Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Efficient optimized design of solar power

tower plants based on successive response


surface methodology

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


..............................................................................................................................................................

Yan Luo1 , Yue Hu2 and Tao Lu1, *


1
School of Mechanical and Electrical Engineering, Beijing University of Chemical
Technology, 15 Beisanhuan East Road, Chaoyang District, Beijing 100029, China; 2 School
of Science, Hubei University of Technology, 28 Nanli Road, Hongshan District, Wuhan
430068, China
............................................................................................................................................
Abstract
Recently the optimal design of the solar power tower (SPT) plants have attracted increasing attention.
In this paper, an improved algorithm combing the successive response surface methodology (SRSM) and
simulated annealing (SA) global algorithm was proposed to achieve the efficient optimization design of
the molten salt SPT plant with 2650 heliostats in Sevilla, Spain. Based on the traditional response surface
methodology (RSM) and the adaptive domain reduction method, the SRSM was established to surrogate
the complex thermo-economic model of the SPT plant to the updated approximation function, which was
a high-order polynomial form to define the relationship between the objective parameter of the levelized
cost of energy (LCOE) and 12 design variables related to the 4 subsystems of the SPT plant. After that, the
SPT plant optimization design was performed by the SA global algorithm on the basis of SRSM. According
to a comparison between the results obtained by the proposed method and the actual model-based global
algorithm, the high accuracy and low computation time of the proposed strategy was proved.

Keywords: solar power tower plants; levelized cost of electricity; successive response surface methodology;
adaptive domain reduction; global optimization
∗ Corresponding author:
lutao@mail.buct.edu.cn Received 4 March 2019; revised 3 May 2019; accepted 8 May 2019
................................................................................................................................................................................

1 INTRODUCTION to obtain the optimal heliostat layout and receiver size with a
minimum LCOE of the central receiver solar plant. Soltani et al.
Compared with other concentrating solar power technology [6] solved the multi-objective optimization problem of a solar
options, solar power tower (SPT) technology is currently a more combined system using the genetic algorithm. Then, the opti-
cost-effective option because of its high operating temperatures mal power block operating parameters and heliostat field layout
and large power station scales [1]. SPT technology has been widely were achieved. Starke et al. [7] optimized a solar-hybrid system
applied in many engineering projects such as PS10, PS20 and with an evolutionary optimization algorithm. Although the global
Gemasolar [2]. To become one of the most appropriate substitutes algorithm can obtain accurate SPT plant optimization design,
of coal-fired power plant the levelized cost of energy (LCOE) of the enormous computation time of the optimization process is
the SPT plant should be further decreased [3]. challenging due to the complex heliostat field model and high-
The SPT plant is generally composed of a solar receiver, helio- dimension optimization process.
stat field, thermal storage system and power block. Based on the Consequently, to solve the above problem, some other strate-
SPT plant subsystem models, the entire SPT plant design is always gies were proposed to reduce the computation time of the opti-
globally optimized to obtain the optimal design parameters of mal SPT plant design. Kistler [8] simplified the heliostat field
each subsystem. Spelling et al. [4] applied evolutionary optimiza- model by dividing the entire heliostat field into 11×11 cells.
tion algorithm to optimize the involved design parameters of the Then, the local enumeration algorithm was used to solve the SPT
SPT plant subsystems, which resulted in a compromise between plant optimization problem aiming at the minimum LCOE [9].
the minimal capital cost and the minimal LCOE of a combined- Ramos et al. [10] used the NSPOC local algorithm to obtain the
cycle SPT plant. Schimitz et al. [5] presented a genetic algorithm optimal design of the SPT plant with the objective parameter
International Journal of Low-Carbon Technologies 2019, 14, 475–486
© The Author(s) 2019. Published by Oxford University Press.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which
permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
doi:10.1093/ijlct/ctz031 Advance Access publication 1 August 2019 475
Y. Lou et al.

LCOE. The results showed that the NSPOC algorithm had similar mation model. Although the RSM approximation model can suc-
optimal solution to that of the genetic global algorithm. Collado cessfully surrogate the numerical problem and greatly reduce the
and Guallar [11] decomposed the high-dimension optimization computation time of the optimization process, it lacks accuracy
problem of the SPT plant design into two successive optimization for highly nonlinear and multi-variable optimization problems.
processes. The first step is to optimize the heliostat field layout by To solve this problem, some researchers [20, 21] applied the
the Campo code [12] with the maximum annual incident energy successive response surface methodology (SRSM) approximation
as the objective function. Then, the optimal solar receiver size model in mechanics field to improve the accuracy of the RSM

