You are on page 1of 9

Applied Thermal Engineering 107 (2016) 1210–1218

Contents lists available at ScienceDirect

Applied Thermal Engineering


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

Research Paper

ANN-based optimization of a parabolic trough solar thermal power plant


T.E. Boukelia a,⇑, O. Arslan b,⇑, M.S. Mecibah a
a
Laboratory of Mechanics, Mechanical Engineering Department, Faculty of Technology Sciences, University of Brothers Mentouri, Constantine 25000, Algeria
b
Mechanical Engineering Department, Engineering Faculty, Dumlupinar University, 43270 Kutahya, Turkey

h i g h l i g h t s

 ANN back-propagation algorithm with three different variants was studied.


 Best approach for modeling of PTSTPP was investigated.
 PTSTPP was optimized using the best ANN model.

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

Article history: Design and optimization of a solar power plant are very complex and require many calculations, data and
Received 12 December 2015 time. From this point of view, artificial neural network (ANN) models are desired options to determine
Revised 12 July 2016 techno-economic performances of this kind of plants. Therefore, the objective of this study is to investi-
Accepted 13 July 2016
gate the feed-forward back-propagation learning algorithm with three different variants; Levenberge
Available online 15 July 2016
Marguardt (LM), Scaled Conjugate Gradient (SCG), and Pola-Ribiere Conjugate Gradient (CGP), used in
the ANN to find the best approach for prediction and techno-economic optimization of parabolic trough
Keywords:
solar thermal power plant (PTSTPP) integrated with fuel backup system and thermal energy storage. The
Artificial neural network
Levelized cost of electricity
obtained statistical parameters showed that LM algorithm with 38 neurons in a single hidden layer looks
Optimization as the best ANN model to predict the annual power generation ðPGnet Þ and levelized cost of electricity
Parabolic trough power plant (LCOE) of the presented PTSTPP. Moreover, the obtained weights from this topology were used in the
LCOE analysis for determining the optimum system design. It is therefore available to get a minimum
LCOE of 8.88 Cent/kWh from the new optimized plant when the plant characteristics are; 34 °C and
850 W/m2 for both design ambient temperature and solar radiation, 23 m for row spacing, 1.7 for solar
multiple, and 2.5 h number of hours for the storage system.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction The selection of the best configuration of a solar plant, and the
optimization of working parameters of such systems are the most
Due to the fastest-increasing demand for energy, the reserves of important aspects in designing a solar thermal power plant [2].
fossil fuels will not satisfy most of the world’s requirements for Several studies have been reported to find an optimum design
this essential material in the future. Furthermore, the using of for these plants with maximum techno-economic benefits. Montes
energy coming from fossil origin, has adverse environmental et al. [3] optimized the solar multiple of stand alone PTSTPP based
effects such as global warming, climate change, greenhouse gas on annual energy generation and levelized cost of electricity
emissions, and depletion of the ozone layer. For these two main (LCOE). While Bonilla et al. [4] used multi-objective algorithm
reasons, it is worth to explore new alternative-friendly renewable approach for calibrating a dynamic model of a direct steam gener-
energy sources. While solar energy can play a key role to solve ation PTSTPP. Ramos and Ramos [5] studied a new method for opti-
these issues, parabolic trough solar thermal power plant (PTSTPP) mizing a central receiver solar plant by study the design of this
is considered as one of the most proven concentrating solar power plant as a function of eleven variables, and described the method
technologies for power generation [1]. to evaluate the impact of each design variable on the performance
of the plant. Avila-Marin et al. [6] presented a parametric analysis
of a medium to large scale central receiver thermal power plants.
⇑ Corresponding authors. The analysis is based on; size and location of these plants, technol-
E-mail addresses: taqy25000@hotmail.com (T.E. Boukelia), oguz.arslan@dpu.edu. ogy integrated between steam and molten salt, storage system in
tr (O. Arslan).

http://dx.doi.org/10.1016/j.applthermaleng.2016.07.084
1359-4311/Ó 2016 Elsevier Ltd. All rights reserved.
T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218 1211

Nomenclature

C env environmental cost according to CO2 rejected (US$) Dtes total number of desired storage hours (h)
C inv total investment cost (US$) gcycles des design point power cycle efficiency
C O&M annual operating and maintenance costs (US$)
crf capital recovery factor Abbreviation
Etes number of hours that storage system can supply energy ANN Artificial neural network
for the operation of the power block at the design input CGP Pola-Ribiere Conjugate Gradient
level (h) COV coefficient of variation
f backup fossil fill fraction ð0 < f backup < 1Þ FBS fuel back-up system
hSFo enthalpy at the outlet solar field (kJ/kg) HTF heat transfer fluid
hTi enthalpy at the inlet of high pressure turbine (kJ/kg) HVAC heating, ventilating, air conditioning
Ib direct normal radiation (W/m2) LCOE levelized cost of electricity
kd annual discount rate LM Levenberge Marguardt
l row spacing between parallel collectors (m) MPE mean percent error
m _ PB mass flow rate of the HTF in the PB (kg/s) NREL National Renewable Energy Laboratory
N depreciation operation time of the system (years) PB power block
PGnet net power generation (kWh) PTSTPP parabolic trough solar thermal power plant
Q_backup thermal energy must be supplied in the FBS (kWh) R2 coefficient of determination
Q FBS thermal energy delivered by the FBS (MWh) RMSE root mean square error
Q Net net electric power output (MWh) SAM Solar Advisor Model
Q PB thermal energy received by the power block (MWh) SCG Scaled Conjugate Gradient
Q SF solar field thermal output (MWh) SF solar field
Q TES thermal energy to/from the storage system (MWh) SM solar multiple
Q Tot total solar energy collected by the solar field (MWh) STEC Solar Thermal Electric Components
Q_ total total energy needed to reach the thermodynamic state TES thermal energy storage system
(kWh) TMY typical meteorological year
T amb ambient temperature at design (°C) TRNSYS transient systems simulation program
wdes design cycle thermal requirement (kW)

