~

Pergamon

Applied Thermal Engineering Vol. 18, Nos. 1-2, pp. 25-45, 1998

PII: S1359-4311(97)00020-3

1997 Elsevier Science Ltd. All rights reserved. Printed in Great Britain 1359-4311/98 $19.00 + 0.00

PERFORMANCE OPTIMIZATION OF DRY-COOLING SYSTEMS FOR POWER PLANTS THROUGH SQP METHODS A. E. Conradie,* J. D. Buyst and D, G. Kr6ger*
*Department of Mechanical Engineering, University of Stellenbosch, SteUenbosch, 7600, South Africa; and tDepartment of Mathematics, University of Stellenbosch, Stellenbosch, 7600, South Africa

(Received 15 January 1997)

Abstract--Inthis study the application o f modern optimization techniques to obtain cost optimal design and performance of dry-cooling systems for power plant applications, is illustrated. The Sequential Quadratic Programming (SQP) method, as well as a SQP decomposition technique are implemented. It is shown that through the proper application of these powerful optimization strategies and careful tailoring of the well-constructed optimization model, direct optimization of complex models does not need to be time-consuming and difficult. The optimal results and trends that one is able to obtain by means of the efficient computational procedures discussed in this paper, pose a challenge and establish new design and performance practices to designers, manufacturers and operators. The degree of optimization that is ultimately achieved in such an economic optimization analysis and its value to a design engineer, are functions of the sophistication of the design program, the expertise of the user, the cost-estimating procedure and the quality of the input data. q~ 1997 Elsevier Science Ltd.
Keywords--Dry-cooling, power plants, engineering optimization, Sequential Quadratic Programming (SQP), SQP decomposition.

NOMENCLATURE .4 A A area, m 2 set of active constraints matrix approximation for the Hessian matrix cost, $, or coefficient constraint function diameter, m energy objective function gravitational acceleration, m/s 2 height, m vector o f equality constraints, or vector unit matrix loss coefficient length, m mass flow rate, kg/s, or number of constraints revolutions per second, s number of variables, or number pitch, m pressure differential, N/m 2 heat transfer rate, W orthogonal matrix search direction temperature, C or K thickness, m vector o f Lagrange multipliers volume flow rate, m3/s, or volume, m 3 velocity, m/s vector of Lagrange multipliers width, in vector of decision variables basis matrix 25

B
C c(x,y) d E f(x,y)

H h(x,y)

1
N

K L
177

P Ap

Q Q
S

T
I

v
v v

W
x

26 y Z V Vx R" e q 0 p Subscripts a A b c ct do eq F f Fb Fd Fr h i 1 max min n o r s sd t tb tg ts u up v w Y Z Superscript T transpose of vector or matrix

A. E. Conradie et al. vector of variables basis matrix first derivative operator (O/ax,) partial derivative operators with respect to x n-dimensional space fraction surface roughness, small positive constant efficiency angle, density, kg/m3, or constant

air, or based on airside area set of active constraints bundle, or bellmouth casing, or condensate cooling tower downstream equality constraints fan fin, or friction fan bays fan drive system fan rows hub, or header inlet, or inside lower, or longitudinal maximum minimum net outlet, or outside root, or row static, or steam steam duct tube, or transversal, or total, or turbine tubes per bundle turbo-generator tower support upper upstream vapor water, or walkway range space null space

INTRODUCTION T h e recent increase in the use o f d r y - c o o l i n g systems for p o w e r plants is a direct result of the g r o w i n g d e m a n d for electricity, the shortage o f water at p o t e n t i a l p l a n t sites a n d stringent env i r o n m e n t a l legislations limiting the use o f wet-cooling systems. A l t h o u g h the design objective is generally to p r o d u c e electric power at the lowest cost, d r y - c o o l i n g systems are, however, associated with a n increase in electric power g e n e r a t i o n cost c o m p a r e d to wet-cooling systems. The significant effect of d r y - c o o l i n g o n the overall p l a n t economics, due to its higher capital a n d o p e r a t i n g costs as well as the t h e r m o d y n a m i c p e r f o r m a n c e o f the turbine, m a k e it necessary to optimize the c o s t - p e r f o r m a n c e trade-off that exists in these systems. Since the p e r f o r m a n c e o f a p a r t i c u l a r d r y - c o o l i n g system a n d the t u r b i n e which it serves are so closely related the complete c o n d e n s i n g system a n d the t u r b i n e can best be considered as o n e integral u n i t in e c o n o m i c opt i m i z a t i o n studies. Several papers have addressed the p r o b l e m o f e c o n o m i c o p t i m i z a t i o n o f d r y - c o o l e d power plants, e.g. [1-9]. A variety o f m e t h o d o l o g i e s are presented to solve these cost m i n i m i z a t i o n problems. M o s t o f these studies tend to limit the n u m b e r o f relevant p a r a m e t e r s due to practical

Performance optimization dry-cooling of systems

27

considerations or in order to simplify the problem. In some cases the optimization was restricted to the adaptation of predesigned commercially available system components. More complex optimization studies are attempted by Ecker and Wiebking [10] and Buys and Krrger [11, 12] where detailed models are developed by using the relevant conservation equations, physical laws and engineering design relations. Prescribed cost structures and practical constraints are employed in the problem formulations. The minimum annual costs are computed by employing modern mathematical programming techniques: geometric programming [10] and Sequential Quadratic Programming (SQP) [11, 12]. Recent advances in dry-cooling technology as well as mathematical programming techniques make it desirable to develop effective design procedures to ensure their optimum performance in power plant applications. A sophisticated equation-based model is presented by Conradie and Krrger [13] to model the natural draft indirect dry-cooling tower and forced draft direct air-cooled condenser performance for power plant applications (refer to Fig. 1). In the present paper this simulation procedure is coupled to a constrained nonlinear programming code to design dry-cooling systems which minimize the power generation cost attributed to the performance of dry-cooling systems for given mean hourly annual temperature frequencies and turbo-generator characteristics. The optimization model takes into consideration all the parameters that may affect the capital and operating cost of the dry-coofing systems and does not prescribe any limits, other than those absolutely essential due to practical limitations and to simulate the systems effectively. The Sequential Quadratic Programming (SQP) method [14] is used as the basis in implementing the constrained nonlinear optimization techniques. The equation-based models and the op-

(a)
- ~ Processfluid (steam) condenser Condensate Circulatingpump
(b) / - - Ste~n

/
Steamturbine

l, ~

Recooled.

I/"--I coolingNattowerdraft ural I/ ~ Heatexchanger ~ - - bundles

U Inll,l |llliilllllllllllll

y I

=' I

Airflow

Circulatingpump

Heatedfluid (water)

/ b
Steamturbine

....

l/ /k ~]['I"/ ~ /

Exhauststeamduct Supplyheader
Heated exchanger

I ]/ I ~''--- I I
"----,
'

~..//~n~.,,~
Cndensatel
c..}-l~

bundles

C~ensate pump ~ a g e

. . I/-- Condensate tank I ~Airflow~ I

Fig. 1. (a) Naturaldraftdry-coolingtowerwithsurfacecondenser(indirectsystem).(b) Forceddraft directair-cooledcondenser(A-framearrangement).

