Você está na página 1de 16

Applied Mathematical Modelling 44 (2017) 727742

Contents lists available at ScienceDirect

Applied Mathematical Modelling


journal homepage: www.elsevier.com/locate/apm

Optimization of ASP ooding based on dynamic scale IDP


with mixed-integer
Yulei Ge a, Shurong Li a,, Peng Chang a, Songlin Lu a, Yang Lei b
a
School of information and control engineering, China University of Petroleum, Qingdao 266580, China
b
Fujian Metrology Institute, Fuzhou 350003, China

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

Article history: The alkali/surfactant/polymer (ASP) ooding is a complex distributed parameter system
Received 13 February 2015 (DPS). In this paper, an optimization model of ASP ooding is developed, which takes net
Revised 19 January 2017
present value (NPV) as the performance index, oil/water seepage continuity equations and
Accepted 14 February 2017
adsorption diffusion equations of displacing agents as the governing equations, physico-
Available online 22 February 2017
chemical algebraic equations and boundary conditions of displacing agents as the con-
Keywords: straint equations. To solve the injection concentration and size of each slug of the model
ASP ooding and terminal ooding time, a dynamic scale iterative dynamic programming with mixed-
Optimization integer (DSMI-IDP) is proposed. The essence of slug size is time, it can only be integer. In
Ddynamic scale IDP DSMI-IDP, the integer truncation is carried out by a proportion method after time normal-
Mixed-integer ization which can convert the free time terminal problem to a xed time terminal prob-
Adjustment factor lem. A dynamic contraction factor and a principle of adjustment factors are introduced to
Dynamic contraction factor realize the dynamic scale. To test the algorithm, three examples are solved by DSMI-IDP.
At last, the DSMI-IDP is applied to optimize an optimization model of ASP ooding. The
solving effect is shown by the comparison with IGA, MIDP and trial and error solutions.
2017 Elsevier Inc. All rights reserved.

1. Introduction

With the exploitation of oil eld continuing, the production ability of old oil eld in our nation is decreasing year by
year. It has entered a stage of high water content and high recovery. Traditional water ooding and chemical ooding cannot
satisfy our demands for oil production and economy prot. The question of how to enhance oil recovery and develop tertiary
oil recovery technique has become an important diculty in oil industry [1]. Alkali/surfactant/polymer (ASP) ooding as an
emerging tertiary oil recovery technique is now used to change the current situation. The synergistic effect among displacing
agents, such as alkali, surfactant and polymer, is fully utilized to further enhance oil recovery [2]. In ASP ooding, surfactant
is injected to cut down surface tension, and polymer is introduced to increase the viscosity of displacing uid, while alkali is
applied to improve the effect between alkali and surfactant. Through this way, the oil production can be enhanced evidently
[3]. ASP ooding is a complex distributed parameter system (DPS), the modeling and optimization are very dicult. Many
researchers have been studying this hot issue.
The mainstream researches of ASP ooding are about numerical simulations and physicochemical reactions, but very
few studies have discussed the model and optimization method. Yuan et al. (1988) proposed a multicomponent model of
chemical ooding considering the mass transformation equations and boundary conditions, to predict the characteristic and


Corresponding author.
E-mail addresses: geyulei@126.com (Y. Ge), lishuron@upc.edu.cn (S. Li).

http://dx.doi.org/10.1016/j.apm.2017.02.032
0307-904X/ 2017 Elsevier Inc. All rights reserved.
728 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

parameter inuence of displacement of reservoir oil [4]. Xian et al. (20 0 0) presented a new ASP ooding model in low
pH. He neglected the micro phase emulsion and brought in a new chemical reaction model [5]. Li et al. (2013) gave a
two-dimensional model for polymer ooding combing with the oil/water seepage property and adsorption of polymer, this
is the rst optimization model for polymer ooding [6]. Zhang et al. (2014) researched the adsorption law among all the
components in ASP ooding and explained the specic interaction relations [7].
Pilot experiment is often used to optimize the production decision, but it is too dependent on the experience of handlers
and is not practical. Zerpa et al. (2005) used eld scale numerical simulation and multiple surrogates to optimize alka-
line/surfactant/polymer ooding process based on UTCHEM. His method enhanced oil production obviously [8]. Lei (2011)
presented a mixed-integer iterative dynamic programming (IDP) to optimize the polymer ooding and obtained good effect
[9].
The optimization of ASP ooding is a complicated optimization problem. It covers not only the traditional optimization
problem, but also a mixed-integer problem because the slug size must be integer. Common methods cannot deal with this.
In addition, the ASP ooding is a complicated distributed parameter system. Its model consists of many high-order partial
differential equations. Regular methods need to solve the HJB equation. However this is usually very dicult to be realized
even not possible sometimes in control domain. To deal with this problem, we adopt the IDP which can avoid solving HJB
equation and simplify the optimizing process. To make IDP method more functional and have better effect, such as handling
the mixed-integer problem, a series of improvements are developed.
In this paper, a new optimization model of DPS for ASP ooding is proposed considering the net present value (NPV),
oil/water seepage continuity equations and adsorption diffusion equations of displacing agents, physicochemical algebraic
equations and boundary conditions of displacing agents. IDP is adopted to optimize this new model to get the optimal
injection concentration, the slug size and the terminal ooding time. Since the essence of slug size is time stage, it has to
be integer. To improve the standard IDP and make it more applicable to the proposed ASP model, a dynamic scale IDP with
mixed-integer (DSMI-IDP) is presented. A dynamic contraction factor and a principle of adjustment factors are introduced
to realize the dynamic scale. Finally, the DSMI-IDP is applied to optimize an optimization model of ASP ooding. The result
demonstrates that the net present value (NPV) gained by DSMI-IDP is higher than that of MIDP (the method proposed by
Lei in [9] we call it MIDP) and trial and error solutions (TES).
This paper is organized as follows. Section 2 presents a new optimization model of ASP ooding which consists of perfor-
mance index, governing equations, physicochemical algebraic equations and constrains. Section 3 gives a new optimization
method that is known as the dynamic scale IDP with mixed-integer (DSMI-IDP). A time normalization method, the dynamic
scale strategy and the integer truncation strategy are brought in. Then three examples are optimized by DSMI-IDP to prove
the effect. In Section 4, we build a specic optimization model of ASP ooding and optimize the optimal injection strategy
by DSMI-IDP.