addition to component’s cost. Furthermore, Spelling et al. [7] chosen as inputs. Arslan [14] developed an ANN model for the opti-
developed a dynamic model of a combined cycle solar thermal mization of Kalina Cycle System-34, and approved the availability
plant working under a variety operating conditions and super- of ANN tool in which Levenberg Marguardt (LM) was used as the
structure layouts. This model is based on a multi-objective training algorithm for the optimization process. In another study,
thermo-economic optimization of both plant performance and Arslan and Yetik [15] used the back-propagation learning algo-
cost. rithm with three different variants: Levenberge Marguardt (LM),
An extensive energy and economic analysis of stand alone para- Pola-Ribiere Conjugate Gradient (CGP), and Scaled Conjugate Gra-
bolic trough solar plant are reported by Desai and Bandyopadhyay. dient (SCG) in the ANN to find the best approach to optimize a
The effects of inlet turbine pressure and temperature, design radi- supercritical ORC-Binary geothermal power plant located in Simav
ation, as well as plant size on overall efficiency and LCOE of this (Turkey). Optimum working conditions of a combustion chamber
plant are studied in order to find the optimal design [2]. Cabello with different fuels were predicted using ANN modeling by Arslan
et al. [8] performed a simplified model based on genetic algorithm [16], then the combustion process was analyzed using exergo-
to optimize a solar power plant to produce the maximum yearly economic study. He found that LM with eight neurons in a single
profit, and analyzed the solar multiple and capacity factor of the hidden layer is the best algorithm to be adapted for this purpose.
new optimized plant. Martín and Martín [9] presented the year- Moreover, Arslan and Yetik [17] applied the back-propagation
round optimization of the operation of a concentrated power plant algorithm with three different variants on ORC-binary geothermal
based on molten salt technology for Almerían (Spain) conditions, power plant to find the most suitable design of such systems.
the optimization procedure is formulated as a multi-period non- The design and optimization of a parabolic solar thermal power
linear programming problem. plant are very complex and require many calculations, data analy-
The artificial neural network (ANN) model is well known as a sis and time consuming. From this point of view, ANN models are
good option for modeling of energy systems, it has the potential desired options in this subject. Therefore, in the present study, it is
to be quicker and more practical alternative compared to other tra- aimed to investigate the feed-forward back-propagation learning
ditional methods [10]. Consequently, it has been applied by algorithm with three different variants; Levenberge Marguardt
numerous researchers. Kalogirou et al. [11–13] investigated some (LM), Scaled Conjugate Gradient (SCG), and Pola-Ribiere Conjugate
applications of ANN method in renewable energy systems field in Gradient (CGP), used in the artificial neural network to find the
different areas such as solar heating and steam generation, HVAC best approach for prediction and techno-economic optimization
systems, solar radiation and wind speed predictions, as well as of PTSTPP integrated with fuel backup system and thermal energy
modeling and control in power generation systems. Bekat et al. storage. The model outputs were selected to be the annual power
[10] predicted the amount of bottom ash formed in a pulverized generation ðPGnet Þ and the levelized cost of electricity (LCOE), while
coal-fired plant using ANN modeling, the ratio of the amount (Bot- five inputs were chosen such as design ambient temperature and
tom ash/Coal burned) was chosen as model output, while moisture solar radiation at design, solar multiple, row spacing between par-
contents, ash contents and lower heating values of the coals were allel collectors, and full load hours of the thermal energy system.
1212 T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218

CON: Condenser, CP: Condenser Pump, DEA: Deaerator, DP: Deaerator Pump, ET: Expansion Tank, FWH: Feed Water Heater, GEN: Generator,
HPT: High Pressure Turbine, LPT: Low Pressure Turbine, PH: Pre-Heater, RH: Re-Heater, SG: Steam Generator, SH: Super-Heater.

Fig. 1. Schematic for the typical studied PTSTPP.

2. Description and modeling of the studied plant Table 1


Technical details of PTSPP of the solar field [25].

