, as
drop pressure and concentration polarization.
5. Conclusion
Furthermore, a detailed objective cost was
The framework with only one reverse osmosis formulated. All the main parts of the network
stage was optimal in the case with highest sea- were taken into account for the cost function:
water salt concentration. Also, this case repre- capital investment (membrane cost, pumping
sents the most expensive cost for the fresh and energy recovery system, intake and pre-
water, as it can be predicted. treatment systems, etc.) and operation and
For the cases solved with seawater salt con- maintenance costs (membrane replacement,
centration between 41,000 and 37,000 ppm, chemical treatment, spares, required and
both stages were required for the optimal con- recovered energy, etc.).
ditions. According the feed salt concentration Both: the equations relating the stages
diminishes, the percentage of brine entering to and the cost functions allow the model to
stage 2 grows. Until the last case (lowest feed choose between one or two reverse osmosis
salt concentration) in which the whole brine of stages.
the first stage enters to the stage 2. The formulated problem is highly noncon-
Obviously, the cost of the fresh water vex. As it was mentioned, several local opti-
diminishes while the seawater salt concentra- mums have been found. Therefore, the
tion diminishes. solution obtained whit any local solver is
The global optimal designs of reverse limited to local guarantee and could be sub-
osmosis networks for seawater desalination stantially different to the global one.

Table 7
Global optimal solutions for different seawater salt concentrations

Seawater salt concentration: C1f 42,000 41,000 40,000 39,000 38,000 37,000 36,000
Operating pressure: Pf1 68.314 66.654 66.943 67.297 67.670 68.042 68.395
Number of permeators, stage 1: NM1 8001 7539 6907 6347 5845 5370 4976
Number of permeators, stage 2: NM2 0 916 1511 2027 2483 2912 3209
Feed flow rate stage1: NM1 Qf1 6696.7 6493.6 5949.0 5467.1 5034.2 4625.2 4285.8
Feed flow rate stage2: NM2 Qf2 — 788.7 1301.4 1745.8 2138.6 2508.1 2764.1
Permeated flow rate stage 1: NM1 Qp1 2000 1887.0 1806.8 1730.2 1656.0 1581.1 1521.8
Permeated flow rate stage 2: NM2 Qp2 — 113.0 193.2 269.8 344.0 418.9 478.2
Brine flow rate stage 1: NM1 Qb1 4696.7 4606.2 4142.3 3736.3 3378.4 3043.7 2764.1
Brine flow rate stage 2: NM2 Qb2 — 675.8 1108.3 1476.7 1794.8 2089.3 2285.9
Percentage of brine entering to stage 2 0 17.12 31.42 46.7 63.3 82.4 100
Permeated concentration, stage 1: Cp1 548.5 530.0 503.1 479.6 458.9 438.9 419.7
Permeated concentration, stage 2: Cp2 — 1237.8 1196.0 1151.3 1104.9 1065.4 1028.7
Cost of fresh water per m3 $/m3 1.0108 1.0028 0.9536 0.9095 0.8695 0.8313 0.7974
Total production: 2000 m3/d. Maximum allowed salt concentration: 570 ppm.
270 M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271