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


and tower height were obtained with minimum LCOE. Similar model.
optimization decomposition strategies [13, 14] were also used to From the above, the approximation model is rarely used in the
conduct the entire SPT plant optimization design. The above opti- SPT field, which is beneficial to overcome the existing problems
mization strategies can reduce the computation time. However, in the SPT plant optimization design. In this paper, aiming at
these approaches are always at the expense of accuracy because of the minimum LCOE, the simulated annealing (SA) optimization
the local algorithm and the decomposed models. Moreover, in our approach was performed to solve the molten salt SPT plant opti-
previous work, the SPT plant optimization problem was solved mization problem on the basis of the SRSM approximation model,
by decoupling those weakly correlated parameters based on the which developed form the RSM model. The proposed strategy
global sensitivity analysis [15], but this method may not applicable is expected to simultaneously guarantee high accuracy and low
when most parameters were strongly correlated. computation time of the SPT plant optimization design.
From these studies, there is a need to develop a strategy with
high accuracy and low computation time for the SPT plant opti-
mization design. As a high-order polynomial function form to 2 CONSTRUCTION OF THE SPT PLANT
define the relation between the design parameters and the objec- OPTIMIZATION MODEL
tive parameter, the response surface methodology (RSM) approx-
imation model [16] is widely used in the optimization design of Figure 1 gives a diagram of the entire molten salt SPT plant system.
various complex systems such as aerospace, chemistry and so on. It is composed of the heliostat field, solar receiver, thermal storage
Sevant et al. [17] combined the partial different equation method system and power block. The number of heliostats in our plant
and the RSM approximation model to a flying wing optimization optimization design is fixed at 2650, which is identical to that of
design problem. It was found that this approach could accurately the Gemasolar plant [11].
explore the global optimum and drastically reduce the computa-
tion time. Liyana-Pathirana and Shahidi [18] demonstrated that
the number of experiments used to obtain the accurate optimum 2.1 Design variables
conditions for the extraction of phenolic compounds from wheat 2.1.1 Heliostat field variables
could be largely reduced based on the RSM approximation model. Figure 2 shows the typical heliostat field layout with a surrounding
Anyanwu et al. [19] found that it was effective to optimize the radial stagger configuration. The heliostat field is composed of
microalgae dewatering efficiency by means of the RSM approxi- several zones with identical azimuth spacing and radial spacing.

Figure 1. Diagram of the entire molten salt SPT plant system.

476 International Journal of Low-Carbon Technologies 2019, 14, 475–486


Efficient optimized design of solar power tower plants

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


Figure 2. Typical heliostat field layout.

The heliostat dimension is 12.84 m by 9.45 m [22]. The optical heat transfer fluid of the receiver and operates at 290–565◦ C. The
error of the heliostat is 2.9 mrad [22]. receiver is designed by two parameters as follows:
To define a typical heliostat field layout with three zones, several V09 R: Receiver radius (m);
design variables should be given as listed below: V10 THT: Tower height (m),
V01–V03 Δεjj : Azimuth spacing between adjacent heliostats of where R is 3–5 m and THT is 125–165 m. The receiver surface
the same row in the jj-th zone (rad). (jj = 1, 2, 3). area A can be approximately defined as A ≈ 2π R(2R + 1) [24].
V04–V06 rjj : Radial spacing between consecutive rows in the
jj-th zone (m). (jj = 1, 2, 3).
2.1.3 Thermal storage system and power block variables
V07–V08 Δγ jj : Transition spacing between the last row of the
The thermal storage system is composed of a hot tank and a
jj-th zone and the first row of the (jj + 1)-th zone (m). (jj = 1, 2).
cold tank. The thermal storage capacity can be determined by the
For a better understanding, Figure 2 also shows a larger version
variable as below:
of the first two rows in a heliostat field. The horizontal distance
V11 H: Equivalent thermal energy supply hours (h),
between the tower and the first heliostat row r1 is described as
where H is 3–12 h.
For the rated power output cycle operation, the inlet feed water
r1 = 0.8THT (1) temperature is 215◦ C, the parameters of the superheat steam that
enters the turbine are 12.5 MPa and 538◦ C, and the condensing
in which THT is the tower height. pressure is 0.008 MPa. In the case of a fixed heliostats number,
To avoid mechanical collisions among the heliostats, the min- the rated generating capacity of the power block is decided by the
imum azimuth spacing εmin between adjacent heliostats of the parameter as follows:
same row and minimum radial spacing rmin between consecu- V12 SM: Solar multiples (−),
tive rows are determined as follows [23]: where SM is 1.3–2.7.

εmin = 2asin [l/ (2r1 )] ≈ l/r1 (2)


2.2 Performance and economic models
rmin = l cos 30◦ = 0.866l, (3) In the present study, the system advisor model software [25] is
applied to provide the annual weather database of Sevilla (37.4◦ N,
where l is the diagonal of the heliostat. Consequently, Δεi ranges
5.9◦ W), Spain. Included in this database are measurements of
from l/r1 to 2l/r1 , Δri ranges from 0.866l to 2.4l and Δγ i ranges
direct normal irradiance (DNI), ambient temperature and wind
from 0.866l to 2.4l.
velocity at hourly intervals for a typical year statistical data. The
typical year developed from a set of data for years 1982–1995, use
2.1.2 Receiver variables data from year 1987 for May, year 1992 for September etc. Con-
The external receiver is applied to absorb the reflected sunlight sequently, 24∗365 cases were used to obtain the statistic analysis
from the heliostat field. The binary nitrate salt is chosen as the of annual performance. Then, based on the mature molten salt

International Journal of Low-Carbon Technologies 2019, 14, 475–486 477


Y. Lou et al.

SPT plant subsystem performance models [26–28], the annual 3 OPTIMIZATION STRATEGY BASED ON
electricity production is obtained by summing up the hourly
power output. Li et al. [26] gave the performance model of the
THE RSM MODEL FOR THE SPT PALNT
heliostat field and optimized the heliostat field layout with given
From the above SPT plant optimization model, it could be
parameters related to the receiver, thermal storage system and
observed that the optimization process requires a large amount
power block. Singer et al. [27] provided the calculation model of
of calculation time because of high-dimensional parameters and
the receiver thermal efficiency by means of iterative method. Zhu
complex performance models. Then, to reduce the calculation