The proposed PTSTPP is of 50 MWe capacity, which referred to a Solar collector field Receiver Collector
relevant size of most existed commercial plants. It consists of: solar Schott PTR Solargenix
2008 SGX-1
field (SF) with large parabolic trough collectors, power block (PB),
thermal energy storage system (TES) and fuel back-up system Aperture width (m) – 5
(FBS) to produce supplement energy to the PB during low/non- Length of collector assembly (m) – 100
Focal length (m) – 1.8
solar time. Fig. 1 presents the schematic for the typical studied Number of modules per assembly – 12
plant. Absorber tube inner diameter (m) 0.066 –
SAM (Solar Advisor Model) developed by the National Renew- Absorber tube outer diameter (m) 0.070 –
able Energy Laboratory (NREL) was chosen as a modeling and sim- Glass envelope inner diameter (m) 0.115 –
Glass envelope outer diameter (m) 0.120 –
ulation software. SAM uses TRNSYS software combined with the
Optical efficiency of the receiver 0.8692 –
Solar Thermal Electric Components (STEC) model library as a tran- Heat loss at design DNI and temperature 166.25 –
sient simulation code, the software was previously validated by (W/m2)
Price [18], and used by many researchers [19–21]. The analytical Solar field layout (number of subsections) H (4)
model used in this software was presented by Wagner and Gilman
[22]. Since the detailed model is exhaustive and hard to be pre-
sented in the paper, just a general modeling of our proposed plant A very important parameter in our study is the solar multiple
was given. (SM). It is defined as the ratio of thermal power supplied by the
For simulation and optimization of the studied plant, typical solar field at design point, to thermal power required by the power
meteorological year (TMY) data of Bechar in the southern of Algeria block at nominal conditions [3], and presented by:
(latitude 31.380 N, longitude 2.150 W, altitude 806 m) with one-hour
Q_ th SF
interval yearly data set (8760 h) was chosen, due to its high mean SMdesign ¼ ð1Þ
Q_ th
point
annual direct normal irradiance Ib (greater than 2500 kWh/m2). PB

2.2. Thermal energy storage (TES) and fuel backup system (FBS)
2.1. Solar field (SF)
TES and FBS are integrated to store heat from solar field during
The solar field with ‘H’ layout [23], is made up of large scale daylight and dispatch it during non/low solar radiation time, this
parabolic trough collector assemblies of SGX1 type, each assembly strategy increases the value of power generated and number of
of 100 m length, 5 m width. These assemblies were placed in par- operation hours or capacity factor compared to stand alone plant.
allel rows in North-South axis horizontal position, and track the Our TES system consists of two tanks of solar salt (60%
sun from East-West [24]. A heat transfer fluid (HTF), Therminol NaNO3 + 40% KNO3). The full load hours of the TES which indicate
VP-1 with a temperature range of 293–393 °C, is heated up by solar the number of hours that thermal storage can supply energy for the
radiation collected by the solar field, and used for thermal heat operation of the power block at the design input level, is given as
transportation to thermal energy storage, fuel backup and finally [22,26]:
to power block for electricity generation and supply it to the grid.
wdes Dtes
The technical details of the solar field considered in this study were Etes ¼ ð2Þ
summarized in Table 1.
gcycles des
T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218 1213

where wdes presents design cycle thermal requirement (kW), Dtes is Table 2
total number of desired storage hours (h), and gcycles des is design The common assumptions and nominal values of the design for the main parameters
in the power block.
point cycle efficiency.
On the other side, the fuel burners or FBS (with natural gas) can Parameter Value
supply thermal energy directly to the HTF in the power block, in Inlet HTF conditions to power block:
order to maintain the plant power generation at design point.  Temperature (°C)/Pressure (bar) 393/20
 Mass flow rate (kg/s) 567.1
Indeed, the backup system provides energy whenever the thermo-
dynamic state required at the turbine inlet is not realized. That Outlet HTF conditions to power block:
 Temperature (°C)/Pressure (bar) 296/15
means when the solar field and TES contributions fall below the
 Mass flow rate (kg/s) 567.1
design point requirement, the FBS supplies energy up to the design
Inlet steam conditions to HP turbine:
point requirement while maintaining a total auxiliary contribution
 Temperature (°C)/Pressure (bar) 371/100
less than fossil fill fraction, which is defined as follows [27]:  Mass flow rate (kg/s) 57.55