2. Optimization model of ASP ooding

The main idea of modeling derives from [9]. We inherit the ow equation for water/oil phase and polymer component.
Adding physicochemical algebraic equations, some constraints and the ow equations for alkali and surfactant, we get the
new optimization model of ASP ooding.

2.1. Performance index

In an optimization model, the performance index is a functional. In this paper, the maximum of NPV is adopted as the
performance index, which means the prot maximization [10].
 tf 
max JNPV = (1 + r )t/ta [Cin (t ) Cout (t )]d dt , (1)
Np 0

where tf denotes the terminal ooding time, R2 is the domain of oil eld, r denotes the discount rate, ta denotes the total
discounting period, Np denotes the number of production wells, Cin (t) denotes the economic benet of oil, Cout (t) denotes the
capital input.
Cin (t ) = o (t )(1 fw )qout , (2)

Cout (t ) = p qinp cinp + a qina cina + s qins cins , (3)


where o (t), p (t), a (t), s (t) denote the coecient of crude oil price [11], polymer, alkali and surfactant, cinp , cina , cins and
qinp , qina , qins denote the injection concentration and rate of polymer, alkali and surfactant respectively, fw denotes moisture
content, qout denotes oil production.
The slug injection strategy is always adopted in ASP ooding. It is a piecewise function as Eq. (4).

vi (k ), t [tk1 , tk ], k = 1, 2, . . . , P 1,
cinp,ina,ins (xi , yi , t ) =   (4)
0, t tP1 , t f ,
where P 1 is the number of slug, vi (k) denotes injection concentration. Fig. 1 shows a three-slug injection scheme.
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 729

1.6

The injection concentration of displacing agents


1.4

1.2

0.8

0.6

0.4

0.2

0 500 1000 1500 2000 2500 3000


Time T/d

Fig. 1. Three-slug injection scheme.

2.2. Governing equations

Let (x, y)  denotes the domain of oil eld, and  be the boundary. There are Nin injection wells and Np production
wells. The locations of all wells are as follows:

Lin = {(xi , yi )|i = 1, 2, . . . , Nin },



Lp = x j, y j | j = 1, 2, . . . , N p .
The state variables of DPS are p, Sw and c, which denote pressure (MPa), water saturation and concentration of displac-
ing agents, such as alkali (%), surfactant (%), polymer (g/L). The DPS can be described as the following non-linear partial
differential equations [6].
The seepage continuity equation for oil phase is

p p
ro h + ro h (1 fw )qout
x x y y

( 1 Sw )
=h . (5)
t Bo

The seepage continuity equation for water phase is



p p Sw
rw h + rw h + qin fw qout = h . (6)
x x y y t Bw
The adsorption diffusion equations of displacing agents is
c KK 
(D p Sw c ) +  rw ( pw w gh ) qw c w
R k B w w

m Sw c (R (1 )Cr )
= + , (7)
t Bw t
where ro = K kro
B o o , r w = K krw
B w R k w , K is the absolute permeability (m2 ), D is the diffusion coecient of polymer (m2 /s), h is
the thickness of reservoir bed (m), r is the rock density (kg/m3 ), o is the oil viscosity (mPa s). p denotes the pressure,
E = { p, a, s} is the set of displacing agents. Eq. (7) can be divided into three equations,  E, p, a, s mean the polymer, alkali
ad surfactant respectively. As to alkali and surfactant, m = denotes the rock porosity, while to polymer, m = p denotes
polymer porosity. Bo and Bw denote the oil and water volume factors. The denitions of other parameters can be found in
[9].
The ux terms of injection wells for Eqs. (5)(7) are dened as

(1 fw )qout , (x, y ) L p ,
qo = (8)
0, (x, y )
/ L p,
730 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742


fw qout , (x, y ) L p ,
qw = qin , (x, y ) Lin , (9)

0, (x, y ) / L p Lin ,

c , (x, y ) L p ,
c w = c in , (x, y ) Lin , (10)

0, (x, y ) / L p Lin ,
where the moisture content of production well fw is given by
1
fw = kro w
. (11)
1+ o krw