28

A.E. Conradieet,al:.

timization problems are simultaneously solved along an infeasible path [15]. Significant computational advantages are achieved when the special mathematical structure of the equation-based models is exploited by adapting the basic SQP code: A SQP decomposition strategy [15-17] is introduced to overcome the increase in matrix storage requirements and computational expense for large optimization problems. The decomposition strategy has a very superior performance when compared to the conventional SQP optimization method and solves the dry-cooling system optimization problem very effectively.

PROBLEM FORMULATION The minimization of the power generation cost attributed to the dry-cooling system's performance for the given temperature frequency data set, involves the variation of the dry-cooling tower operating and geometrical variables that will minimize the ratio of its total annual cost to the net energy output of the turbo-generator set to which it is coupled. The objective function is Cpower = Ctotal/En (S/kWh) (1)

where the total annual cost is estimated from the procedures presented in Appendix A, and the total net annual power output is computed as described in [13]. Studies have shown that performing an optimization study while repeatedly simulating the physical model requires prohibitive computational effort. Therefore, the simulation model is formulated as part of the optimization problem and solved simultaneously (infeasible path integrated approach) [15]. The dimensionality of the problem is substantially increased in this way and little information is recoverable if the solution algorithm fails. For the dry-cooling system optimization problem the decision variables and constraints are divided into distinct classes (Appendix B), i.e. independent or geometrical variables, dependent or operating variables, geometrical constraints, operating constraints and conservation equations. The dry-cooling system optimization problems can be formulated as nonlinear constrained mathematical programming problems as follows: Minimizex,y f (x,y) subject to the constraints ci(x,y) = 0, i = 1..... meq ci(x,y)>0, i -----meq + 1. . . . . h(x,y)=O,
mt

(2)

where f = objective function (e.g. total annual cost/net annual power output), x = (n-r)-component vector of independent variables (e.g. geometrical variables), y = r-component vector of dependent variables (e.g. operating variables), ci= general constraints (e.g. geometrical or operating constraints), h = r-component vector of equality constraints (e.g. conservation equations needed for model simulation), r = number of equality constraints, meq=number of equality constraints, mt = total number of general constraints, n = total number of variables. It is assumed that the problem is formulated in such a way that, for fixed x values, it is possible to determine y (r-component vector) by solving the equality constraints, h(x,y) = 0, that describe the physical model to be optimized. This procedure is known as the sequential or feasible path method. However, we use the infeasible path integrated approach where x and y are adjusted simultaneously and all the equations need only to be satisfied at the final solution. The variables and constraints of the minimization problem should be carefully scaled to ensure that all the quantities used in the minimization algorithm have approximately equal magnitudes. The scaling of the objective function and constraints are scaled according to the method proposed by Gill et al. [18].

Performance optimization of dry-coolingsystems

29

The dry-cooling system optimization problems have special mathematical characteristics. Variations of the operating variables will affect the operating constraints and the conservation equations only and not the geometrical constraints, whereas variations of the geometrical variables influence all the constraints. For each temperature data set there is a corresponding set of operating variables, conservation equations and operating constraints. Only one set of geometrical variables and constraints are defined. The operating variables of a specific temperature data set affect the operating constraints belonging to the same temperature data set only. The operating variables do not effect the geometrical constraints. However, the geometrical variables effect all the operating constraints of the different temperature data sets. SOLUTION METHODS Various computational methods do exist for solving nonlinear constrained mathematical programming problems. However, it will be extremely inefficient to choose a method without considering the features of the problem that allow it to be solved more efficiently. Problem (2) will be solved by means of the Sequential Quadratic Programming (SQP) method [14], because it has been recognized as one of the most efficient algorithms for solving small to moderately sized nonlinear constrained optimization problems [19]. In a computational environment where one function evaluation can correspond to a full simulation, SQP methods are in general favored due to the fewer function and gradient evaluations required [1 1, 12].

Sequential quadratic programming (SQP)

SQP is motivated by a Newton method for the solution of the K u h n - T u c k e r optimality conditions of a constrained nonlinear programming problem [14]. This can be shown to be equivalent to the solution of a sequence of quadratic programming (QP) subproblems. At each major SQP iteration the values of (x,y) are fixed and the resulting QP subproblem [problem (2) refers] has the form: Minimizes Vf (x,y)Ts + 0.5sXBs subject to the linearized constraints ci(x,y) -1- Vci(x,y)Ts = 0, i = 1..... meq Ci(x,y) q- VCi(X,y)Ts~>0, i = meq q- 1..... mt h(x,y) + Vh(x,y)Xs = 0 (3)

where s is the search direction and B is an approximation to the Hessian of the Lagrange function of problem (2). Convexity of problem (3) is guaranteed by calculating B using the BFGS matrix update formula with Powell damping [20]. At each iteration, the following first-order K u h n - T u c k e r optimality conditions for the QP subproblem must be satisfied: Vf (x,y) + Bs - Vc(x,y)u - Vh(x,y)v = 0 ci(x,y) q- ~Ci(X,y)Ts : 0, i E A

ui[ci(x,y) + VCi(X,y)Ts] = 0, i ui >_0, i :

meq

= meq +

1. . . . .

rnt

"q- 1.... ,mt (4)

h(x,y) + Vh(x,y)Ts = 0,

where v and u are the Lagrange multipliers of the equality and general constraints respectively

30

A.E. Conradieet aL

and A is the set of general constraints which are active at the current iteration (satisfied as equalities). The active set consists of m general active constraints. Let CA present the vector of general active constraints at each iteration. The above first-order necessary conditions for the QP subproblem can be presented in matrix form as follows:
0 = (5)

where VCA and Vh are n m and n r matrices, respectively. B is a dense n n matrix that must be updated, factored and stored at each iteration.

SQ P implementation
Powell has written two FORTRAN subroutines VF02AD [20] and VMCWD [21,22] that implement the Wilson, Han and Powell SQP algorithm. Both these implementations are used as the basis for the present SQP implementations. Several modifications were made to these subroutines which will be discussed below [23]. Quadratic programming routine. The solution of the strictly convex (positive definite) QP subproblem is a major calculation in SQP methods and can affect their overall efficiency. Several numerical procedures can be used to solve the QP subproblem. The feasible point primal QP algorithm normally used with Powell's SQP implementation is replaced by a dual QP algorithm for convex quadratic problems based on the method of Goldfarb and Idnani [24], but modified to employ stable matrix updating procedures [25] as well as to make effective use of the specialized features of the QP subproblems associated with the SQP methods [26-28]. Evaluation of derivatives. VF02AD and VMCWD require that the gradients of the objective and constraint functions be calculated even during line search iterations. Since this information is only required on the last line search point (the new point), a modification was made so that the values of the objective function and the constraints and their corresponding gradients can be calculated only when required. The first-order derivatives that are needed by the SQP method are calculated by means of the finite difference derivative approximations outlined in Dennis and Schnabel [29]. Auxiliary subroutines. The user is responsible for the calculation of the objective and constraint functions, as well as their gradients. The natural structure of the optimization problems are exploited in these routines. An extremely efficient function and gradient evaluation procedure is obtained as a result. Large-scale SQP problems. When the constrained nonlinear optimization problems become large, there are basically two ways to implement SQP methods: either decomposition techniques can be used or the natural problem structure can be exploited [15, 17]. The natural problem structure has been exploited to a certain degree in the infeasible path integrated approach. However, the reasons for advocating decomposition are that the QP subproblems can be very expensive to solve, the storage requirements can become quite large and the problems have a small number of degrees of freedom (difference between the number of variables and the number of active constraints). Decomposition techniques use the equality constraints to 'eliminate' the dependent variables and reduce the size of the QP subproblem that must be solved at each iteration. The resulting reduced Hessian SQP method is formulated and solved in the independent variable space only, while the move in the dependent variable space is obtained directly from a Newton step for the solution of the model equations. The model equations are not exactly solved during each iteration, but an approximation to their solution is computed. A step in the direction of the solution is given such that convergence of the model equations and the optimization problem will be reached simultaneously.