Q_ backup ¼ m
_ PB ðhTi  hSFO Þ ð3Þ Inlet steam conditions to LP turbine:
 Temperature (°C)/Pressure (bar) 370/16.5
where m_ PB and hTi are mass flow rate of the HTF in the PB (kg/s), and  Mass flow rate (kg/s) 46.01
enthalpy at the inlet of high pressure turbine (kJ/kg) respectively. Outlet steam conditions from LP turbine:
And, the needed fossil fuel fraction is given by:  Temperature (°C)/Pressure (bar) 45.01/0.096
 Mass flow rate (kg/s) 38.35
Q_ backup HPT isentropic efficiency 0.85
f backup ¼ ð4Þ
Q_ tot LPT isentropic efficiency 0.88
Mechanical efficiency for HPT/LPT 0.99
where f backup is fossil fill fraction ð0 < f backup < 1Þ, and Q_ total is total Leaving losses at each stage of HPT/LPT (kJ/kg) 4
Feedwater pump isentropic efficiency 0.80
energy needed to reach the required thermodynamic state (kWh). Condensate pump isentropic efficiency 0.80
Terminal temperature difference in the feedwater 5
2.3. Power block (PB) heaters and the condenser (°C)
Primary pressure drop in the feedwater heaters 0.05
and the condenser (bar)
The power block of our study is considered to operate using a
Generator efficiency 0.97
steam regenerative Rankine cycle of 50 MW. The HTF passes Pressure drop (PH, SH and RH) in steam/HTF sides (bar) 0.5/1.6
through a series of heat exchangers such as super-heaters, evapo- Terminal temperature in SH and RH (°C) 23
rators and finally a pre-heater (Fig. 1). While the thermal efficiency Approach temperature difference in the SG (°C) 3
Overall efficiency of PB 0.3726
of the power cycle is increased by including closed type feed water
heat exchangers (FWH 1-5), the optimum quantity of it is based on
economical optimization. Therefore, according to the size of our
cycle, it is advisable to use five heaters [28].
Table 3
In this study, as the analysis of hourly and annual power gener- Costs data for the economic model [3,25,26,31].
ation using PTSTPP technology like all concentrating solar power is
highly affected by the quality of off-design performance character- 1. Direct capital cost
 Site improvement 30.00 $/m2
ization. The power cycle was evaluated and simulated using com-  Solar field 270.00 $/m2
mercial software EBSILON professional 10.06 [29], by considering it  HTF system 80.00 $/m2
as a control volume that contains all components of the PB pre-  TES system 80.00 $/kWht
sented in Fig. 1, and excludes the heat rejection system (which  FBS system 60.00 $/kWe
 PB system 830.00 $/kWe
modeled by SAM with other components of the plant). The values
 Balance of the plant 105.00 $/kWe
of the overall thermal efficiency of the PB calculated by EBSILON  Contingency (of the direct capital cost) 7%
were incorporated as inputs in SAM to simulate the techno-
2. Indirect capital cost
economic performance of the whole plant that includes SF, TES,  EPC and Owner cost (of the direct capital cost) 11%
FBS, and PB with heat rejection system (Fig. 1).  Total land cost 10,000.00 $
Table 2 shows the common assumptions and nominal values of  Sales Tax applies to 80% of direct cost 5%
the design for the main parameters in the considered plant. At dif- 3. Operation and maintenance costs
ferent part loads, the off-design of the plant in the power block is  Fixed cost by capacity 65.00 $/kW-yr
also considered in the software, and for any given cycle design,  Variable cost by generation 3.00 $/MWh
 Fossil fuel cost 3.00 $/MMBtu
the thermal performance is a function of three independent vari-
ables: heat transfer fluid inlet temperature and mass flow rate,
and condenser steam pressure. Deviation among these three vari-
ables from their design-point values adversely impacts the heat-
to-power conversion efficiency of the cycle [22]. The optimization model inputs was set according to previous studies and databases
of the studied plant is based only on working parameters related to [3,25,26,31]. These inputs were summarized in Table 3. The eco-
solar field and thermal energy storage, while the optimized design nomic assessment was performed using the software SAM, where
of the power block was proposed based on previous studies the LCOE is calculated as:
[1–4,9,19].
crf  C inv þ C O&M  C env
LCOE ¼ ð5Þ
2.4. Economic model PGnet

It is well known that the levelized cost of electricity (LCOE) is where C inv is total investment cost (US$), C O&M is annual operating
the most commonly used indicator to study the feasibility of any and maintenance costs (US$), C env is environmental cost according
solar thermal power plant [30]. This parameter can be significantly to CO2 rejected (US$), and PGnet is annual net power generation
affected by the assumptions and inputs of the economic model, our (kWh). While crf is capital recovery factor, and calculated as:
1214 T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218