The initial conditions and boundary conditions are


p(x, y, t )|t=0 = p0 (x, y ) , Sw (x, y, t )|t=0 = Sw
0
(x, y ),
(12)
c (x, y, t )|t=0 = c (x, y ), (x, y ) ,
0

  
p  Sw  c 
= 0 , = 0 , = 0. (13)
n   n   n  

2.3. Physicochemical algebraic equations

The descriptions for volume factor and porosity are


Bo = Bor [1 + Co ( p pr )],
Bw = Bwr [1 + Cw ( p pr )],
(14)
= r [1 + Cr ( p pr )],
p = fa ,
where fa is the effective pore volume coecient, Co , Cw and Cr denote the compressibility factors of oil, water and rock,
respectively. pr denotes the reference pressure (MPa), r , Bor and Bwr denote the porosity, the oil and water volume factors
under the condition of the reference pressure, respectively.
The relative permeability for oil phase and water phase kro , krw can be described as
Krw,ro = A (1 Sw )B (Sw C )D , (15)
where A, B, C, D denote the identication coecients, the specic method is shown in [12] and [13].
The viscosity of displacing uid is affected by polymer primarily. After dissolved in water, the viscosity and the perme-
ability reduction factor Rk will be changed as follows [14]:
w = w0 [1 + ( p1 c p + a p2 c2p + a p3 c3p )]Csep
sp
, (16a)

(Rkmax 1 ) brkp c
Rk = 1 + , (16b)
1 + brkp c
where Csep denotes the salinity, ap1 , ap2 , ap3 , sp, Rkmax , brkp are the constant parameters. w0 , w denote the viscosity of
water and displacing uid, respectively.
The interfacial tension of oil/water from surfactant is
ln wo = ai + biCT , (17a)

Cs (1 Sw )
CT = Cs Sw + . (17b)
Ks wo
The relation between residual oil saturation Sor and capillary number Ncap is

Sor = CSor1 + CSor2 ln Ncap Sor0 , (18)
w
where Ncap = vw
wo , CSor1 , CSor1 denote the coecient, vw denotes the ow velocity of water phase.
If the oil saturation is less than the residual oil saturation at current capillary number, the residual oil will not be dis-
placed, that is kro = 0, ifSw 1 Sor .
When there is alkali in displacing system, alkali will react with acidic compounds and produce a new surfactant which
can inuence the interfacial tension of oil/water, that is wo = wo (Cs )Ast (Ca ). Where wo (Cs ) is the surfactant interfacial
tension of oil/water without alkali. Ast (Ca ) is the tension multiplier with alkali.
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 731

Alkali can decrease the adsorbing capacity of polymer and surfactant distributed on the surface of rock, the original
Langmuir isothermal adsorption equation will be renewed as
Crla = Cl Aad (Ca ), (l = s, p), (19)
where Aad (Ca ) denotes the multiplier of alkali.

2.4. Constraints

The usage condition of displacing agent is

t f  
q in c in d dt m max , (20)
0 
where m max (kg) denote the upper mass bound of displacing agents.
The injection concentration condition is
0 c in c max , (21)
where c max denote the maximal concentration of displacing agents.
The terminal constraint of moisture content is

fw t=t f = 98% . (22)
The optimization model of ASP ooding covers all the above equations. It is a complicated distributed parameter system,
the state parameter consists of pressure p, water saturation Sw and grid concentration c ,  E. Traditional optimization
method cannot solve this problem. Here we propose a new dynamic scale IDP with mixed-integer (DSMI-IDP) based on IDP.

3. Dynamic scale IDP with mixed-integer

According to the slug injection strategy, the whole ooding process is divided into P stages. The former P 1 stages
are slugs of ASP ooding, and the Pth is water ooding. In order to be optimized by IDP, the performance index can be
discretized as Eq. (23a).


P
max JNPV = (1 + r )N (k)/ta [Cin (k ) Cout (k )], (23a)
k=1
 
N (k )1
 
Nout

Cin (k ) = o (n ) t n
1 n
fw, j qout , (23b)
n=0 j=1


Nin
Cout (k ) =  s(k ) cin (k )qin . (23c)
E, j=1

where n = 0, 1, , N(k) 1 denotes the discrete time steps, N(k) denotes the steps of time k, Nin , Nout denote the number
of injection wells and production wells respectively, tn denotes the time step length, qin = x yqin (n, k), qout = x yqout
(n, k), x, y denote the space step length in direction x and y, s(k)=tk tk 1 , k = 1, 2, , P denotes the size of kth slug.
In order to simplify calculation, we have to normalize the time and convert the free time terminal problem to a xed
time terminal problem.

3.1. Time normalization

Eqs. (5)(7) can be described as the below general form [9],


g (u )
= f(u, ux , uy , uxx , uyy , v), (24)
t
where u = (p, Sw , cp , ca , cs )T denotes the state vector in domain , ux = ux , uxx = xu2 , uy = uy , uyy = yu2 , v denotes the
2 2

control vector.
Bring in the normalized time variable , we dene t at kth slug [tk 1 , tk ] which meets the below condition
t = s(k )P . (25)
Integrating to both sides of Eq. (25), we can get
tk tk1 = s(k )P (k k1 ). (26)
732 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

Fig. 2. Diagram of iterative dynamic programming.

Every normalized slug size is

k k1 = 1/P. (27)
Let 0 =0, then k = k/P , k = 1, 2, . . . , P . Thus, time domain is changed from [0, tf ] to [0, 1]. At kth new slug [ k 1 , k ],
Eq. (24) is transformed into (28).
g (u )
= s(k )P f(u, ux , uy , uxx , uyy , v). (28)

Add slug size s(k) to control variables, then original optimization problem of varying time interval and free time terminal
is converted into a problem with same time interval and xed time terminal. We can use IDP [15,16] to solve this. The
standard IDP can be described as Fig. 2, in which three state grids and three admissible controls are adopted to express the
basic principle, that is M = 3, R = 3.
According to Fig. 2, we can nd that IDP is similar to dynamic programming (DP). Both of them start from the last
discrete time stage and solve the optimal decision of every time stage until to the rst time stage. But the time stage of
standard IDP needs to be determined before optimization, and its time node is xed which cannot be changed. It cannot be
used to deal with the mixed-integer problem and slug optimization problem in the optimization model of ASP ooding. In
view of these situations, we improve the standard IDP and develop the DSMI-IDP.