Decomposition of large-scale SQP problems

The computational performance of the SQP methods when applied to large constrained nonlinear programming problems can be considerably improved through a suitable change of bases representation [15, 17]. The new bases vectors are obtained by partitioning the search space into

Performance optimizationof dry-coolingsystems

31

two subspaces which are spanned by the columns of matrices Z and Y, respectively, where Z is chosen so that its columns span the null space of Vh(x,y)T, i.e. Vh(x,y)TZ = 0 (where Z is a n ( n - r ) matrix and Y is a n r matrix). It is assumed that the rank of Vh(x,y)T is r. Several distinct choices of the bases matrices, Z and Y, can be made such that [Y Z] is non-singular (spans the entire search space). These choices include orthonormal, orthogonal and coordinate bases [15]. After the decomposition, the matrix to be updated is a projection of B on to a space whose dimension is given by the number of degrees of freedom of the problem. This number can be conveniently small. The actual projected Hessian matrix is expected to be positive definite at the solution, which is sufficient to guarantee optimality. A reduced Hessian SQP decomposition strategy with coordinate bases is implemented that is well suited to take advantage of the underlying mathematical structure of the optimization problem [15, 16]. This method is based on Powell's original SQP implementation [20]. The variables are partitioned into the dependent variables, yEW, and the independent variables, xeR (" "). The decomposition method that corresponds to the coordinate, bases method is obtained when the following choice is made for Z and Y: Z=

[ _(VyhT)' -,Vxh T ] , Y : iol

s = Ysv + Zsz

(6)

The search direction, s, can be expressed as the sum of its components in the two subspaces, i.e. (7)

Let Q be a non-singular matrix of order (n + m + r) given by

=io o i o]
0 0

(8)

Premultiplying Equation (5) by Q r and substituting for s Equation (7) yields: I yTBy ZTBy
- V e v, Y

yTBZ ZTBZ
- Ve T Z

--yTvc A
--ZTVcA 0

_yTVh0
0

--yTvf-]

-zTvr /

(9)

-Vh~Y

The Lagrange multipliers of the equality constraints, v, can be obtained from the solution of (the first row of Equation (9)): yTBYsv + yTBZsz
- y T V c A u --

yTVh v = - - y T v f .

(10)

Exact values of these multipliers are only required at the solution of the nonlinear programming problem. As the algorithm converges, s ---, 0 and Equation (10) can be simplified to -yXVhv = - y T v f
+

yTVeAu.

(ll)

The second and third rows of Equation (9), namely Z I B Z s z + (ZTBYsv + z T v f ) - ZTVCAU = 0 CA + Vc~Ysy + VeTZsz : 0 (12)

32

A.E. Conradie e t

al.

are the optimality conditions of the following reduced QP subproblem to be solved at each iteration: Minimizesz (ZTBYsy + Z TVf)TS z -~ 0.5sTZTBZsz subject to the constraints

Ci -J- VcTi (Zs Z + Ysy) : 0, i : I ..... meq

ci +

Vc~(Zsz + Y s y ) > 0 , i = meq + 1..... mt

(13)

where s is obtained from the last row of Equation (9): sy = (VhTy)-th. (14)

The search direction for the dependent variables satisfies the set of linearized equality constraints. Problem (13) may be solved to obtain Sz and u (Lagrange multipliers of the general constraints). The Jacobian matrix of the equality constraints, h, can be written as Vh=

[Vxhq V, hJ"

(15)

Duc to the particular choicc of Y, yTVh is now simply equal to Vyh. Any sparsity and structure of the equations that describc the modcl is thus maintained and it is thus straightforward to tailor this algorithm to take advantage of thc particular structure of certain problems classes which can grcatly improvc the pcrformancc of SQP. An additional bencfit is that coordinatc bases require less computational effort per iteration than thc other decomposition methods. Equation (I l) now simplifics to
--Vyhv : --Vyf + and Equation (14) simplifies to sv : -(VyhT)-lh. (17)

VyeAU

(16)

The matrix ZTBy contained in the objective function (Equation (13)) can become too large to store or to compute when r becomes large. The term ZTBYsy can usually be neglected in the objective function when orthonormal or orthogonal bases are used [15, 17]. The original coordinate bases algorithm proposed by Locke e t a l . [16] also did not include this term. However, the performance of the coordinate bases method using an objective function for the QP subproblem that omits this term, depends to a large extent on the partitioning of the variables and can lead to inconsistent performance of the algorithm [15, 17]. Schmid and Biegler [15] propose an improved'coordinate bases SQP algorithm which constructs an approximation of the term ZTBYSy to enhance the performance of the original algorithm. However, due to the fixed variable partitioning employed in the dry-cooling system optimization problems, the matrix ZTBy will be omitted from Equation (13). The Hessian matrix of the reduced QP subproblem, ZTBZ, can be approximated by means of the BFGS quasi-Newton update formula as proposed by Powell [16, 17,20]. ZTBZ is of much smaller dimension than the original full Hessian matrix B. For the particular practical application under consideration (minimization of the power generation cost), the vector of operating variables can be expanded as follows:

Y . [Yl .Y2 . .

yp]T

rY l (1) " " y(lq) y~l) " " y(q) . . .y(1). [ _ 2 . .

y(pq)]T,

(18)

where q is the number of variables contained in each vector, yj, j = 1,-.., p and p x q = r. The partitioning of the variables into dependent and independent variables is also fixed in this par-

Performance optimization of dry-cooling systems

33

ticular practical application The performance evaluation model's equality constraints (conservation equations) can be expanded as follows: h : [hi h z ' " h p ] T :

[h'l'l'"h', q) h~l) " "'h(2q)'" 'h(pl) "" "h(q)] T-

(19)

The structure of the performance evaluation model is such that hi, j = 1.....p, is a function of (x,y/) only. This can be mathematically expressed as: . h l ( x , y / ) = 0 , or hj(~)(x ,Yi ,YJ ' ' ' " Y i (q)/ = 0 , J = 1..... p ; v = ~ (1), (2) Equation (2) can thus be reformulated as Minimizex,y~ ..y,, f ( x , y 1,Y2,''" ,Yp) subject to the constraints 1..... q. (20)

ci(x,Yl,y2,...,yp)

= 0, i = 1,---,meq i : meq + 1 , " .,mr (21)

ci(x,yl,Y2,...,yp)>O,

h/(x,yj) : 0, j - - 1,.--,p. Equation (15) can now be rewritten as I Vxhl V~h2 ---

Vxhp 1

Vh = L V y h j =

Fv~hl i
L

V.h:
0

o]
h2

(22)

Vylh p
(23)

and

yTVh = Vyh :

Furthermore, sv can be obtained by solving the following systems of equations

[Vy Olls lhil 1

0 Vy:h2 T 0

Vylhl

0 Vy: h2

s~)

= _

(24)

Vyphp

which decomposes into p smaller systems: Vy hTs~ ) = - h i , i = 1,... ,p. Z can be determined by solving (VyhV)A = - V x h x for A = -(VyhV)-'Vxh T, i.e.