kd  ðkd þ 1Þ
N 1
crf ¼ ð6Þ f ðzeÞ ¼ ð11Þ
½ðkd þ 1Þ  1
N 1 þ eze
where ze is weighted sum, and presented in terms of bias ðbÞ,
weight ðwÞ, and output ðyÞ as:
3. ANN modeling of the studied plant
X
n
ANN method proposes a mathematical model to tackle com- zej ¼ wij yi þ bj ð12Þ
plex, noisy and incomplete data, in order to create a tool with i¼1
learning capabilities and used for prediction and optimization, as
well as classification problems. This method performs dynamic The different algorithms were performed using MATLAB software
or non-linear problems at a high speed, and can be a good choice with an increased number of neurons, to define the output accu-
for modeling energy conversion systems [11]. rately in a single hidden layer for 1000 epochs. The obtained results
ANN is carrying out in two main steps including training and by MATLAB were exported to Excel spreadsheets to perform addi-
testing of data. The unit element of an ANN is the neuron. As in nat- tional calculations and evaluations. The summarized results of the
ure, the network function is specified largely by the connections statistical performance evaluation were presented in Tables 4 and
between the elements [32,33]. To obtain the relationships between 5 for training and testing data respectively.
inputs and outputs, several learning algorithms are available. While the ideal values of statistical tests such as RMSE, MPE,
While the most commonly used algorithm is the feed-forward and COV are 0 or closer to 0, the best values of R2 should be 1 or
back-propagation learning algorithm, the most widely used vari- closer to 1. As shown in Tables 4 and 5, LM algorithm is very satis-
ants to be adapted for this algorithm in the energy systems field factory compared to other algorithms, and can be used for the pre-
are Levenberge Marguardt (LM), Scaled Conjugate Gradient (SCG), feasibility of this type of plants with an acceptable accuracy. More-
and Pola-Ribiere Conjugate Gradient (CGP). ANN with a feed- over, the LM training model with 38 neurons in a single hidden
forward back propagation algorithm learns by changing the con- layer looks as the best ANN model to predict PGnet and LCOE of
nection weights, and these changes are stored as knowledge. the presented PTSTPP. Indeed, this model presents the highest
In this study, design ambient temperature ðT amb Þ, design solar coefficients of determination with values of 0.9999 for power gen-
radiation ðIb Þ, solar multiple ðSMÞ, row spacing between parallel eration and 0.9998 for LCOE. Furthermore, the lowest values of
collectors ðlÞ and finally full load hours of the thermal energy sys- other statistical parameters (RMSE, MPE and COV) were derived
tem ðDtes Þ were selected as inputs with the ranges of 20–38 °C, by the same model as 0.6108, 0.1185 and 0.2877 for PGnet and as
750–950 W/m2, 1–4, 10–31 m, and 0–12 h, respectively. The calcu- 0.0358, 0.1574 and 0.2995 for LCOE. The architecture of the best
lated annual power generation ðPGnet Þ and LCOE using SAM and topology (LM 40) is shown in Fig. 2.
EBSILON software packages were determined as outputs. The data Moreover, the comparison of simulated PGnet data using SAM
of 1024 designs of the studied plant (with the inputs supposed and and those obtained by this best model of ANN is shown in
outputs calculated) are normalized in the range of (0–1). Then, 70% Fig. 3a. While Fig. 3b presents the comparison of the simulated
of these data in a number of 717 designs was selected for training LCOE data using SAM and those of ANN best model. The LCOE
while the remained was used for testing. The ANN model with and PGnet are varying as the design of the parabolic trough solar
back-propagation algorithm is using supervised learning, which thermal power plant is changing, every design of 1024 ones corre-
means that we provide the model with the data of the inputs sponds to a change in the value of one of the five inputs presented
and outputs we want the network to compute, and then the errors in the paper. As the LCOE is affected by the annual power genera-
between actual results (calculated by the software products) and tion as mentioned in Section 2.4, the last factor is highly influenced
those expected (determined by ANN) are calculated. The goal of by the variations in different inputs included in the paper, they
this process is to reduce these errors to the minimal. affected in the optical and thermal efficiencies of the plant. These
The performance of the studied PTSTPP was calculated using variations in the LCOE and PGnet ranging from high values of more
four different statistical parameters, namely, root mean square than 350 GWh for annual power generation and more than
error (RMSE), mean percent error (MPE), coefficient of variation 21 Cent/kWh for levelized cost of energy, to low values of less than
(COV) and the coefficient of determination (R2). These indicators 100 GWh for PGnet and less than 9 Cent/kWh for LCOE.
were formulated in terms of output value ðyoutput Þ, target value
actual Þ and pattern ðnÞ as:
ðyactual Þ, average of target ðy
4. Optimization of the plant
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 Xn  2
RMSE ¼ youtput  yactual ð7Þ As mentioned above, five parameters (Tamb, Ib, SM, l, and Dtes)
n i¼1
were taken into consideration to find the optimum design. These
five parameters were chosen because the physical behavior of
!
1X n
youtput  yactual CSP is highly affected by these parameters as;
MPE ¼ ð8Þ 1. Ambient temperature at design Tamb influences the cooling
n i¼1 youtput
section of the power block then the whole system (condenser
Pn   and PB performances vary significantly as the ambient tempera-
i¼1
output ðyactual  y
youtput  y actual Þ ture varies). This parameter is affiliated directly with the heat
COV ¼  100 ð9Þ
n rejection system (condenser-cooling tower) and determines the
condenser’s heat exchange surface area, design, and performances
2 32
Pn   [25,26]. 2. Direct solar radiation at design (Ib) and solar multiple
6 
i¼1 youtput  youtput ðyactual  yactual Þ
 7
2
R ¼ 4qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 ð10Þ (SM), they both define the aperture area of the solar field. However,
Pn  2 Pn
i¼1 youtput  youtput
 
i¼1 ðyactual  yactual Þ
2 SM is related to the power block demand as mentioned in Sec-
tion 2.1, while design Ib is related to solar potential of the location
As mentioned above, LM, SCG and CGP variants of feed-forward where the plant is simulated. 3. Row spacing between parallel col-
back-propagation learning algorithm were applied, and the loga- lectors (l) has a direct influence on the solar field efficiency and its
rithmic sigmoid (Logsig) was used: thermal energy output. An increase in this distance results in the
T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218 1215

Table 4 Table 5
Comparison of error analysis of training data for ANN topologies. Comparison of error analysis of testing data for ANN topologies.