3.2. DSMI-IDP

DSMI-IDP can deal with mixed-integer problems and has better accuracy than standard IDP. In DSMI-IDP, dynamic scale
strategy is used to adjust the injection concentration, contraction factor and slug size dynamically. Integer truncation is
introduced to deal with mixed-integer. The specic process to optimize the optimization model of ASP ooding with DSMI-
IDP is as follows.

1) Initialization: divide the whole process of ASP ooding into P stages, the size of every slug is 1/P on the normalized time
domain [0, 1]. The optimization variables cover injection concentrations of displacing agents v(k)(including cp , ca ,
cs ) and slug size s(k), v(k ) = [v 1 (k ), v 2 (k ), . . . , v Nin (k )], k = 1, 2, . . . , P 1. E. The number of initial discrete state
grid is M. Select R values for every grid. Set dynamic contraction factor , maximal iteration lmax , initial control feasible
region r in , in and the initial control value. Start from generation l = 1.
 in , = in .
2) Set the current control domain r l =r l

3) In current feasible region, generate M 1 injection concentrations and slug sizes with uniform strategy [17] based on the
optimal control strategy got from previous generation. The uniform strategy is shown below.

2l
vl i (k ) = v(l1) (k ) r ( ) (k ).
l1
(29)
(M 1 ) 
Solve state equations in (28) from = 0 to = 1 with full implicit nite difference method [18], get M state trajectories.
At the beginning of each stage, there are M state values u which correspond to the M state grids.

4) From the beginning of Pth stage, that is to say from the time of P1 = (P 1 )/P , generate R slug sizes for every state
grid as Eq. (30).

sl (P ) = s(l1) (P ) + (l1) (P ). (30)

where denotes the random number generated from [ 1, 1], s(l 1) (P)
denotes the optimal slug size of previous iteration.
The slug size got from Eq. (30) is real number, it may be not integer. To satisfy the integer condition, we bring in the integer
truncation strategy.
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 733

Let sl (P ) = sl (P ) + s, where sl (P ) denotes the integer part, s denotes the decimal part. Then the integer truncation
strategy can be described as follows:

sl (P ), 0 v < 0.7v max ,
sl ( P ) = (31)
sl (P ) + 1, 0.7v max v v max .
The possibility of trunc is higher than that of rounding by a boundary proportion, this can make the performance index
become less than before because of the expensive displacing agents.
Note that v (P ) 0, solve Eq. (28) with nite difference method based on R integer slug sizes and injection concentra-
tion in every state grid from P 1 to f = 1. Then calculate the performance index in this stage, save the optimal slug size
through comparison.

5) From the beginning of (P 1)th stage, that is to say from the time of P2 = (P 2 )/P , generate R displacing agents
concentration and slug size for every state grid as Eqs. (32a) and (32b).
(l1 )
vl  (P 1 ) = v(
l1 )
( P 1 ) + r (P 1 ), (32a)

sl (P 1 ) = s(l1) (P ) + (l1) (P 1 ), (32b)


(l1 )
where denotes a diagonal matrix with the dimension Nin Nin , v (P 1 ) denotes the optimal injection concentration
of previous iteration. and are the adjustment factor. We dene an adjusting principle as follows:

rand (1, 1), , 0 l 0.3lmax ,
{ , } = rand (0.5, 0.5), 0.3lmax < l 0.7lmax , (33)

rand (0.2, 0.2), 0.7lmax < l lmax .
where rand denotes the operation of generating a random number in a certain interval. Since the price of displacing agents
is expensive, to make the NPV as less as possible, the proportion is changed with iterations. At the initial stage of iteration,
proportion is relatively big to ensure that the result can approach the extreme point fast. At the later stage, proportion is
relatively small to help realize the accuracy demand. The random strategy can ensure the high search capability.
When generating injection concentrations which not conform to Eq. (21), we deal with the control variables as follows.

0, v i ( P 1 ) < 0
v i ( P 1 ) = , i = 1, 2, . . . , Nin . (34)
c max , v i (P 1 ) > c max

Solve Eq. (28) with nite difference method based on R integer slug sizes and injection concentrations in every state grid
from P 2 to P 1 . Then select the optimal injection concentration and slug size which are proximal to the terminal point
of the front difference from P 1 , and solve Eq. (28) with them until f = 1. Calculate the performance index between P 2
and f = 1, save the optimal slug size and injection concentration on every state grid in slug (P 1) through comparison.

6) Shift the time step forward, repeat step 5) until the rst slug, namely 0 . Compute the optimal performance index of the
whole normalized time domain in Eq. (23), and save the corresponding injection concentration and slug size.
7) Shrink the feasible domain of decision variables:
(l+1 )
r (k ) = (l+1) r
l
(k ), k = 1, 2, . . . , P, (35a)

(l+1) (k ) = (l+1) l (k ), k = 1, 2, . . . , P, (35b)


where (l + 1)
denotes the dynamic contraction factor. Take the optimal injection concentration and slug size which are
obtained from step 6) as the center of feasible domain in next iteration.
Dynamic contraction factor is another refection of dynamic scale. The contraction factor in [9] is xed, this restricts the
search capability. To search more solutions, here the linear searching strategy is brought in. At the early searching, is
small to accelerate search. But with the iteration going on, becomes bigger which can improve the accuracy of solutions.
The specic formula is shown below,
l+1
(l+1) = 0 + (max 0 ) , (36)
lmax
where max and 0 denote the maximum and minimum of dynamic contraction factor, they are given by experience, usually
be 0.9 and 0.7, respectively.