VyIh T i

..

Vy2hT

"-"

Ii l
T

A 2

IV~hT7 V~h~/
(25)

. Vyphl,
"

L V~h~ J

34 Thus,

A.E. Conradieet al.

Z = [I Al A2 " - Ap].

(26)

The above problem simplifications are possible due to the specific structure of the original engineering optimization problem. The block diagonal structures in Equations (24) and (25) provide great computational advantages (e.g. solution techniques and storage requirements). More detail of the SQP-based algorithms and their implementation can be found in [23].

COMPUTATIONAL RESULTS In this section we present a numerical illustration of the advantages, both computationally and functionally, of the application of formal minimization techniques to cost-optimal dry-cooling system design. Both the conventional (modified) SQP and the SQP decomposition techniques are illustrated for large-scale optimization problems (Equation (2)). The SQP-based computational procedures are implemented in different computer programs. The programs were written in double precision FORTRAN 77, developed and run on an ALPHA 3000-800 computer. NDOPT1 and FDOPT1 implement the conventional SQP algorithm (VMCWD), while NDOPT2 and FDOPT2 implement the coordinate bases SQP decomposition strategy (FD denotes forced draft, ND denotes natural draft). All the different implementations require the same input data and produce the same output. The objective function, constraint and gradient evaluations as well as the model simulations are performed similarly. The following acronyms and definitions are introduced to measure the computational efficiency of the solution techniques: 1. TOTCPU: the total CPU time (in seconds) used until the required convergence is reached. 2. QPCPU: the total CPU time (in seconds) used to solve the QP subproblems of the SQP methods. 3. NFEVAL: the number of objective function, constraint and their corresponding gradient evaluations performed during program execution. 4. NITER: number of major SQP iterations. The example problems and the simulation models used for both the forced draft direct aircooled condenser and the natural draft indirect dry-cooling tower are described in detail in [13]. An extensive list of input data is required for the dry-cooling system optimization. These can be conveniently categorized into economical data, site data, turbo-generator performance characteristics, fan performance characteristics and heat rejection system data. With the minimization techniques we compute the combination of operating and geometrical variables that will minimize the ratio of the total annual cost of the dry-cooling system (attributed to its performance) to the annual net energy output of the turbo-generator set to which it is coupled. The cost estimation procedure and the list of decision variables and constraints are listed in Appendix A and Appendix B, respectively. For these example problems the following information is applicable: 1. Natural draft indirect dry-cooling tower: Geometrical variables used: Hs, //3, ds, d3, t"/tb(max),Lt, 0b, do, it, tr, dr, tf, Pf, Pt. The total number of variables and constraints are 150 and 491, respectively, while the number of dependent variables and model simulation equations are both equal to 136. 2. 3. 4. 5. Forced draft direct air-cooled condenser: Geometrical variables used: H3, dF, OF, NF, ntb(max), Lt, Ob, do, It, tr, dr, tf, Pf, Pt. The total number of variables and constraints are 286 and 700, respectively, while the number of dependent variables and model simulation equations are both equal to 238.

P e r f o r m a n c e o p t i m i z a t i o n o f d r y - c o o l i n g systems Table 1. Optimum natural draft indirect dry-cooling tower system Initial value
H3 H5 d3 d~ ncu(,nax) L, On
cto

35

Optimum
18.84 m 136.4 m 97.64 m 70.8 m 31 9.52 m 60.0 27.26 mm 1.2 mm 1.0 m m 86.63 0.205 m m 2.224 mm 96.77 mm 2083.90 G W h 2877.2l G W h 0.0524 S/k W h

tt
t~ dlt~ Pf P, E,, Qt C o. . . .

13.67 m 120.0 m 82.958 58.0 m 39 15.0 m 30.75 25.4 m m 1.9 m m I. 1 m m 57.2 m m 0.5 m m 2.8 mm 58.0 m m 2074.34 G W h 2886.39 G W h 0.0530 S/k W h

Tables l and 2 list the results of the example problems for the proposed cost-estimating procedure. The power generation cost is reduced by 1.13% in the case of the natural draft indirect dry-cooling tower and 2.57% in the case of the forced draft direct air-cooled condenser. The optimum geometry and performance characteristics of the dry-cooling systems ensure the turbine performs in a very efficient region of its performance characteristics. When proper cost coefficients are available, the performance and annual costs of the optimal forced draft direct aircooled condenser and the natural draft indirect dry-cooling tower can be directly compared to each other. In both examples, fairly realistic and practical results were obtained. It should again be stressed that realistic results depend very strongly on a realistic cost estimation model and cost coefficient values. The operating point optimizations performed by Buys and Kr6ger [11, 12] found the optimum fin thickness to be impractically thin. The present study uses a more extensive simulation model and cost-estimation procedure which contribute to differences in the optimal solution. In this study the fact that the performance of dry-cooling systems decrease at high ambient temperatures is recognized. Furthermore, the auxiliary power consumption (fans and/or pumps) is also taken into account. For fixed turbo-generator characteristics the cost-optimal dry-cooling tower that causes the cheapest power to be produced for a given annual frequency of ambient temperatures is computed, while taking the above-mentioned dry-cooling system characteristics into account. This can be viewed as a system optimization, because the interaction of the drycooling system and the turbo-generator unit forms an integral part of the optimization process, Designing a dry-cooling system in this way is different from the conventional means of design that is purely based on fixed ambient and process fluid conditions. Finding the best design over a range of operating conditions will definitely result in a more realistic and reliable design.
Table 2. Optimum forced draft air-cooled condenser system Initial value
H3 30.0 m

Optimum
35.91 m 10.48 m 12.0 36.0 75.0 rpm 131.0 10.98 m 35.0 36.8 m m 1.2 mm 1.0 mm 92.33 mm 0.216 m m 3.06 mm 1.75 mm 96.1 I mm 2079.41 GWh 2877.25 G W h 0.053 S/k W h

dl:
0v Nv ntb(.,~xl Lt 0~ do tt tr dr tf Pc~l) Pf(21 P~ En Qt Cp . . . . .