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


et al. [28] presented the heat balance calculation method of the
time in the optimization process, the approximation function was
power block.
built to express the relationship between the SPT plant design
Finally, the economic model of the SPT plant is constructed and
variables and the LCOE instead of (5).
the LCOE can be determined as follows [29]:

 
Cinvest Idebt (1+Idebt )n
+ Iinsurance + CO&M
3.1 Design of experiment
(1+Idebt )n −1 Before the construction of the approximation model for the SPT
LCOE = , (4)
W plant design, a design of experiment (DOE) should be imple-
mented to generate a set of samples in the range of the SPT
where I debt and I insurance are the debt and insurance interest, plant design variables. In this study, the optimal latin hypercube is
respectively, n is operation life of the SPT plant and equal to chosen for the DOE because of its low computational cost and uni-
30 years, W is the yearly power output, Cinvest is the SPT plant form samples, which result in better approximations [32]. Then,
capital cost (which covers direct and indirect investments) [30], the numerical analysis is begun to achieve the actual response
CO&M is the yearly operation and maintenance cost and accounts LCOE on the basis of the thermo-economic models of the SPT
for 1% of total investment per year [31]. plant.

2.3 Mathematical optimization model 3.2 Construction of the RSM model


As a result, the objective parameter LCOE is a function of 12 After the response results of these sample points are obtained,
design variables as below: the approximation model is used to analyze the results. Among
different approximation models, the RSM is an effective statis-
tical method to achieve the relationship between the response
Y = f (x1 , x2 , . . . , xd ) , (5) and design parameters [33]. Therefore, in this study, the RSM
is applied to construct approximation model for the SPT plant
where Y is the objective parameter LCOE, xi (i = 1, 2, . . . , d) are design. The relationship between the 12 design variables and the
the design variables and d is design variables number. response parameter LCOE is fitted with a high-order polynomial
To minimize the objective parameter LCOE, the mathematical using the RSM. The approximate function constructed by the
model of the SPT plant optimization problem is developed as RSM can be represented as below:
below:
Objective : Minimize Y 
N

, (6) Ŷ = βii ϕii (x), (7)


Subject to : xiL ≤ xi ≤ xiU ii=1

where xiL and xiU are the lower and upper limits of the entire design
space for the i-th design variable, respectively. Table 1 lists the where Ŷ is the approximate LCOE, x is the vector of the SPT plant
entire design space of the SPT plant design variables. design variables and can be expressed as x = (x1 , x2 , . . . , xd ), βii
is the unknown coefficient, N is the number of terms of ϕii (x),
ϕii (x) is the basis function of x and the form of a fourth-order
Table 1. Entire design space of the SPT plant design variables. basis function is

Lower Upper Lower Upper


bound bound bound bound 
d 
d 
d
Ŷ = β1 + βi+1 xi + βi+d+1 xi2 + βi+2d+1 xi3
ε 1 (rad) l/r1 2l/r1 γ 1 (m) 0.866l 2.4l i=1 i=1 i=1
ε 2 (rad) l/r1 2l/r1 γ 2 (m) 0.866l 2.4l
ε 3 (rad) l/r1 2l/r1 R (m) 3 5 
d 
d 
d
r1 (m) 0.866l 2.4l THT (m) 125 165 + βi+3d+1 xi4 + β −i2 +23i +i’+4d−11 xi xi’ .
2
r2 (m) 0.866l 2.4l H (h) 3 12 i=1 i=1 i’>i
r3 (m) 0.866l 2.4l SM (−) 1.3 2.7
(8)

478 International Journal of Low-Carbon Technologies 2019, 14, 475–486


Efficient optimized design of solar power tower plants

The unknown coefficients vector β = (β1 , β2 , . . . , βN )T can be 4 OPTIMIZATION STRATEGY BASED ON


obtained by the least square method and the detailed steps are as THE SRSM MODEL FOR THE SPT PLANT
follows:
(1) m sample points are generated by the DOE. Generally, the traditional RSM model lacks accuracy for multiple
(2) The sample points x(j) ( j = 1, 2, . . . , m) are used to conduct variable and highly nonlinear system. Developed from the RSM
the numerical analysis to obtain the vector of the actual model, the SRSM model is recently built to obtain a higher preci-
T sion approximation function with reduced design space.

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


response values Y = (Y(x(1) ), Y(x(2) ), . . . , Y(x(m) )) .
(3) The sum of the squares of the errors between the actual The basic principle of the SRSM model is the application of a
m subregion of the entire design space to determine an approximate
responses and the approximate responses is E(β) = optimum by the iterative optimization method [35]. For each
j=1
iteration process, an adaptive domain reduction method is used

N
to diminish the design space size and the RSM model is used to
[Y(x(j) ) − βii ϕii (x(j) )]2 . To minimize the value of E(β), β
ii=1 construct a new approximation function.
is obtained by solving the differential equation E’ (β) = 0.
(4) Based on the equation of E’ (β) = 0, vector β of the unknown
−1 4.1 Adaptive domain reduction method
coefficients is finally obtained by β = (X T X) (X T Y), where Considering that an efficient reduction strategy of the design
X is calculated by space is the key to construct the SRSM model, an adaptive domain
⎡ reduction method is proposed to update the subregion of the