8) Add the iteration l = l + 1, go to step 3). Continue implementing steps 3)5) until l > lmax . Then save the values of optimal
injection concentration and slug size, nish the optimization process of DSMI-IDP.

In regard to the integral constraints of displacing agent usage Eq. (21) and terminal state constraint Eq. (22), we adopt
penalty function method and construct the augmented performance index [19,20]. The solving process is the same as above.
734 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

0.5

-0.5
Analytical solution
-1 DSMI-IDP

Control v
-1.5

-2

-2.5

-3

-3.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time t

Fig. 3. Optimal control trajectories for LQR problem.

3.3. Algorithm test

The characteristic of DSMI-IDP is the integer truncation strategy. Since it is dicult to nd a classical and widely recog-
nized mixed-integer example, to test the effect of DSMI-IDP we ignore the integer truncation strategy and optimize three
classical optimal control problems based on MATLAB R2011b.

Example 1. LQR optimal control problem [21].

Consider below LQR problem,



1 5
min J = u(t )2 + v(t )2 dt,
2 0
s.t .u (t ) = u(t ) + v(t ), (37)

u(0 ) = 2, u t f = 1.

Solve the Riccati equation [21], then we can get the analytical solution of Eq. (37),

u(t ) = 1.4134 e 2t
+ 8.4831 104 e 2t
,


v(t ) = 3.4122 e 2t
+ 3.5138 104 e 2t
.

Set the stage P = 20, admissible control R = 3, 0 = 0.7, max = 0.9, grid number M= 1. Solve the LQR problem with DSMI-
IDP, the convergence precision is =1 10 4 . Get the optimal performance index J = 2.6367while the analytical solution is
J = 2.6180. The optimal control curve is shown in Fig. 3. The control is very close to the analytical solution. The relative error
is only 0.71%. DSMI-IDP has good effect in solving LQR problem.

Example 2. Bang-Bang optimal control problem [22].

Consider below Bang-Bang problem,

min J = t f ,
s.t .u 1 (t ) = u2 (t ),
u 2 (t ) = v2 (t ),
(38)
u1 ( 0 ) = 1, u2 ( 0 ) = 1,
 
u1 t f = 0, u2 t f = 0,
|v(t )| 1.
Solve the analytical solutions of Eq. (38) by maximum principle,

t 2 /2 + t + 1, t ts ,
u1 (t ) =
t 2 /2 t f t + t 2f /2, t > ts ,
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 735

1 Analytical solution
DSMI-IDP
0.8

0.6

0.4

0.2

Control v
0

-0.2

-0.4

-0.6

-0.8

-1

0 0.5 1 1.5 2 2.5 3 3.5


Time t

Fig. 4. Optimal control trajectories for Bang-Bang problem.


t + 1, t ts ,
u2 (t ) =
t t f , t > ts ,

1, t ts ,
v(t ) =
1, t > ts ,
 
where ts = 1 + 3/2, the terminal time t f = 1 + 2 3/2.
Set the stage P = 2, admissible control R = 3, 0 = 0.7, max = 0.9, grid number M= 1. Solve the LQR problem with DSMI-
IDP, the convergence precision is =1 10 4 . Get the optimal performance index J = 3.4495. The optimal control curve is
shown in Fig. 4. The control is very close to the analytical solution. DSMI-IDP has good effect in solving bang-bang optimal
control problem.

Example 3. Continuous stirred tank reactor (CSTR) optimal control problem [23].

The governing equations of CSTR optimal control problem consist of below two nonlinear differential equations:
 25u 
1
u 1 = (2 + v )(u1 + 0.25) + (u2 + 0.5 ) exp , (39)
u1 + 2
 25u 
1
u 2 = 0.5 u2 (u2 + 0.5 ) exp , (40)
u1 + 2
where u1 and u2 are the state variables, u1 denotes the temperature deviation of dimensionless equilibrium state, u2 de-
notes the concentration deviation of dimensionless equilibrium state. v(t) denotes the control value of coolant ow injected
to the reactor. The above two equations describe the thermal mass equilibrium process of a rst-order irreversible chemical
reaction in continuous stirred tank reactor. Make the below performance index be the minimum value by searching the
optimal control v (t).
 tf 
min J = u21 + u22 + 0.1v2 dt, (41)
0

where the terminal time is tf = 0.78, the state initial condition is u(0) = [0.09, 0.09]T , the interval of control value is 0 v(t)
5.0.
Solve the CSTR problem with DSMI-IDP. To demonstrate the improved effect, we introduce the method proposed by Lei in
[9] (we call it MIDP in the following content) and the iterative genetic algorithm (IGA) [24]. The parameters are as follows:
DSMI-IDP: Stage number P = 13, admissible control R = 3, 0 = 0.7, max = 0.9, grid number M= 1, convergence precision
=1 10 4 .
MIDP: Stage number P = 13, admissible control R = 3, = 0.85, grid number M= 1, convergence precision =1 10 4 .
IGA: Stage number P = 13, population number 20, crossover probability 0.9, mutation probability 0.4, convergence preci-
sion =1 10 4 .
The results are shown in Table 1. The control curves of MIDP, IGA and DSMI-IDP are demonstrated in Fig. 5.
According to Table 1 and Fig. 5, DSMI-IDP can get the global optimal solution effectively. In Fig. 5, we can easily nd that
the whole searching process can be divided into three stages. The searching speed of the rst stage is the fastest, because
736 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

Table 1
Results for the optimal control problem of CSTR.