9.145 xn
16.0 ~' 100.0 rpm 153.0 10.0 30.0' 38.1 mm 1.5 mm 1.1 mm 69.9 m m 0.35 mm 3.63 mm 2.54 m 76.2 mm 2058.35 G W h 2877.14 G W h 0.0544 S/kWh

36

A. E. C o n r a d i e et al. Table 3. Performance of optimization algorithms

Program name NDOPTI NDOPT2 FDOPT 1 FDOPT2

NITER 83 148 164 267

NFEVAL 89 150 167 269

TOTCPU [s] 339.33 141.5 6094.8 1234.5

QPCPU [s] 272.66 1.22 5613.3 12.36

Although this method of optimal design involves a considerable amount of computation, the efficient SQP-based computational algorithms developed and implemented in this study ensure that this procedure is not excessively time consuming, The performance of the SQP techniques is listed in Table 3. The computational efficiency trends observed in this table can be explained as follows: I. FDOPT1 and NDOPTI: the optimization problems are solved in the total decision variable (dependent and independent) space and require the solution of large QP subproblems. The order of the Hessian of the Lagrange function is equal to the total number of decision variables. Large computational overheads and storage are needed to solve these optimization problems. The solution of the QP subproblems consumes most of the CPU time required to solve the problem. 2. FDOPT2 and NDOPT2: the optimization problems are solved in the independent variable space and significantly reduce the size of the QP subproblems. The order of the corresponding reduced Hessian matrix is equal to the number of independent variables. The computational overhead needed to solve the QP subproblems is drastically reduced. These drastic reductions in QPCPU indicate that little can be done to further improve the QP algorithm's performance. Almost all the CPU time is consumed by the function and gradient evaluations. More SQP iterations and function and gradient evaluations are needed than in the conventional case. Although the order of the Hessian matrix is enormously reduced, additional storage is required for the vectors and matrices used in this decomposition method. Numerical comparisons based on the solution of a number of dry-cooling system optimization problems demonstrate the effectiveness of the SQP decomposition strategies, compared to conventional SQP implementations. On average, the solution of the QP subproblems (the most time-consuming part of the SQP algorithm) require 300 times less CPU time, while the total CPU time is reduced by a factor of 6 [23]. Similar trends have been observed by [15, 17]. The significant improvements that the reduced Hessian SQP method has over the original SQP methods illustrate the benefit of tailoring the solution algorithm to take advantage of the mathematical structure of the model. The optimization is not complete when the solution is obtained; in fact the solution only serves as the basis of the most important parts of the study: solution validation and sensitivity analysis. The information about the state of the system in the neighborhood of the solution provides key insights into the following important factors: the constraints active at the solution that limit further improvement of the system, the dominant cost terms that should be refined, and the prescribed parameters that can be improved.

CONCLUSIONS In this study we develop and implement computational procedures which are very efficient and practical when applied to sophisticated economic design and performance optimization computations of dry-cooling systems as found in power plant applications. The computer programs were extensively tested on a variety of dry-cooling system problems and provide a reliable practical tool with which realistic answers and trend information can be obtained. Realistic optimization results can only be obtained if the cost-estimating procedure represents realistic operating and capital cost structures. SQP methods have great potential for routinely solving complex engineering optimization problems. The infeasible path integrated approach was very reliable and efficient when applied to the optimization problems in this study. The modifications to Powell's SQP implementation,

Performance optimization of dry-cooling systems

37

VMCWD [22], i.e. the new QP routine, more efficient derivative evaluation and the partial exploiting of the problem structure in the auxiliary routines all improved the performance of VMCWD. This modified version of VMCWD performed efficiently on dry-cooling system optimization problems with a small to moderate number of variables. For larger optimization problems the storage requirements for the Hessian matrix as well as the computational expense involved in solving QP subproblems can become prohibitive. In this study we overcome these problems by developing and implementing the coordinate bases SQP decomposition strategies [15, 16] These decomposition strategies are particularly suited to take advantage of the block diagonal structure of the Jacobian matrix. The resulting optimization problem is solved in the independent decision variable space only. Numerical comparisons based on the solution of a number of dry-cooling system optimization problems demonstrate the effectiveness of the reduced SQP decomposition strategies, compared to conventional SQP implementations. Although more function and gradient evaluations are used by these decomposition strategies, the total CPU time is significantly reduced. The coordinate bases decomposition strategy is computationally the cheapest decomposition method and allows us to exploit the particular structure of the model when calculating s, [15]. This method is therefore preferred for large-scale optimization problems. The coordinate bases algorithm suffers no ill-effects related to the fixed variable partitioning implemented in the drycooling system optimization problems. Although several authors claim that this method can give inconsistent results, [15, 17, 30] it solves our large-scale optimization problems very accurately, efficiently and reliably. It is worthwhile considering an approach that takes advantage of the problem structure during its solution because of the potential to outperform the conventional approaches. The resulting tailored solution procedures are very effective for performing rather sophisticated performance evaluation and optimization computations in a relatively short time.
Acknowh, dgements--The financial support from the Water Research Commission is gratefully acknowledged.

REFERENCES 1. J. P. Rossie and E. A. Cecil, Research on dry-type cooling towers for thermal electric generation: Part 1 and II. EPA Report, Water Pollution Control Research Series 16130 EES (1970). 2. E. C. Smith and M. W. Larinoff, Power plant siting, performance, and economics with dry cooling tower systems. Proceedings of the American Power Conference 32, 544-572 (1970). 3. L. G. Hauser, K. A. Oleson and R. J. Budenholzer, An advanced optimization technique for turbine, condenser, cooling system combinations. Proceeding of the American Power CoJ!ference 33, 427-445 (1971). 4. E. S. Miliaras, Power Plants with Air-cooled Condensing Systems. MIT Press, Cambridge, MA (1974). 5. B. C. Fryer, A review and assessment of engineering economic studies of dry cooled electrical generating plants. BNWL (1976). 6. M. Choi and L. R. Glicksman, Computer optimization of dry and wet/dry cooling tower systems for large fossil and nuclear power plants. MIT Energy Laboratory Report No. MIT-EL 79-034. Cambridge, MA (1979). 7. K. W. Li and W. Sadiq, Computer-aided optimization of cooling systems, A S M E International Computers hi Engineering CoJl~,rence 3, 165 171 (1985). 8. G. A. Veck and P. J. E. Rubbers, Economic evaluation of dry cooling: the options. VGB Conference, 9 13 November 1987. Johannesburg, South Africa, pp. 80-89 (1987). 9. G. Lovino, H. Hattoni, L. Hazzocchi, R. Senis and R. Vanzan, Optimal sizing of natural draft, dry cooling towers for ENEL combined cycle power plants. In A Futurejor Energy, FLOWERS "90, Proceedings of the Florence World Energy Research Symposium, Firenze, Italy, 28 May-1 June 1990 (Edited by S. S. Steceo and M. J. Moran). Pergamon Press, New York, pp. 669-680 (1990). 10. J. G. Ecker and R. D. Wiebking, Optimal design of a dry-type natural-draft cooling tower by geometric programming. Journal q[" Optimization Theory and Applications 26(2), 305-323 (1978). 11. J. D. Buys and D. G. Kr6ger, Dimensioning heat exchangers for existing dry cooling towers. Energy Conversion & Management 29(1), 63 71 (1989). 12. J. D. Buys and D. G. Kr6ger, Cost-optimal design of dry cooling towers through mathematical programming techniques. A S M E Journal ~f Heat Tran.~'[~'r 111(2), 322-327 (1989). 13. A. E. Conradie and D. G. Kr6ger, Performance evaluation of dry-cooling systems for power plant applications. Applied Thermal Engineering 16(3), 219 232 (1996). 14. R. Fletcher, Practical Methods c~['Optimization, 2nd ed. J. Wiley and Sons, New York (1987). 15. C. Schmid and L. T. Biegler, Acceleration of reduced Hessian methods for large-scale nonlinear programming. Computers and Chemical Engineering 17(5/6), 451-463 (1993). 16. M. H. Locke, A. W. Westerberg and R. H. Edahl, An improved successive quadratic programming optimization algorithm for engineering design problems. AIChE Journal 29(5), 871-874 (1983).