Algorithm Statistical results R2 RMSE MPE COV Algorithm Statistical results R2 RMSE MPE COV
PGnet PGnet
LM 20 0.9998 0.9857 0.2480 0.4643 LM 20 0.9998 0.8963 0.2908 0.4488
30 0.9999 0.7913 0.1906 0.3727 30 0.9999 0.7217 0.2200 0.3614
36 0.9999 0.7219 0.1651 0.3400 36 0.9999 0.7219 0.1651 0.3400
38 0.9999 0.6108 0.1185 0.2877 38 0.9999 0.6108 0.1185 0.2877
40 0.9999 0.6143 0.1161 0.2893 40 0.9999 0.5561 0.1590 0.2784
50 0.9999 0.6416 0.1337 0.3022 50 0.9999 0.6157 0.1804 0.3083
SCG 10 0.9897 7.0786 2.2154 3.3341 SCG 10 0.9844 7.9775 2.6911 3.9947
16 0.9875 7.7963 2.3788 3.6722 16 0.9836 8.2672 2.7083 4.1397
18 0.9900 6.9709 2.1687 3.2834 18 0.9847 7.9149 2.5273 3.9633
20 0.9904 6.8280 2.1114 3.2161 20 0.9855 7.7119 2.5042 3.8617
24 0.9893 7.1902 2.2205 3.3867 24 0.9849 7.9023 2.5444 3.9570
30 0.9883 7.5236 2.3947 3.5437 30 0.9883 7.5236 2.3947 3.5437
CGP 8 0.9863 8.1515 2.4715 3.8395 CGP 8 0.9807 8.8730 2.8531 4.4431
10 0.9902 6.8838 2.1539 3.2424 10 0.9867 7.3770 2.4427 3.6940
12 0.9923 6.1175 1.9183 2.8814 12 0.9906 6.2000 2.0344 3.1046
16 0.9874 7.8211 2.4112 3.6838 16 0.9789 9.2932 2.9788 4.6535
18 0.0437 313.5156 119.4946 147.6703 18 0.0706 300.2056 118.7459 150.3259
20 0.9839 8.8325 2.7301 4.1602 20 0.9839 8.8325 2.7301 4.1602
LCOE LCOE
LM 20 0.9986 0.0888 0.3633 0.7436 LM 20 0.9992 0.0681 0.4544 0.5670
30 0.9993 0.0631 0.2732 0.5289 30 0.9980 0.1056 0.4650 0.8729
36 0.9993 0.0624 0.2613 0.5231 36 0.9990 0.0750 0.4981 0.6242
38 0.9998 0.0358 0.1574 0.2995 38 0.9995 0.0524 0.2740 0.4362
40 0.9997 0.0382 0.1461 0.3203 40 0.8744 0.9079 0.6988 7.5595
50 0.9998 0.0326 0.1827 0.2734 50 0.9905 0.2318 0.6146 1.9303
SCG 10 0.9737 0.3816 2.3606 3.1973 SCG 10 0.9750 4.8890 2.5958 10.2485
16 0.9631 0.4522 2.8080 3.7883 16 0.9602 0.4699 3.3618 3.9120
18 0.9762 0.3631 2.2249 3.0418 18 0.9748 0.3737 2.6759 3.1111
20 0.9689 0.4152 2.6576 3.4782 20 0.9551 0.5036 3.6463 4.1931
24 0.9786 0.3441 2.1794 2.8825 24 0.9743 0.3802 2.6342 3.1656
30 0.9700 0.4078 2.5543 3.4167 30 0.9700 0.4078 2.5543 3.4167
8 0.9511 0.5206 3.2569 4.3611
CGP 8 0.9385 7.6281 4.0998 15.9906
10 0.9672 0.4262 2.6275 3.5709
10 0.9643 5.8052 3.1965 12.1693
12 0.9724 0.3911 2.4763 3.2762
12 0.9743 0.3795 2.7273 3.1595
16 0.9724 0.3911 2.4330 3.2766
16 0.9658 0.4356 3.0184 3.6267
18 0.9749 0.3728 2.3411 3.1233
18 0.9705 0.4047 2.9723 3.3692
20 0.9625 0.4558 2.8706 3.8183
20 0.9625 0.4558 2.8706 3.8183

decrease of the shading effect, as a consequence in optical effi-


ciency of the solar field. On the other hand, the increase of this dis-
tance causes an increasing in the thermal losses of the piping lines,
therefore on thermal performance of the solar field. 4. The total
number of the storage hours or TES capacity (Dtes) has an impor-
tant influence on the annual power generation of the plant, as this
last parameter is greater for higher thermal storage capacities,
therefore it delivers the power block with more energy during
low or non-solar times, thus enhancing its dispatch and potential
to operate for longer time intervals.
Furthermore, the increasing in the values of these parameters
causes an increasing in the investment costs of the plant, thus on
the LCOE. The optimization involves finding the combination of
these inputs that minimize the LCOE which is the most commonly
used indicator to study the feasibility of any solar thermal power
plant as stated before.
In the optimization process, the data obtained by means of the
weights of the best ANN model (LM with 38 neurons) were used in Fig. 2. The architecture of the best ANN topology.
the LCOE equation. It can be noted from Fig. 3b, that the optimum
design of the plant located within the first 200 designs which cor-
responding to variation ranges of 32–38 °C for Tamb, 17–31 m for l,
750–950 W/m2, 1–3, and 0–7 h for both Ib, SM, and Dtes respec- From the results of these economical values, it can be seen that
tively. The LCOE variations for different design solar radiation, solar the optimum plant design is the one that can get the minimum
multiple, and number of hours of storage are presented in Figs. 4– LCOE. Taking the current variation of these inputs values into
8. While for the two other parameters; design ambient tempera- account, a benefit ranging between 8.99 and 14.07 dollar Cent/
ture and row spacing, the optimum points were calculated using kWh can be obtained for solar radiation at design of 750 W/m2
the ANN model weights in Excel spreadsheet, and presented in as presented in Fig. 4. For Ib = 800 W/m2, this range becomes 8.88
Table 6. to 13.40 dollar Cent/kWh as showed in Fig. 5. The same observa-
1216 T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218