ϕ1 x(1) . . . ϕN x(1) design variables with iterations. The specific methodology is
⎢ ⎥ explained as follows.
X=⎣ ... ... ... ⎦.
(m)
(m)
The subregion of the initial iteration is defined as the entire
ϕ1 x . . . ϕN x
design space of each variable. For the k-th (k = 2, 3, . . . ) iter-
ation, the lower and upper bounds of its subregion for the i-th
L and xU , respectively. Then,
design variable are assumed to be xi,k
To measure the reliability of the approximation model of the i,k
SPT plant design, the coefficient R-square is applied to conduct opt
the k-th subregion is approximated, and the k-th optimum xi,k is
an error analysis and calculated as follows: obtained.
For the (k + 1)-th (k = 2, 3, . . . ) iteration, its subregion size
M 


2 is reduced from the previous subregion. Generally, a smaller dis-
opt
Y x(J) − Ŷ x(J) tance between the k-th optimum xi,k and the (k-1)-th optimum
J=1 opt opt
R2 = 1 − , (9) xi,k−1 indicates that xi,k is closer to the actual optimal point.
M


2
Y x(J) − Y Thus, the size of the (k + 1)-th subregion can be more rapidly
J=1 reduced. Based on the above analysis, a ratio parameter δi,k+1 is
opt opt
proposed to estimate the deviation between xi,k and xi,k−1 and
represented as
where M is the sample size for the error analysis, Y(x( J) ) and
 opt 
Ŷ(x( J) ) are the actual and approximate response value of the  x − xopt 
 i,k i,k−1 
J-th sample point, respectively, for the error analysis and Y is the δi,k+1 =  U , (11)
 xi − xiL 
average value of all actual responses.
Then, the value of R-square is between 0 and 1. As R-square where xiL and xiU are the lower and upper bounds of the initial
gets closer to 1, the approximation model for the SPT plant design subregion for the i-th design variable.
becomes more accurate. To accelerate the domain reduction process, different contrac-
tion approaches are used on the basis of δi,k+1 . Thus, the k-th
subregion is divided into three zones to determine which of these
3.3 Optimization method contraction approaches to be adopted. The expressions of these
Based on the construction of the traditional RSM, the mathemat-
zones are defined as follows:
ical model of the SPT plant optimization problem is described as
below: ⎧

⎨zone 1 : δi,k+1 < a1
Objective : Minimize Ŷ zone 2 : a1 ≤ δi,k+1 < a2 , (12)
. (10) ⎪

Subject to : xL ≤ x ≤ xU zone 3 : δi,k+1 ≥ a2

Then, the SPT plant global optimization problem is solved by where a1 and a2 are constants and indicate the deviation between
opt opt
the SA algorithm [34], which has the advantage of high accuracy. xi,k and xi,k−1 .

International Journal of Low-Carbon Technologies 2019, 14, 475–486 479


Y. Lou et al.

The possible adaptive domain reduction method for a typical


two-variable SRSM problem is shown in Figure 3, and the corre-
sponding illustration is as follows:
opt
(1) If the k-th optimum xi,k is very close to the (k-1)-th optimum
opt
xi,k−1 and falls into zone 1, the (k + 1)-th subregion is greatly
shrinked to b1 times of the entire design space and centered

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


opt
on the location of xi,k . (Figure 3a).
opt
(2) If the k-th optimum xi,k belongs to zone 2, the (k + 1)-th sub-
region is first reduced to b2 times of the entire design space
opt
with xi,k as the center point, and then further diminished by
intersecting with the k-th subregion (Figure 3b).
opt
(3) If the distance between the k-th optimum xi,k and the
opt opt
(k-1)-th optimum xi,k−1 is very large and the point xi,k falls
into zone 3, both size and center of the (k + 1)-th subregion
remain as the k-th subregion (Figure 3c).
Consequently, the bounds of the (k + 1)-th subregion for the
i-th design variable are expressed as follows:

 L 
opt
L
xi,k+1 = Max xi,k − 12 λi,k+1 xiU − xiL , xi,k

 U  , (13)
opt
L
xi,k+1 = Min xi,k + 12 λi,k+1 xiU − xiL , xi,k

L
where xi,k+1 U
and xi,k+1 are the lower and upper bounds of the
(k + 1)-th subregion for the i-th design variable; λi,k+1 is the
contraction parameter of the (k + 1)-th subregion for the i-th
design variable and can be determined by

⎨b1; (Case 1)
λi,k+1 = b2; (Case 2) , (14)

+ ∞; (Case 3)
where the values of b1 and b2 (b1 < b2) are set up for the specific
problem.

4.2 Optimization procedure


Based on the adaptive domain reduction method, the iterative
steps of the SPT plant optimization design are provided, and
Figure 4 shows the corresponding flow chart.
(1) Create the initial approximation function. To reduce the
computation time, the RSM is applied to evaluate the relation
between the objective parameter LCOE and the design
variables. The initial approximation function Ŷ1 (x1 ) is
established with sample points generated by the DOE in the
entire design space. Figure 3. Possible adaptive domain reduction for a typical two-variable SRSM
(2) Solve the optimization problem based on the initial ap- problem.
proximation function. Based on the initial approximation
function Ŷ1 (x1 ), the minimum LCOE and corresponding
opt
optimal variables x1 of the SPT plant design are obtained Generally, there is a relatively large deviation between the
by the SA global algorithm. The mathematical model of the opt opt
approximate result Ŷ1 (x1 ) and the actual value Y1 (x1 )
initial iterative optimization is expressed as follows:
for multiple variable and strongly nonlinear system. There-
Objective : Minimize Ŷ1 (x1 ) fore, the accuracy of the approximation function should be
. (15)
Subject to : xL ≤ x1 ≤ xU improved to obtain the actual optimal solution.