38

A.E. Conradie et al.

17. S. Vasanthrajan and L. T. Biegler, Large-scale decomposition for successive quadratic programming. Computers and Chemical Engineering 12(11), 1087-1101 (1988). 18. P. E. Gill, W. Murray and M. H. Wright, Practical Optimization. Academic Press, London (1981). 19. W. Hock and K. Schittkowski, A comparative performance evaluation of 27 nonlinear programming codes. Computing 30, 335-358 (1983). 20. M. J. D. Powell, A fast algorithm for nonlinearly constrained optimization calculations. In Numerical Analysis, Dundee 1977, Lecture Notes in Mathematics 630 (Edited by G. A. Watson). Springer-Verlag, Berlin, pp. 144-157 (1978). 21. M. J. D. Powell, Extensions to subroutine VFO2AD. In System Modeling and Opthnization, Lecture Notes in Control and Information Sciences, No. 38 (Edited by R. F. Drenick and F. Kozin). Springer-Verlag, Berlin, pp. 529-538 (1982). 22. M. J. D. Powell, VMCWD: a Fortran subroutine for constrained optimization. Report DAMTP-1982/NA4, University of Cambridge (1982). 23. A. E. Conradie, Performance optimization of engineering systems with particular reference to dry-cooled power plants. Ph.D dissertation, Department of Mechanical Engineering, University of Stellenbosch, Stellenbosch, South Africa (1995). 24. D. Goldfarb and A. ldnani, A numerically stable dual method for solving strictly convex quadratic programs. Mathematical Programming 27, 1 33 (1983). 25. P. E. Gill, G, H. Golub, W. Murray and M. A. Saunders, Methods for modifying matrix factorizations. Mathematics of Computation 2g(126), 505 535 (1974). 26. P. E. Gill, W. Murray, M. A. Saunders and M. H. Wright, Some issues in implementing a sequential quadratic programming algorithm. SIGNUM Newsletter 20(2), 13 19 (1985). 27. P. E. Gill, W. Murray, M. A. Saunders and M. H. Wright, Considerations of numerical analysis in a sequential quadratic programming method. In Numerical Analysis, Guanajuato, 1984, Lecture Notes in Mathematics 1230 (Edited by J. P. Hennart). Springer-Verlag, Berlin, pp. 46-62 (1986). 28. C. Schmid and L. T. Biegler, Quadratic programming methods for reduced Hessian SQP. Computers and Chemical Engineering 18(9), 817-832 (1994). 29. J. E. Dennis and R. B. Schnabel, Numerical Methods .fin" Unconstrabwd Optimization and Nonlinear Equations, Prentice-Hall, Englewood Cliffs, NJ (1983). 30. L. T. Biegler, J. Nocedal and C. Schmid, A reduced Hessian method for large-scale constrained optimization. SIAM Journal of Optim. 11, 74-81 (1994). 31. P. A. Ard, C. H. Henager, D. R. Pratt and L. E. Wiles, Cost and cost algorithms for dry cooling tower systems, BNWL-2123, Richland, WA (1976). 32. K. K. Humphreys, Jelen's Cost and Optimization Engineering, 3rd ed. McGraw-Hill, New York (1991). 33. E. Richter, Untersuchung der Str6mungsvorg/inge am Austritt von Naturzugkiihltiirmen und deren Einfluss auf die Kfihlwirkung. Dr. Ing. Thesis, TU Dresden (1969). 34. D. G. Kr6ger, Fan performance in air-cooled steam condensers. Heat Recovery Systems and CHP 14(4), 391399 (1994). 35. D. E. Briggs and E. H. Young, Convection heat transfer and pressure drop of air flowing across triangular pitch banks of finned tubes. Chemical Engineering Progress Symposium Series 59(41), 1 10 (1963). 36. K.K. Robinson and D. E. Briggs, Pressure drop of air flowing across triangular pitch banks of finned tubes. Chemical Engineering Progress Symposium Series 62(64), 177-184 (1966).

APPENDIX

A: D R Y - C O O L I N G

SYSTEM

COST ESTIMATION

PROCEDURE

Capital and operating cost estimation procedures based on cost information obtained from different sources [31,32], are used for both forced draft and natural draft dry-cooling systems. Analytical functions express the annual cost in terms of the various geometrical and operating parameters of the dry-cooling systems. Cost weighting factors are introduced on several occasions in the proposed cost structure to take into consideration estimation uncertainties, overheads and other related costs. The proposed cost estimation structure is not necessarily unique and simpler or more complex models may be employed in practical designs. Typical values of the cost coefficients are shown in brackets. Cooling system capital cost estimation Heat exchanger bundle cost (1) Tube cost per unit tube length, Ct Ct = CtmM, + Ct,, (S/m). M t is the tube mass per unit tube length (kg/m), Ctu is the tube added unit cost (2 S/m) and Ctm is the tube material cost (0.8 $/kg). (2) Cost of fins per unit tube length, Cf Cf : CfmMf + CfuLfb (S/m). Cfu is the fin material added unit cost per unit tube length (0.2 S/m), Co,1 is the fin material cost (4 $/kg), M~ is the fin material mass per unit tube length (kg/m) and Lib is the length of fin strip per unit tube length (m/m).

Performance optimization of dry-cooling systems (3) Total finned tube cost per unit tube length, Cft Cft = (Ct + Cf)Wrt (S/m). Wft is the finned tube cost weighting factor (1.3-forced, 1.2-natural) (4) Heat exchanger bundle header and frame cost, Ch Ch = Cft Ltntb Wh (S/bundle). Wh is the header and frame cost weighting factor (0.2). (5) Bundle assembly cost, Cb,,
Cba : Cbat,~ltb (S/bundle).

39

Cb~,t is the assembly cost per finned tube (25 S/finned tube). (6) Total heat exchanger bundle cost, Ch~
Che : (CftLtntb + Ch q- Cba)Whe (S/bundle).

Wh~ is the heat exchanger bundle cost weighting factor (installation) (l.3-forced, 1.2-natural). Fan system capital cost It is assumed that the height of the fan casing is 0.1 dye, the inlet bell height is 0.15 d w and the inlet bell diameter is 1.2 riFe. (1) Fan capital cost, CF
CF = CFt, AF + CFf (S/fan).

Cv, is the fan unit cost (110 $/m2), Cvf is the fan fixed cost (650 S/fan) and AF is the fan sweep area (m2). (2) Fan casing and inlet bell cost, Cvc
CFc : (,4Fc -{- AFb)CFc b (S/fan).