(a)
400
ANN
350 Real
Power Generation (GWh)

300

250

200

150

100

50
0 100 200 300 400 500 600 700 800 900 1000
Pattern Fig. 5. Variation of the LCOE for different solar multiple, and number of hours of
storage for Ib = 800 W/m2.

(b)
22
ANN
20 Real

18
LCOE (Cent/kWh)

16

14

12

10

8
0 100 200 300 400 500 600 700 800 900 1000
Pattern

Fig. 3. The comparison of ANN prediction and simulated design results; (a) PGnet Fig. 6. Variation of the LCOE for different solar multiple, and number of hours of
and (b) LCOE. storage for Ib = 850 W/m2.

Fig. 4. Variation of the LCOE for different solar multiple, and number of hours of
Fig. 7. Variation of the LCOE for different solar multiple, and number of hours of
storage for Ib = 750 W/m2.
storage for Ib = 900 W/m2.

tion can be noted in Figs. 6–8 for Ib = 850 W/m2, Ib = 900 W/m2, and mum solution to get for 34 °C, 850 W/m2, 23 m, 1.7, and 2.5 h for
Ib = 950 W/m2 with different ranges of (8.82–12.85, 8.84–13.00, design ambient temperature and solar radiation, row spacing, solar
8.88–13.35 dollar Cent/kWh) respectively. It is therefore, the opti- multiple and number of hours for TES respectively, which gives a
T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218 1217

350 60
Q Tot
300 Q PB
40
Q SF

Power to/from TES (MWh)


250 Q FBS
Q Net 20

Power (MWh)
200 Q TES
0
150

-20
100

-40
50

0 -60
0 3 6 9 12 15 18 21 24
Time (Hour)
Fig. 8. Variation of the LCOE for different solar multiple, and number of hours of
Fig. 9. Mean hourly power balances variations through the year of the new
storage for Ib = 950 W/m2.
optimized plant.

Table 6
Table 8
Design ambient temperature and row spacing distance optimum points for different
Investment costs of the new optimized plant.
solar radiation.
Total installed cost 303.8 M$
Ib (W/m2) 750 800 850 900 950
1. Direct capital cost 258.42 M$
Tamb (°C) 33 33 34 34 34  Site improvement 12.41 M$
l (m) 19.5 22 23 25 25  Solar field 111.74 M$
 HTF system 33.11 M$
 TES system 29.52 M$
 FBS system 3.30 M$
Table 7  PB system 45.66 M$
Percentage errors between the calculated values of LCOE and PGnet using ANN model  Balance of the plant 5.78 M$
and those obtained by SAM software.  Contingency (of the direct capital cost) 16.91 M$
Ib design 750 800 850 900 950 2. Indirect capital cost 45.35 M$
Operation and maintenance costs 55.14 M$
LCOE
Results with SAM 8.84 8.86 8.92 8.94 8.92
Results with ANN 8.99 8.88 8.82 8.84 8.88
Error (%) 1.67 0.23 1.13 1.13 0.45 solar radiation. On the other side, the FBS and TES systems feeds
PGnet the plant with the necessary energy to operate it during the night
Results with SAM 194.586 194.544 196.616 190.389 192.312 time. For the economic assessment of the new optimized plant, and
Results with ANN 192.772 194.230 198.698 191.807 191.581 as presented in Table 8, the main cost driver of the new system
Error (%) 0.941 0.162 1.048 0.739 0.382
installation is determined by the solar field with a percentage of
more than 33.7% compared to the total installed cost. While the
O&M costs that includes natural gas and water for the operation,
minimum LCOE of 8.82 Cent/kWh and annual power generation of as well as preventive maintenance of the plant, share an important
198.698 GWh. proportion of the total investment cost of the plant with more than
As shown in Table 7, the percentage errors between the calcu- 55 million dollars.
lated values of LCOE and PGnet using ANN model and those obtained
by SAM software of this optimum designs, varies between very low
values of 0.23% and 0.162% to 1.67% and 1.048 for both LCOE and
PGnet respectively. This means that the investigated ANN model is 5. Conclusion
very satisfactory to be used for the pre-feasibility of this type of
plants with a high accuracy. A highly unique, flexible ANN algorithm was presented to pre-
The mean hourly power balances variations through the year of dict and optimize a parabolic trough solar thermal power plant
the new optimized plant are reported as functions of working time using Thermic oil as HTF and integrated with TES and FBS.
of the plant and plotted in Fig. 9, in terms of: total solar energy col- The results obtained using SAM software were used to train sev-
lected by the solar field Q Tot , solar field thermal output Q SF , thermal eral ANN algorithms, such as LM, SCG, and CGP. The trained algo-
energy received by the power block Q PB , thermal energy delivered rithms were then tested and evaluated using four different
by the FBS Q FBS , net electric power output Q Net , and thermal energy statistical indicators; RMSE, MPE, COV and R2. The obtained statis-
to/from the TES Q TES (in blue1). Moreover, the investment costs of tical parameters showed that LM algorithm is very satisfactory
the new optimized plant are summarized in Table 8. compared to other algorithms, and can be used for the pre-
As it can be seen in Fig. 9, Q Tot , Q SF and Q PB variations are pro- feasibility of this type of plants with an acceptable accuracy. More-
portional to direct normal irradiance received by the solar field, over, the LM training model with 38 neurons in a single hidden
as the energy flows in the system increase by the increasing of layer looks as the best ANN model to predict PGnet and LCOE of
the presented PTSTPP. This model presents the highest coefficients
of determination (0.9999 for power generation) and (0.9998 for
1
For interpretation of color in Fig. 9, the reader is referred to the web version of LCOE). Furthermore, the lowest values of other statistical parame-
this article. ters (RMSE = 0.6108, MPE = 0.1185, and COV = 0.2877 for PGnet ) and
1218 T.E. Boukelia et al. / Applied Thermal Engineering 107 (2016) 1210–1218