480 International Journal of Low-Carbon Technologies 2019, 14, 475–486


Efficient optimized design of solar power tower plants

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019

Figure 4. Flow chart of the SPT plant optimization procedure based on the SRSM.

(3) Update the approximation function and repeat the opti- is used to update the approximation function. For the
mization. The adaptive domain reduction method is applied second iteration, a new approximation function Ŷ2 (x2 ) is
to reduce the design space of variables and the RSM model constructed based on the updated subregion calculated by

International Journal of Low-Carbon Technologies 2019, 14, 475–486 481


Y. Lou et al.

(13). Then the optimization is repeated to obtain the new sample points were generated by the DOE to construct the quartic
opt
optimal point x2 . Similarly, the approximation function is RSM model of the SPT plant design. The actual LCOE acquired
reconstructed and the optimization process is repeated with by numerical analysis are compared with the approximate values
gradually reduced domain. The mathematical model of the obtained by RSM as shown in Figure 6. The actual LCOE was
k-th iterative optimization is represented as follows: approximately consistent with the approximate LCOE across the
entire design space and R-square of the RSM model was 0.9373.
Objective : Minimize Ŷk (xk ) After the traditional RSM model was constructed, the SA global

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


. (16) algorithm was used to solve the SPT plant optimization prob-
Subject to : xLk ≤ xk ≤ xU
k
lem. Figure 7 shows the convergence curve of the SPT plant
optimization design on the basis of the RSM. The approximate
(4) Obtain the final optimal solution. Conduct step (3) until
the solution is converged. The convergence criterion is the
relative difference of approximate objective values in the last
two iterations and can be expressed as follows:

  opt   
 Ŷ x opt 
 k k − Ŷk−1 xk−1 
    < ξ, (17)
 opt 
 Ŷk−1 xk−1 

where ξ is the convergence control parameter, and its value is


opt opt
considered as 0.01; Ŷk−1 (xk−1 ) and Ŷk (xk ) are the approximate
response values for the optimal points of the (k-1)-th and k-th
iterations, respectively.

5 RESULTS AND DISCUSSION


Figure 5. Variation of R-square with the sample size for the RSM model
Evolved from the traditional RSM model, the SRSM model of the
construction of the SPT plant design.
SPT plant system was constructed to optimize the design variables
with the objective function of minimum LCOE.

5.1 Optimization results of the SPT plant design based


on the RSM model
To reduce the computation time in the SPT plant optimization
process, the quartic RSM model was first built to express the rela-
tionship between the SPT plant design variables and the response
LCOE in the entire design space.
The accuracy and computation time of the traditional RSM
model generally vary with the number of sample points generated
for numerical analysis. To select a proper sample size for the RSM
model construction, the coefficient R-square was used to conduct
an error analysis according to (9). The variation of R-square with
the sample size for the RSM model construction of the SPT plant
design is shown in Figure 5. The coefficient R-square approached
the equilibrium status after ∼600 sample points were generated
for numerical analysis to construct the RSM model. When the
sample size increased from 600 to 2400, R-square increased by
only 1.47%, so the improvement in accuracy of the RSM model
was small. Moreover, the computation time was enhanced by
three times because the computation time of the RSM model con- Figure 6. Comparison between the actual LCOE by numerical analysis and the
struction was proportional to the sample size. Consequently, 600 approximate LCOE by RSM.

482 International Journal of Low-Carbon Technologies 2019, 14, 475–486


Efficient optimized design of solar power tower plants

Table 3. Optimization history of the SPT plant design based on the SRSM.
Design variables First Second Third Fourth
iteration iteration iteration iteration

ε 1 (rad) 1.000l/r1 1.000l/r1 1.019l/r1 1.005l/r1


ε 2 (rad) 1.108l/r1 1.000l/r1 1.009l/r1 1.000l/r1
ε 3 (rad) 1.348l/r1 1.356l/r1 1.338l/r1 1.324l/r1
r1 (m) 0.866l 0.888l 0.866l 0.866l

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


r2 (m) 0.914l 0.969l 0.990l 1.007l
r3 (m) 1.636l 1.682l 1.636l 1.633l
γ 1 (m) 1.724l 1.141l 1.214l 1.122l
γ 2 (m) 1.545l 2.306l 1.477l 1.354l
R (m) 3.5 3.4 3.4 3.4
THT (m) 148 138 136 135
H (h) 12.0 12.0 12.0 12.0
SM (−) 2.7 2.7 2.7 2.7
Approximate LCOE 21.23 21.67 22.06 22.03
(c/kWhe )
Actual LCOE (c/kWhe ) 22.23 22.06 22.05 22.02
R 2 (−) 0.9373 0.9598 0.9862 0.9917
Figure 7. Convergence curve of the SPT plant optimization process based on the
RSM.