Cv~b is the fan casing and inlet bell unit cost (10 $/m2), Ave is the fan casing circumference area (m 2) and Avb is the inlet bell circumference area, approximated by a conical frustrum (m2). Ave = 0.1xd~-c (m -~) Avh "~ 1.17zdFc~O..O225d~-c - 0.01 ~i~_,~= 0.122984rcd~,c (m2). (3) Fan safety screen cost, Cv~ Cw = Cv~Aw (S/fan) CF~ is the safety screen cost (2.5 S/m-),' A w is the safety screen area (m 2) and A w = 0.25rc(l.2&~)-~ = 0._36rcd?.~ (m2). (4) Electric motor cost, C~,,, Cem = C~mf + s~mC~m~P~. (S/motor). C~,nl is the electric motor fixed cost (400 S/motor), C~mu is the electric motor unit cost (120 $/kW) and P~ is the electric power input to the electric motor (kW). A safety factor, s~m (I.25), against electric motor undersizing is also introduced. (5) Speed reducer cost, C~..

40
C~r = (CF + Cen~)Wsr (S/speed reducer).

A . E . Conradie et al.

W~r is the speed reducer cost weighting factor (0.4). (6) Electric wiring and switching cost, Cws C,,,s = Cem Wws (S/fan). Wws is the wiring and switching cost weighting factor (0.1). (7) Total fan system cost, Cvt
CF, = (Cv + Cvc + Cvs + C~m + Gr + Cw~)WF (S/fan). WF is the fan system cost weighting factor (1.25).

Forced draft cooling tower structure (1) Cost of land, excavation, foundations and construction, Cc Cc = (C1 jr_ Court3 + CFpl)mFpl mc (S). Ct is the land, excavation and foundation cost (15 $/m2), Ccu is the construction unit cost (17.5 $/m3), Cvpl is the fan platform cost (75 $/m2), AFpI is the fan platform area (m 2) and W~ is the construction cost weighting factor (1.25). Natural draft cooling tower shell (hyperbolic reinforced concrete) The cooling tower shell will be approximated by means o f two conical frustums, one above and the other below the throat. A constant shell thickness is assumed. The volume of the conical frustrum can be calculated as follows:
H )(a~ + dtds + d~) + O.O83333rr(nt - n , ) ( d t + +

with H s - H t = 0.25(Hs-H3), H t - H 3 = 0.75(tI5-H3) and d5 = 1.05 dr. These relationships are typical in most of the existing hyperbolic cooling towers found in practice. The volume of concrete in the tower shell can be calculated with the aid of the above equation. It is further assumed that the tower base angle, #5b, is fixed at 70 and that the tower support length can be expressed as Lt~ = H3/sin(~bb). (1) Cost of land, excavation and foundation, Ci
Ci = G , A h ($),

(-'lu is the land, excavation and foundation cost (70 $/m 2) and Ab is the tower base area (m2). A b = 0.25re(d3 + 2H3/tan qSb)2. (2) Cost of the tower shell, CCt

C~t, is the cost of the reinforced concrete used in the shell (250 $/m 3) and Vt~ is the volume of reinforced concrete in the tower shell (m3). (3) Cost of the heat exchanger bundle platform, Ch~v~ Chepl = CpluAhepl ($). Cpl~ is the platform unit cost (100 $/m 2) and Ahcpl is the heat exchanger platform area (m2). (4) Cost of the tower supports, Cu Cts ~ Cts.uLtsdtsnts (S). Ct~u is the tower support unit cost (175 $/m3).

Performance optimization of dry-cooling systems

41

(5) Total construction cost, Cc

Cc = (G + Cc~ + Chep~ + G~) wc ($)

We is the structural maintenance cost weighting factor (1.25). Steam/condensate distribution system costs (direct system) (1) Distribution system cost, Cud

Csd = (Ch~ + Cw) W~d ($).

W~a is the steam/condensate distribution system cost weighting factor (0.2). Circulation system costs (indirect system) Two 50%-duty pump units will be considered in this analysis. (1) Pump capital cost, Cpunlv

Cpmp = 2(Cpf + C~oP,,,) ($).

Cpf is the pump fixed cost (2300 S/pump), Cpu is the pump unit cost (180 $/kW), Pw is the required pumping power
(kW). (2) Electric motor cost, C~,~1 Cem = 2(Cemf + semCemuPw) ($).

C~mf is the electric motor fixed cost (400 S/motor) and C~m~ is the electric motor unit cost (120 $/kW).
A safety factor, s~,, (1.25), against electric motor undersizing is also introduced. (3) Electric wiring and switching cost, Cw.~ C,,~ = C~m W,,~ ($). Ww~ is the wiring and switching cost weighting factor (0.1) (4) Total pump system cost, Cm~

Cp~, = (Cv.,,,p + ~,,, + Cw~) Wp~ ($).

Wp~ is the pump system cost weighting factor (1.25). (5) Piping and valves cost, Cp,, Cv,~ = ChJzb Wp,. ($). Wp,, is the piping and valves cost weighting factor (0.25) Steam condenser cost (indirect system) Condenser cost, C~

C~ = Cc~,vi~,lW~,. ($).
W~c is the steam condenser cost weighting factor (0.075).

42 Cooling system operation cost estimation

A . E . Conradie et al.

Fuel cost The total annual fuel cost, Cft, can be expressed as
I Cf! -~- CfavqplamPg~.

qplant is the overall plant efficiency, Pg is the gross power output of the turbo-generator (MW), Cr~,v, levelized unit cost of the fuel (0.00597 S/M J) and ~ is the number of operating hours (h). Fan operating cost The annual operating cost of the fan, Cvo, can be expressed as CFo = CeavPeT. Cc~,,, levelized electricity cost (0.10571 S/kWh), Pe is the input power to the fans (kW) and z is the number of operating hours (h). Pump operating cost The annual operating cost of the pump, Cpo, can be expressed as Cpo = &Ce~,v~. C~av levelized electricity cost (0.10571 S/kWh), Pe is the input power to the pumps (kW) and ~ is the number of operating hours (h). Fixed charges The cost o f the fixed charges or the annual owning cost of an investment as a percentage of the investment (consisting of interest, taxes, insurance and depreciation.), CFCR, can be expressed as F C R Ccapita I-

CFC R =

F C R (0.2) is the levelized fixed charge rate and C~,p,taJ is the total capital cost ($). Maintenance costs The maintenance cost weighting factors are assumed to be levelized values over the operating lifetime of the plant. (1) Heat exchanger bundle maintenance cost, Chum
Ch,,, = lYhe,nCt~ (S/bundle),

WI,~,,, is the heat exchanger bundle maintenance cost weighting factor (0.01). (2) Fan system maintenance cost. C~m
Cvm = Cvt Wvm ($/thn). Wr:,,, is the fan system maintenance cost weighting factor (0.03).

(3) Structural maintenance cost, ('c,,,

C~,,, - C~ W~m ($).