In the present work, a novel global opti- cci — Indirect capital cost ($)
mization algorithm to solve nonconvex pro- ccprm — Permeator capital cost ($)
blem was used to find the global solution of ccswip — Seawater intake and pre-treatment
the problem. The algorithm is deterministic, capital cost ($)
so the global optimality is guaranteed. Differ- cmax — Maximum allowed salt concentra-
ent cases are solved showing the performance tion (ppm)
of the model and the algorithm. Global opti- coc — Capital charge operational cost ($/y)
mal designs were obtained. coch — Chemical treatment operational cost
($/y)
coe — Energy operational cost ($/y)
6. Symbols
p
coom — Operation and maintenance cost
Ps — Average pressure in the fiber bore ($/y)
(atm) corp — Permeator replacement operational
b
Ps — Average pressure on the shell side of cost ($/y)
the fiber bundle (atm) cos — Spares operational cost ($/y)
Qbs — Brine flow rate (m3/h) cost — Cost of fresh water per m3 ($/m3)
Csm — Concentration at the membrane cpm — Permeator cost ($)
wall (ppm) Dif — Diffusion coefficient (m2/s)
Csb — Concentration in the high pressure dp — Specific surface diameter (m)
side (ppm) fc — Load factor
Csp — Concentration of the permeate ks — Mass transfer coefficient (m/s)
(ppm) L — Length of fiber bundle (m)
Csf — Feed concentration (ppm) NMs — Number of membranes
Qfs — Feed flow rate (m3/h) Pout — Average pressure used to calculate
f
Ps — Feed pressure (atm) the cost of the ERS (atm)
Qsp
— Permeate flow rate (m3/h) prodc — Total plant production (m3/h)
Vsw — Permeation velocity (m/h) Qfeed — Total seawater feed flow rate (m3/h)
JsS — Salt flux (kg/m2h) ro — Outer fiber radius (m)
si
Us — Superficial velocity at the inner Res — Reynolds number
radius of the fiber bundle (m/s) ð2  r0  UsS  b =b Þ
Usso — Superficial velocity at the outer ri — Inner fiber radius (m)
radius of the fiber bundle (m/s) Ri — Inner radius of the fiber bundle (m)
UsS — Superficial velocity (m/s) Ro — Outer radius of the fiber bundle (m)
" — Void fraction Scs — Schmidt number (b/bDif)
Jsw — Water flux (kg/m2h) Shs — Sherwood number (2kro/Dif)
A — Pure water permeability constant TCC — Total capital cost ($)
(kg/m2satm) w — Auxiliar variables
Am — Membrane area (m2) ybk — Auxiliar variables
AOC — Annual operating cost ($/y) ycj — Auxiliar variables
B — Salt permeability constant (m/s)
cccw — Civil work capital cost ($)
cceq — Total equipment cost ($) Acknowledgements
ccers — Energy recovery system capital cost ($) The authors acknowledge financial sup-
cchpp — High pressure pumps capital cost ($) port from ‘Agencia Nacional de Promoción
M.G. Marcovecchio et al. / Desalination 184 (2005) 259–271 271

Cientı́fica y Tecnológica’ (ANCyT), and K.S. Spliegler and A. Laird, eds., 2nd ed., Part
‘Consejo de Investigaciones Cientı́ficas y Téc- B, Academic Press, New York, 1980.
nicas’ (CONICET), Argentina. [7] A. Malek, M.N.A. Hawlader and J.C. Ho,
Design and economics of RO seawater desalina-
tion. Desalination, 105 (1996) 245–261.
References [8] N.G. Voros, Z.B. Maroulis and D. Marinos-
[1] M. Sekino, Study of an analytical model for Kouris, Short-cut structural design of reverse
hollow fiber reverse osmosis module systems. osmosis desalination plants. J. Membr. Sci., 127
Desalination, 100 (1995) 85–97. (1997) 47–68.
[2] N.M. Al-Bastaki and A. Abbas, Modeling an [9] A.M. Helal, A.M. El-Nashar, E. Al-Katheeri
industrial reverse osmosis unit. Desalination, and S. Al-Malek, Optimal design of
126 (1999) 33–39. hybrid RO/MSF desalination plants. Part I.
[3] N.M. Al-Bastaki and A. Abbas, Predicting the Modeling and algorithms. Desalination, 154
performance of RO membranes. Desalination, (2003) 43–66.
132 (2000) 181–187. [10] S. Kimura and S. Sourirajan, Analysis of data in
[4] A. Chatterjee, A. Ahluwalia, S. Senthilmurugan and reverse osmosis with porous cellulose acetate
S.K. Gupta, Modeling of a radial flow fiber module membranes. AIChe J., 13 (1967) 497.
and estimation of model parameters using numerical [11] N.M. Wade, Distillation plant develop-
techniques. J. Membr. Sci., 236 (2004) 1–16. ment and cost update. Desalination, 136 (2001)
[5] M. Cheryan, Ultrafiltration and Microfiltration. 3–12.
Handbook. Technomic Publishing Co., Inc., PA, [12] A. Brooke, D. Kendrick, A. Meeraus and
USA, 1998. R. Raman, GAMS Language Guide, Release 2.25,
[6] L. Dresner and J.S. Johnson Jr., Hyperfiltration Version 92. GAMS Development Corporation,
(reverse osmosis). in: Principles of Desalination, 1997.

You might also like