Table 2. Optimal solution of the SPT plant design based on the RSM.
new RSM model. Then, the SPT plant optimization design was
Design variables Optimal values Design variables Optimal values performed by the SA algorithm based on the new approximation
ε 1 (rad) 1.000l/r1 γ 1 (m) 1.724l
function.
ε 2 (rad) 1.108l/r1 γ 2 (m) 1.545 Table 3 lists the optimization history of the SPT plant design
ε 3 (rad) 1.348l/r1 R (m) 3.5 on the basis of the SRSM. The optimal solution converged after
r1 (m) 0.866l THT (m) 148 four iterations. Thus, the total number of required sample points
r2 (m) 0.914l H (h) 12.0 was 2400 for numerical analysis to construct the SRSM model.
r3 (m) 1.636l SM (−) 2.7
Approximate LCOE 21.23 Actual LCOE 22.23
In the iterative optimization process, the optimum value of some
(c/kWhe ) (c/kWhe ) design variables varied greatly, while the optimal values of other
design variables almost unchanged. For instance, the optimal
γ 1 varied from 1.724l to 1.122l with iterations, but the opti-
mal H remained unchanged. The reason may be that the vari-
LCOE decreased from 25.73 c/kWhe to 21.23 c/kWhe after 10138 ation of the actual response LCOE was highly nonlinear with
optimization steps. Although the number of optimization steps γ 1 in the entire design space, which resulted in the low accu-
was large, the optimization procedure was completed in a very racy of the RSM model of the initial iteration. Thus, the SRSM
short time because of the application of the RSM model. Table 2 model was applied to decrease the nonlinear effect with reduced
lists the optimization results based on the RSM. The approximate design space.
and actual LCOE at the optimal point were 21.23 c/kWhe and The error analysis results of the approximation function are
22.23 c/kWhe , respectively. Because the potential LCOE reduction also shown in Table 3. With the number of iterations increased,
of the SPT plant was a few cents per kilowatt hour [36], the R-square of the approximation function enhanced from 0.9373
deviation between the approximate and the actual LCOE was to 0.9917 and the deviation between the approximate and actual
large, which indicated that the developed RSM model could not LCOE at the optimal point decreased from 1 c/kWhe to 0.01
be accurately used to calculate the response LCOE. Therefore, c/kWhe . Hence, the accuracy of the approximation function was
despite the low computation time, the constructed RSM model improved during the iterative optimization process. Then, an
should be updated by the SRSM to improve the accuracy of the actual minimum LCOE of 22.02 c/kWhe was obtained based on
approximation function. the SRSM model, which was a more preferable optimal solution
than that based on the traditional RSM model constructed in the
entire design space.
5.2 Optimization results of the SPT plant design based The updating history of the SPT plant design variables subre-
on the SRSM model gion based on the SRSM is shown in Table 4. For this case, the
The SRSM model was constructed by the iterative optimization values of a1 and a2 in (10) were 0.1 and 0.5, respectively, as well
method to further improve the precision of the SPT plant opti- as b1 and b2 in (12) were set to 0.4 and 1, respectively. The history
mization design results. For each iteration, 600 sample points indicated that the subregion of the design variables was gradually
generated from the updated subregion were set to reconstruct a reduced with the iterations. For example, the subregion of ε2

International Journal of Low-Carbon Technologies 2019, 14, 475–486 483


Y. Lou et al.

Table 4. Updating history of the SPT plant design variables subregion based on the SRSM.
Design variables First iteration Second iteration Third iteration Fourth iteration

ε 1 (rad) (1–2)l/r1 (1–1.5)l/r1 (1–1.2)l/r1 (1–1.2)l/r1


ε 2 (rad) (1–2)l/r1 (1–1.608)l/r1 (1–1.5)l/r1 (1–1.209)l/r1
ε 3 (rad) (1–2)l/r1 (1–1.848)l/r1 (1.156–1.556)l/r1 (1.156–1.538)l/r1
r1 (m) (0.866–2.4)l (0.866–1.633)l (0.866–1.195)l (0.866–1.173)l
r2 (m) (0.866–2.4)l (0.866–1.681)l (0.866–1.276)l (0.866–1.276)l

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


r3 (m) (0.866–2.4)l (0.869–2.4)l (1.376–1.989)l (1.376–1.943)l
γ 1 (m) (0.866–2.4)l (0.957–2.4)l (0.957–1.908)l (0.957–1.521)l
γ 2 (m) (0.866–2.4)l (0.866–2.312)l (0.866–2.312)l (0.866–2.312)l
R (m) 3–5 3–4.5 3–3.8 3–3.7
THT (m) 125–165 128–165 128–158 128–144
H (h) 3–12 7.5–12 10.2–12 10.2–12
SM (−) 1.3–2.7 2–2.7 2.4–2.7 2.4–2.7

Figure 8. Convergence curves of the SPT plant optimization design based on the SRSM. (a) Iteration 1. (b) Iteration 2. (c) Iteration 3. (d) Iteration 4.

reduced from (1–2)l/r1 to (1–1.209)l/r1 after four iterations. Fig- 5.3 Comparisons of the proposed strategy and actual
ure 8 shows the convergence curves of the SPT plant optimization model-based algorithms
design based on the SRSM. Although the number of optimization To demonstrate the precision and computation time of the pro-
steps was ∼10000 for each iteration, the optimization process was posed strategy, the SA global algorithm and modified method of
not time-consuming due to the application of the approximation feasible directions (MMFD) local algorithm based on the actual
function. SPT plant analysis model were also adopted to make comparisons

484 International Journal of Low-Carbon Technologies 2019, 14, 475–486


Efficient optimized design of solar power tower plants

Table 5. Optimal solutions of different strategies for the SPT plant design.
Design variables Present strategy Actual model-based SA algorithm Actual model-based MMFD algorithm