1~.,,~ is the structural maintenance cost weighting factor (0.0005). (4) Distribution system maintenance cost. C~a,,,

Performance optimization of dry-cooling systems

43

Csdm = Csd H/sdm ($)-

W~dm is the steam/condensate distribution system maintenance cost weighting factor (0.005). (5) Pump system maintenance cost, Corn

CpmCp~ Wpm (S),

Wpm is the pump system maintenance cost weighting factor (0.03). (6) Piping maintenance cost, Cp~m C~vm = Cp,, Wp,,m ($). Wp,~m is the piping and valves maintenance cost weighting factor (0.005). Total costs The total annual cost of the cooling system is:
Ctotal ~ Coperating @

CFCR (S/annum).

The cost of power generation (attributed to the cooling system) is: C o...... = Ctotal/En (S/kWh) where E,, is the net annual power output. APPENDIX B: D E C I S I O N VARIABLES AND CONSTRAINTS

The dry-cooling system simulation equations are listed in [13]. The decision variables and constraints that are required by the dry-cooling system optimization problems are presented here. The operating variables, conservation equations and operating constraints must always be present during the optimization process. The geometrical variables can either be variable or fixed at a constant value during the optimization process. The corresponding geometrical constraints can either be applicable or completely ignored. The feasibility inequality constraints introduced by the operating variables (bounds), will not necessarily be active at the optimum solution. Natural draft indirect dry-cooling towers Operating variables: m~,, T,4, T,~o, T~,,. Geometrical variables: Hs, H3, ds, d3, ntblmax), Lt, 0bFeasibility inequality constraints (operating constraints):
117a

~>EI

Till -it- g'2

Ta4 ~ Z~i -/i-~

T~,4 + Two > T,,I + T,,i + ~:4 T,,I + ~:s-< Tt, -< Ttvu - ~:5 The small constants, e.i, are used to introduce the equality sign into the feasibility inequalities. The bounds on the operating variables are safeguards against the violation of the physical laws. General operating constraints To prevent cold inflow [33], the following condition is proposed as an inequality constraint

(m:,/As)2/[p~,s(p~,6 - 0,,s)gds] < 3.05.

To prevent fouling inside the tubes, a lower velocity limit can be specified: v,, > v,,i, (Vwl Conservation equations Q~, = Q,,. = Q~ (energy - heat transfer equation). 1 m/s).

Ap~t,d~,rt -- Ap~y,tcm (momentum -- draft equation),

Geometrical constraints

44

A . E . Conradie et al. The hyperbolic cooling tower dimensions are limited to the practical values: O.05<H3/Hs<O.15 0 . 6 4 < d 3 / ( H 5 - H3) _< 1.13 1.37_<a(~/d5 < 1.77. The heat exchanger bundle width is limited by transport requirements. Who -- Wb > 0 , (Wbu = 3 m). The finned tube length is limited by structural, transport and construction requirements. Ltu - Lt ~ 0 , The bundle semi-apex angle cannot exceed 9 0 . 90 - 0b _>0.
(Ltu = 2 5 m ) .

The following assumptions and relations are used to obtain values for the geometrical parameters linked to the optimization variables: n b = u~tnci~[4L~(Wb sin 0b + n~P, cos 0b)] where %t is the fraction of the tower area covered by heat exchanger bundles. Iltb : ntb(max)//r -- nr/2; Wb = ,"/tb(max)Pt. The following parameters are assumed to be constant during the optimization process: CDts, tits, /'/ts, /'is, :~ct, g/r, k.'f. Forced draft direct air-cooled condensers Two rows o f finned tubes with different performance characteristics will be considered in this analysis. Operating constraints: m~,, Taoo), Tao(2), Tvi, Ttv, m~(l), mc~2). Geometrical constraints: H> dF, 0v, Nv, ntb0nax), Lt, 0b. Feasibility inequality constraints (operating constraints):
Dial _]_ t;[ .~/71 a
~Ola u _

~;1

Illc(D ~"/;2, Illc(2) ~E2 T~i(l~ + e3 _</~ao~t)~< Taol2i - g3 Tao(2) -< Tvi -- g3 Tvi ~ Try - g4 Twl 4- g5 --<Try ~< T,,.u - ~:5. General operating constraints The tip speed of the fan blade is usually limited to 60 m/s to control noise. 60 - rcd~ Nv >_O. The steam .velocity at the finned tube inlet is limited to 120 m/s. Vvi(i) <~ 120, Vvi(2 ~<_120. ) Conservation equations Qa(]) ~ Qcl), Qa~21 = Qc(2) nv(Qa(I) + Q~(2)) = Qtg (energy - heat transfer equation) Ap~y~te,,1 = A p w ( m o m e n t u m - draft equation). Geometrical constraints The ratio o f the fan unit's width and the fan unit's length to the fan diameter must be kept within practical limits to assure a reasonable heat exchlmger normal approach velocity (2-4 m/s). The ratio of the total heat exchanger frontal area to the fan casing area usually ranges from 1.8 to 2.6. 1.2dv _<2Lt sin Ot, _< 1.5dv

Performance optimization of dry-cooling systems

45

1.2dr ___0.5nb Wb _< 1.5dv.

The fan blade angle must lie within the limits imposed by the fan performance characteristics. 0Fmin <.~OF < 0Fro.... The heat exchanger bundle width is limited by transport requirements.

wbu- wb>0 (who =3m).

The correlations of the downstream loss coefficient [34] are valid within the following limits: 20<0b<35 0.17886 -- (0.5dshLt)(sin 0b + LwLt)_>0 0.09033 - Lw/ Lt >_0 0,303 - dsh/Lt >0 (Lw = 0). The following assumptions and relations are used to obtain values for the geometrical parameters linked to the optimization variables: ntb(2) = ntb(max)~ Wb ~ ntb(maxlPt

dvc -----1.005dF; dvh = 0,165dv; Hvc = 0.1dvc; Hb ----0.15dye; db = 1.2dvc.

The following parameters are assumed to be constant during the optimization process: Lw, Kts, dsh, Kso, Kup, Koo, nv, nVb, nFr, qVd, nr, rib. Finned tubes R o u n d tubes with circular fins are used in the optimization study. The airside performance characteristics as a function of the geometrical variables can be found in [35, 36]. Geometrical variables: do, tt, tr, dr, tf, Pf, PtFor round tubes with circular fins, the following geometrical inequality constraints are imposed for both the natural draft indirect dry-cooling tower and the forced draft direct air-cooled condenser: The finned tubes must be kept from touching. Pt - df >_0. The fin diameter must exceed the fin root diameter (extended airside surface area is required).
dr - d r > _ 0 .

Due to structural considerations, the fin root thickness and the tube thickness must exceed their lower limits. tr - trl _>_0, I t - It[ > 0 (lrl = 1.0 m m , /tl = 1.2 ram). For effective cleaning and where airside fouling is of significance, the fin pitch must exceed the imposed lower limit. Pf - Pn>-0 (Pn = 1.75 mm). The fin thickness is limited by the manufacturing processes. tf -- tn >_0 (tfl = 0.05 mm). The fin pitch must exceed the fin thickness. Pf - If_>0. The geometrical parameters that are linked to the geometrical variables are: PL = 0.866Pt; d i = do - 2tt; dr = do + 2tr.