Algorithm Performance index Operation time (s) convergence precision

MIDP 0.1338 10.6 1 10 4


IGA 0.1330 42.8 1 10 4
DSMI-IDP 0.1326 9.2 1 10 4

4.5
MIDP
4 DSMI-IDP
IGA
3.5

3
Control v

2.5

1.5

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Time t

Fig. 5. Comparison of optimal control trajectories for CSTR problem.

we need nd the roughly range of the optimal solution as fast as possible. While the searching speed of the third stage is
the slowest, because at the later period of searching, we need slow down and search the optimal solution as intensively
as possible. This phenomenon depends on the design and implementation of DSMI-IDP. No matter the performance index
or the operation time, DSMI-IDP can acquire the best performance index with the least time among the three methods.
For comparing conveniently, the non-uniform grid is adopted in IGA. The result with IGA is very close to that of DSMI-IDP,
but it consume more time because of the property of intelligence algorithm. For MIDP, it lacks the dynamic scale adjusting
process than DSMI-IDP which causes the diculty for optimal solutions. So DSMI-IDP has the best performance in the three
methods.
From the above, DSMI-IDP which is proposed in this paper has less operation time and better accuracy than common
methods. It can be used to optimize the optimization model of ASP ooding.

4. Optimization of ASP ooding with DSMI-IDP

In this section, we establish the optimization model of ASP ooding with four injection wells and nine production wells
based on the presentation in Section 2. The displacing agents cover alkali, surfactant and polymer. All wells are distributed
uniformly as Fig. 6. The rst letter S of all wells represents production well, and I represents injection well. The data of ow
property is shown in Table 2.
The domain of two-dimensional oil eld is 630 630m3 , the grid of direction x, y is 21 21. The depth of upper surface is
2420m, the porosity of every layer is 0.3 and the pore volume is 3.8881 105 m3 . The permeability distribution in direction
xis shown in Fig. 7. The initial water saturation, initial oil saturation and initial pressure are shown in Figs. 810. Figs. 610
are got from the CMG platform which is a kind of reservoir numerical simulation software.
The whole process covers three parts: water ooding, ASP ooding and water ooding. ASP ooding starts at fw = 97%
in the rst part of water ooding. The time for ASP ooding is 30 0 0 days. Displacing uid is injected to oil eld through
four injection wells by way of three-slug-mode, that is P=4, in which the last slug is water ooding. The water injec-
tion rate of every injection well is 83m3 /d. Supposing the price of all displacing agents is  = a = 16.67$/kg  = s = 2.5$/kg
 = p = 3.33$/kg, the price of oil is o = 610$/m3 . The concentration range of displacing agents is cp [0, 3](g/L)ca [0, 2](%)
cs [0, 1](%). The maximum of displacing agents is mpmax = 10 0 0 0 0 0kgmamax = 50 0 0 0 0kgmsmax = 50 0 0 0 0kg.
DSMI-IDP is used to optimize the ASP ooding optimization model based on MATLAB R2011b, the detailed process is
described as Section 3. To explain the optimization effect of DSMI-IDP, we compare the result with that of MIDP [9], IGA
[24] and trial and error solutions (TES) [6]. To apply IGA in this optimization of ASP ooding, the same integer truncation
strategy is added to deal with the mixed-integer problem. Select the middle value of injection concentration as the trial and
error solutions, that is cp = 1.5g/L, ca = 1%, cs = 0.5%.
The crude oil price is predicted by the method of dynamic correcting SVRin [11]. The unit is $/t, and the predicting
result is shown in Fig. 11.
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 737

Fig. 6. The distribution diagram of well position.

Parameter selection of DSMI-IDP: grid number M=3, R = 15, 0 = 0.7, max = 0.9, lmax = 40, r in (rpin = 1, rain = 0.7,
rsin = 0.4), in = 100, r = 0.12.
Parameter selection of MIDP: grid number M=3, R = 15, = 0.8, r in (rpin = 1, rain = 0.7, rsin = 0.4), in = 100, r = 0.12.
Parameter selection of IGA: grid number M = 3, population number 20, crossover probability 0.9, mutation probability
0.4.
The convergence precision for the three methods is =1 10 4 . The optimization result is shown in Figs. 1216.
In Figs. 1216, TES means the trial and error solutions. The red , blue , black and cyan denote the terminal
ooding time for DSMI-IDP, TES, MIDP and IGA, respectively. The red , blue , black and cyan R denote the terminal
injection time of displacing agents, separately.

Table 2
Flow data.

Sign Mean Value Sign Mean Value

Oil o Density (kg/m3 ) 950 o Viscosity (mPa s) 50


Co Compression coecient (1/MPa) 5 10 6 Bo Volume coecient 1
Water w Density (kg/m3 ) 10 0 0 w Viscosity (mPa s) 0.458
Cw Compression coecient (1/MPa) 4 10 6 Bw Volume coecient 1
Polymer a Adsorption constant (cm3 /g) 0.03 b Adsorption constant (L/g) 3.8
D Diffusion coecient (m2 /s) 1 10 5 fa Attainable pore volume coecient 1
Rkmax Permeability reduction factor 1.15 brkp Permeability reduction factor 1.2
ap1 Viscosity constant 15.426 ap2 Viscosity constant 0.4228
ap3 Viscosity constant 0.2749
C a2+
Others NaCL Concentration (g/m3 ) 0.017 Concentration (g/m3 ) 0.0 0 045
Mg2+

Fig. 7. Permeability distribution in x direction.