ε 1 (rad) 1.005l/r1 1.000l/r1 1.000l/r1


ε 2 (rad) 1.000l/r1 1.003l/r1 1.017l/r1
ε 3 (rad) 1.324l/r1 1.315l/r1 1.303l/r1
r1 (m) 0.866l 0.868l 0.866l
r2 (m) 1.007l 1.011l 1.015l

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


r3 (m) 1.633l 1.637l 1.710l
γ 1 (m) 1.122l 0.994l 1.451l
γ 2 (m) 1.354l 1.224l 1.576l
R (m) 3.4 3.4 3.9
THT (m) 135 133 145
H (h) 12.0 12.0 7.8
SM (−) 2.7 2.7 2.2
LCOE (c/kWhe ) 22.02 22.01 23.01
Number of numerical analyses 2400 10137 368

of the SPT plant optimization design results. In general, the SA 6 CONCLUSIONS


algorithm obtains high precision optimization results in spite
of enormous computation time [36]. The MMFD algorithm is A novel optimization strategy with high accuracy and low com-
a classic gradient-based local optimization strategy and always putation time was proposed to acquire the minimum LCOE of the
obtains the local optimal solution quickly [37]. molten salt SPT plant. This method makes a comprehensive use
Table 5 shows the optimization solutions of different strategies of the SRSM model and the SA algorithm, which were applied for
for the SPT plant design. The minimum LCOE of the proposed the approximation function construction and design parameters
strategy, actual model-based SA algorithm and actual model- optimization, respectively.
based MMFD algorithm were 22.02 c/kWhe , 22.01 c/kWhe and
23.01 c/kWhe , respectively. The optimal values of some design (1) The approximation model constructed by the traditional
variables of the actual model-based MMFD algorithm signifi- RSM model for the SPT plant design lacked accuracy, which
cantly deviated from those of the actual model-based SA algo- led to the deviation between the approximate and the actual
rithm, such as γ 1 and H. As a result, the optimal solution of the LCOE at the optimal point was up to 1 c/kWhe .
actual model-based MMFD algorithm was worse than that of the (2) Developed from the traditional RSM model, the SRSM
actual model-based SA algorithm because the MMFD algorithm model was constructed with the proposed adaptive domain
always obtained a local optimal solution. However, the proposed reduction method for the SPT plant. By utilizing the
strategy had notably similar minimum LCOE and optimum val- SRSM model, the accuracy of the approximation model
ues of the design variables to the actual model-based SA algo- was improved and R-square of the approximation function
rithm, which confirmed the accuracy of the proposed strategy. increased from 0.9373 to 0.9917.
The computation time of different strategies for the SPT plant (3) Based on the SRSM approximation model, the SA global algo-
optimization design was generally proportional to the number of rithm was used to conduct an iterative optimized design for
SPT plant numerical analyses. For the traditional optimization the SPT plant. The actual minimum LCOE is 22.02 c/kWhe
strategies, such as the actual model-based SA algorithm and the after four iterations, which was more favorable than that
actual model-based MMFD algorithm, the number of numerical based on the traditional RSM model.
analyses was equal to the number of optimization steps. For (4) Compared with the actual model-based SA global algorithm,
the proposed strategy, the computation time of the optimization the proposed strategy achieved a notably similar optimal
process could be negligible after the approximation model was design and reduced the computation time by ∼75%. But it
constructed, so the number of numerical analyses was the same as should be noted that the computation time may be increased
the number of sample points used to construct the approximation with increasing number of design parameters to ensure the
model. Then, the number of numerical analyses for different accuracy of the optimization results.
strategies are listed in Table 5. It could be seen that the number
of numerical analyses for the proposed strategy, actual model-
based SA algorithm and actual model-based MMFD algorithm
were 2400, 10137 and 368, respectively. The computation time of
ACKNOWLEDGMENTS
the proposed strategy was about 75% lower than that of the actual The financial supports for this research project from the National
model-based SA algorithm. Therefore, the proposed strategy is Natural Science Foundation of China (No. 51806009) and the
a preferable optimization method with high accuracy and low Fundamental Research Funds for the Central Universities (No.
computation time. ZY1814) are gratefully acknowledged.

International Journal of Low-Carbon Technologies 2019, 14, 475–486 485


Y. Lou et al.

REFERENCES [20] Kurtaran H, Eskandarian A, Marzougui D, et al. Crashworthiness design


optimization using successive response surface approximations. Comput
[1] Boudaoud S, Khellaf A, Mohammedi K, et al. Thermal performance Mech 2002;29:409–21.
prediction and sensitivity analysis for future deployment of molten salt [21] Pajunen S, Heinonen O. Automatic design of marine structures by
cavity receiver solar power plants in Algeria. Energy Convers Manag using successive response surface method. Struct Multidiscipl Optim
2015;89:655–64. 2014;49:863–71.
[2] Behar O, Khellaf A, Mohammedi K. A review of studies on central receiver [22] Noone CJ, Torrilhon M, Mitsos A. Heliostat field optimization: a
solar thermal power plants. Renew Sust Energ Rev 2013;23:12–39. new computationally efficient model and biomimetic layout. Sol Energy

Downloaded from https://academic.oup.com/ijlct/article-abstract/14/4/475/5542446 by Universidade Federal da Bahia user on 17 December 2019


[3] Avila-Marin AL, Fernandez-Reche J, Tellez FM. Evaluation of the potential 2012;86:792–803.
of central receiver solar power plants: configuration, optimization and [23] Collado FJ. Preliminary design of surrounding heliostat fields. Renew
trends. Appl Energy 2013;112:274–88. Energ 2009;34:1359–63.
[4] Spelling J, Favrat D, Martin A, et al. Thermoeconomic optimization of a [24] Ortega JI, Burgaleta JI, Téllez FM. Central receiver system solar power
combined-cycle solar tower power plant. Energy 2012;41:113–20. plant using molten salt as heat transfer fluid. J Energ Eng 2008;130:
[5] Schmitz M, Schwarzbözl P, Buck R, et al. Assessment of the potential 024501.
improvement due to multiple apertures in central receiver systems with [25] Solar Advisor Model (SAM). National Renewable Energy Laboratory.
secondary concentrators. Sol Energy 2006;80:111–20. Version 2017.9.5. https://sam.nrel.gov/download/version-2017-9-5.html
[6] Soltani R, Keleshtery PM, Vahdati M, et al. Multi-objective optimization of [26] Li C, Zhai R, Yang Y. Optimization of a heliostat field layout on annual
a solar-hybrid cogeneration cycle: application to CGAM problem. Energy basis using a hybrid algorithm combining particle swarm optimization
Convers Manag 2014;81:60–71. algorithm and genetic algorithm. Energies 2017;10:1924.
[7] Starke AR, Cardemil JM, Escobar R, et al. Multi-objective optimiza- [27] Singer C, Buck R, Pitz-Paal R, et al. Assessment of solar power tower
tion of hybrid CSP+ PV system using genetic algorithm. Energy driven ultrasupercritical steam cycles applying tubular central receivers
2018;147:490–503. with varied heat transfer media. J Sol Energy 2010;132: 041010.
[8] Kistler BL. 1986. A User’s Manual for DELSOL3: A Computer Code for
[28] Zhu MS, Liu Y, Lin ZZ, et al. 1994. Engineering Thermodynamics. Beijing:
Calculating the Optical Performance and Optimal System Design for Solar
Tsinghua University Press. In Chinese.
Thermal Central Receiver Plants. Albuquerque, USA: Sandia National
[29] Montes MJ, Abánades A, Martínez-Val JM. Performance of a direct steam
Laboratories.
generation solar thermal power plant for electricity production as a func-
[9] Lipps FW, Vant-Hull LL. A cellwise method for the optimization of large
tion of the solar multiple. Sol Energy 2009;83:679–89.
central receiver systems. Sol Energy 1978;20:505–16.
[30] Turchi CS, Heath GA. 2013. Molten salt power tower cost model for the
[10] Ramos A, Ramos F. Strategies in tower solar power plant optimization. Sol
system advisor model (SAM). In National Renewable Energy Laboratory.
Energy 2012;86:2536–48.
USA: Golden.
[11] Collado FJ, Guallar J. Two-stages optimised design of the collector field of
[31] Montes MJ, Abánades A, Martínez-Val JM, et al. Solar multiple optimiza-
solar power tower plants. Sol Energy 2016;135:884–96.
tion for a solar-only thermal power plant, using oil as heat transfer fluid in
[12] Collado FJ, Guallar J. Campo: generation of regular heliostat fields. Renew
the parabolic trough collectors. Sol Energy 2009;83:2165–76.
Energ 2012;46:49–59.
[32] Devanathan S, Koch PN. Comparison of meta-modeling approaches for
[13] Carrizosa E, Domínguez-Bravo C, Fernández-Cara E, et al. Optimization
optimization. ASME 2011 International Mechanical Engineering Congress
of multiple receivers solar power tower systems. Energy 2015;90:2085–93.
and Exposition. ASME 2011, pp. 827–35.
[14] Saghafifar M, Gadalla M. Thermo-economic analysis of conventional
combined cycle hybridization: United Arab Emirates case study. Energy [33] Ekren O, Ekren BY. Size optimization of a PV/wind hybrid energy conver-
Convers Manag 2016;111:358–74. sion system with battery storage using response surface methodology. Appl
[15] Luo Y, Lu T, Du XZ. Novel optimization design strategy for solar power Energy 2008;85:1086–101.
tower plants. Energy Convers Manage 2018;177:682–92. [34] Kirkpatrick S, Gelatt CD, Vecchi MP. Optimization by simulated anneal-
[16] Laurent PJ. 1972. Approximation et Optimization. Paris: Hermann. ing. Science 1983;220:671–80.
[17] Sevant NE, MIG B, Wilson MJ. Aerodynamic design of a flying wing using [35] Stander N, Craig KJ. On the robustness of a simple domain reduc-
response surface methodology. J Aircr 2000;37:562–9. tion scheme for simulation-based optimization. Eng Computation 2002;
[18] Liyana-Pathirana C, Shahidi F. Optimization of extraction of phenolic 19:431–50.
compounds from wheat using response surface methodology. Food Chem [36] Kolb GJ, Ho CK, Mancini TR, et al. 2011. Power Tower Technology
2005;93:47–56. Roadmap and Cost Reduction Plan. Albuquerque, USA: Sandia National
[19] Anyanwu R, Rodriguez C, Durrant A, et al. Optimisation of tray drier Laboratories.
microalgae dewatering techniques using response surface methodology. [37] Vanderplaats GN, Moses F. Structural optimization by methods of feasible
Energies 2018;11:2327. directions. Comput Struct 1973;3:739–55.

486 International Journal of Low-Carbon Technologies 2019, 14, 475–486

You might also like