738 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

Fig. 8. The distribution diagram of initial water saturation.

Fig. 9. The distribution diagram of initial oil saturation.

Fig. 10. The distribution diagram of initial pressure.

According to Figs. 1214, the optimized injection concentrations for three displacing agents are different. As regard to
the four injection wells, the results of MIDP and IGA are similar to that of DSMI-IDP, except the very slight error of slug
size and concentration. DSMI-IDP has the shortest terminal ooding time and the longest injection time of displacing agent
among the three methods. The effects of all the above three methods are much better than that of TES. To demonstrate the
difference further, we calculate the average moisture content and cumulative oil production of nine production wells. Then
draw the curves as Figs. 15,16.
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 739

1100

1000

900

800

oil price $/t


700

600

500

400

300

200
0 500 1000 1500 2000 2500 3000
Time t/d

Fig. 11. The price of crude oil.

2
Cp(g/L)

0
0 500 1000 1500 2000 2500 3000
Time T/d (I4-251)

2
Cp(g/L)

0
0 500 1000 1500 2000 2500 3000
Time T/d (I4-256)

2
Cp(g/L)

0
0 500 1000 1500 2000 2500 3000
Time T/d (I5-263)

2
Cp(g/L)

0
0 500 1000 1500 2000 2500 3000
Time T/d (I5-234)

Initial injection IGA DSMI-IDP MIDP

Fig. 12. The injection concentration comparison of polymer.

In Figs. 15 and 16, no matter the average or the cumulative oil production, MIDP, IGA and DSMI-IDP are with lower
moisture and more oil production than that of trial and error solutions obviously. The result of DSMI-IDP is the best among
all the methods because of the shortest terminal ooding time and best ooding effect. The result of IGA is very close
to that of DSMI-IDP, which is because of the advantage of intelligent algorithm in searching optimal solution. But the non-
uniform grid cannot obtain the optimal slug size. For MIDP, the contraction factor and adjustment factor cannot be regulated
dynamically, the optimization effect is a little worse than that of DSMI-IDP and IGA, but is better than that of TES.
740 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

Ca(%)
1

0
0 500 1000 1500 2000 2500 3000
Time T/d (I4-251)

Ca(%)
1

0
0 500 1000 1500 2000 2500 3000
Time T/d (I4-256)

2
Ca(%)

0
0 500 1000 1500 2000 2500 3000
Time T/d (I5-263)

2
Ca(%)

0
0 500 1000 1500 2000 2500 3000
Time T/d (I5-234)

Initial injection IGA DSMI-IDP MIDP

Fig. 13. The injection mass concentration comparison of alkali.

Table 3
Comparison of optimization result.

TES DSMI-IDP MIDP IGA

Terminal ooding time(d) 2970 2880 2910 2889


Terminal ASP injection time(d) 1260 1170 1140 1165
Polymer consumption(t) 493.02 619.01 596.46 619.11
Alkali consumption(t) 328.68 407.66 392.10 408.22
Surfactant consumption(t) 164.34 102.19 98.86 102.08
Oil production(m3 ) 73,134.03 81,088.47 80,172.79 80,776.24
NPV(million $) 37.080 40.351 40.135 40.223

To distinguish the effect, we compute the relevant data. The specic optimization result of four methods is shown in
Table 3.
According to Table 3, TES is the worst one among the three methods with the least oil production and NPV and the
longest terminal ooding time. The result of IGA is the best except DSMI-IDP. Comparing with IGA, after optimizing the
injection strategy by DSMI-IDP, we add the oil production to more than 312.23m3 and increase the NPV income by 0.128
million dollars. The oil recovery is enhanced evidently. The optimization effect of DSMI-IDP is better than that of MIDP, IGA
and TES for both the injection concentration and slug size.

5. Conclusions

In this paper, a new optimization model of ASP ooding is established considering the performance index, governing
equations, physicochemical algebraic equations and some constraints. To solve this optimization problem of DPS, a DSMI-
IDP is proposed to optimize the injection concentration, slug size and terminal ooding time. The main conclusions obtained
from this study are as follows.

(1) A new optimization model of ASP ooding is established. It is reasonable for considering seepage theory and the
interaction among displacing agents, the performance index and constraints as well.
Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742 741

0.6

0.4

Cs(%)
0.2

0
0 500 1000 1500 2000 2500 3000
Time T/d (I4-251)

0.4
Cs(%) 0.2

0
0 500 1000 1500 2000 2500 3000
Time T/d (I4-256)

0.4
Cs(%)

0.2

0
0 500 1000 1500 2000 2500 3000
Time T/d (I5-263)

0.6

0.4
Cs(%)

0.2

0
0 500 1000 1500 2000 2500 3000
Time T/d (I5-234)

Initial injection IGA DSMI-IDP MIDP

Fig. 14. The injection mass concentration comparison of surfactant.

98

96
The average fw of 9 production wells (%)

94

92

90

88

86

84 TES
IGA
82 DSMI-IDP
MIDP
80
0 500 1000 1500 2000 2500 3000
Time T/d

Fig. 15. The comparison of average moisture content for 9 production wells.

(2) DSMI-IDP is proposed by bringing in a mixed-integer strategy and a dynamic scale strategy to deal with the integer
problem and improve algorithm performance.
(3) As to the optimization problem of ASP ooding, we solve the optimization model by DSMI-IDP. NPV is increased sub-
stantially which veries the validity of DSM-IDP. Furthermore, the improved IDP in this paper is suitable to optimizing
ASP ooding.
742 Y. Ge et al. / Applied Mathematical Modelling 44 (2017) 727742

4
x 10
9
TES

Cumulative oil production of 9 production wells (m3)


8 IGA
DSMI-IDP
7 MIDP

0
0 500 1000 1500 2000 2500 3000
Time T/d

Fig. 16. The comparison of cumulative oil production of all production wells.

Acknowledgment

This work is supported by National Nature Science Foundation of China under Grant No. 60974039 and No. 61573378,
Natural Science Foundation of Shandong province under Grant No. ZR2011FM002, the Fundamental Research Funds for the
Central Universities under Grant No. 15CX06064A.

References

[1] H.L. Chang, Z.Q. Zhang, Q.M. Wang, et al., Advances in polymer ooding and alkaline/surfactant/polymer processes as developed and applied in the
Peoples Republic of China, J. Petroleum Technol. 58 (2) (2006) 8489.
[2] G. Shutang, G. Qiang, Recent progress and evaluation of ASP ooding for EOR in Daqing oil eld, in: Proceedings of the EOR Conference at Oil & Gas
West Asia. Society of Petroleum Engineers, SPE-127714-MS, 2010, pp. 17.
[3] W. Kon, M.J. Pitts, H. Surkalo, Mature wateroods renew oil production by alkaline-surfactant-polymer ooding, Proceedings of the Eastern Regional
Meeting. Society of Petroleum Engineers 26 (4) (2002) 355362.
[4] S.Y. Yuan, N.VAN Quy, Multidimensional numerical simulation of chemical ooding (part one), ACTA Peroei Sinica 9 (1) (1988) 5160.
[5] CH.G. Xian, ZH.X. Lang, H. Cheng, The mathematical model and its application of ASP ooding, J. Univ. Petroleum 24 (2) (20 0 0) 6169.
[6] SH.R. Li, X.D. Zhang, Optimal Control of Polymer Flooding For Enhanced Oil recovery, China University Of Petroleum Press, Dongying, 2013, pp. 1923.
[7] L.B. Zhang, H.Y. Cai, Q. Wang, H.F. Fan, M. Tang, H.H. Chen, Study on static adsorption of each components during the process of ASP ooding,
Petroleum Geology Recovery Ec. 21 (2) (2014) 3237.
[8] L.E. Zerpa, N.V. Queipo, S. Pintos, et al., An optimization methodology of alkalinesurfactantpolymer ooding processes using eld scale numerical
simulation and multiple surrogates, J. Petroleum Sci. Eng. 47 (3) (2005) 197208.
[9] Y. Lei, SH.R. Li, X.D. Zhang, Q. Zhang, L.L. Guo, Optimal control of polymer ooding based on mixed-integer iterative dynamic programming, Int. J.
Control 84 (11) (2011) 19031914.
[10] Y.L. Ge, SH.R. Li, S.L. Lu, et al., Spatial-temporal ARX modeling and optimization for polymer ooding, Math. Problems Eng. 2014 (2014) 110.
[11] SH.R. Li, Y.L. Ge, Crude oil price prediction based on a dynamic correcting support vector regression machine, Abstract and Applied Analysis 2013 (1)
(2013) 666686.
[12] Y.L. Ge, SH.R Li., Computation of reservoir relative permeability curve based on RBF neural network, CIESC J. 64 (12) (2013) 45714577.
[13] Y.L. Ge, SH.R. Li, K.X. Qu., A novel empirical equation for relative permeability in low permeability reservoirs, Chin. J. Chem. Eng. 22 (11) (2014)
12741278.
[14] CH.ZH. Yang, et al., Enhanced Oil Recovery For Chemical Flooding, 10, Petroleum Industry Press, Beijing, 2007, pp. 328344.
[15] R. Luus, Optimal control by dynamic programming using systematic reduction in grid size, Int. J Control 51 (5) (1990) 9951013.
[16] R. Luus, Application of dynamic programming to very high-dimensional non-linear optimal problems, Hung. J. Ind. Chem. 17 (1990) 523543.
[17] L.L. Guo, Scheme Design of Polymer Flooding Via Dynamic optimization, China University of Petroleum Press, Dongying, 2012, pp. 109113.
[18] K. Aziz, A. Settari, Fundamentals of Reservoir Simulation, Elsevier Science &Technology Publishers, New York, 1986.
[19] R. Luus, Iterative Dynamic Programming, Chapman and Hall/CRC, London,UK, 20 0 0.
[20] R. Luus, Handling inequality constraints in optimal control by problem reformulation, Ind. Eng. Chem. Res 48 (2009) 96229630.
[21] G.CH. Zhang, The Optimal Control Method, Chengdu University Of Science And Technology Press, Chengdu, 1991.
[22] X.S. Cai, Optimization and Optimal Control, Tsinghua University Press, Beijing, 1983.
[23] I.L. Lopez Cruz, L.G. Van Willigenburg, G. Van Straten, Ecient differential evolution algorithms for multimodal optimal control problems, Appl. Soft
Comput. 3 (2) (2003) 97122.
[24] F. Qian, F. Sun, W.M. Zhong, et al., Dynamic optimization of chemical engineering problems using a control vector parameterization method with an
iterative genetic algorithm, Eng. Optim. 45 (9) (2013) 11291146.

Você também pode gostar