(RMSE = 0.0358, MPE = 0.1574, and COV = 0.2995 for LCOE) were [9] L. Martín, M. Martín, Optimal year-round operation of a concentrated solar
energy plant in the south of Europe, Appl. Therm. Eng. 59 (1) (2013) 627–633.
derived by the same model.
[10] T. Bekat, M. Erdogan, F. Inal, A. Genc, Prediction of the bottom ash formed in a
In the second part of the study, the obtained weights from the coal-fired power plant using artificial neural networks, Energy 45 (1) (2012)
best ANN topology were used in the LCOE analysis for determining 882–887.
the optimum system design. It is therefore available to get a min- [11] S.A. Kalogirou, Applications of artificial neural-networks for energy systems,
Appl. Energy 67 (1) (2000) 17–35.
imum LCOE of 8.88 Cent/kWh from the new optimized plant when [12] S.A. Kalogirou, M. Bojic, Artificial neural networks for the prediction of the
the plant characteristics are; 34 °C and 850 W/m2 for both design energy consumption of a passive solar building, Energy 25 (5) (2000) 479–491.
ambient temperature and solar radiation, 23 m for row spacing, [13] S.A. Kalogirou, C.C. Neocleous, C.N. Schizas, Artificial neural networks for
modelling the starting-up of a solar steam-generator, Appl. Energy 60 (2)
1.7 for solar multiple, and 2.5 h number of hours for TES. As the (1998) 89–100.
yield and techno-economic behavior of any solar thermal power [14] O. Arslan, Power generation from medium temperature geothermal resources:
plant can be affected by the difference in solar resources between ANN-based optimization of Kalina cycle system-34, Energy 36 (5) (2011)
2528–2534.
a site and another. Therefore, the optimal design can be slightly [15] O. Arslan, O. Yetik, ANN based optimization of supercritical ORC-binary
changed from a site and another. However, the paper can present geothermal power plant: Simav case study, Appl. Therm. Eng. 31 (17) (2011)
a good methodology for prediction and optimization of a solar 3922–3928.
[16] O. Arslan, ANN-based determination of optimum working conditions of
thermal power plant based on ANN approach. residential combustors with respect to optimum insulation, Energy Source
This algorithm can be used easily for the design and optimiza- Part A 36 (23) (2014) 2603–2612.
tion of complex solar thermal power plants. ANN can be solved [17] O. Arslan, O. Yetik, ANN modeling of an ORC-binary geothermal power plant:
simav case study, Energy Source Part A 36 (4) (2014) 418–428.
many configurations and fluids other than the presented in the
[18] H. Price, A parabolic trough solar power plant simulation model, in: ASME
paper. A Solar power plant design without the ANN is very complex 2003 international solar energy conference, Hawaii, 2003, pp. 665–673.
and time consuming. Thus, the ANN is more advantageous in a [19] S.A. Kalogirou, Solar thermoelectric power generation in Cyprus: selection of
solar power plants topic. The LM variant of ANN algorithm can the best system, Renew. Energy 49 (2013) 278–281.
[20] W. Pierce, P. Gauché, T. Von Backström, A.C. Brent, A. Tadros, A comparison of
be employed to determine the techno-economic performances of solar aided power generation (SAPG) and stand-alone concentrating solar
such plants, thereby reducing the engineering efforts. power (CSP): a South African case study, Appl. Therm. Eng. 61 (2) (2013) 657–
The present optimization study may hold good, by using this 662.
[21] V. Poghosyan, M.I. Hassan, Techno-economic assessment of substituting
number of designs. However, a rigorous study, including a higher natural gas based heater with thermal energy storage system in parabolic
number of designs can improve the accuracy of the ANN obtained trough concentrated solar power plant, Renew. Energy 75 (2015) 152–164.
results. Furthermore, as a part of future work, it is advisable to per- [22] M.J. Wagner, P. Gilman, Technical Manual for the SAM Physical Trough Model.
<nrel.gov/docs/fy11osti/51825.pdf> (last accessed 07/04/2015).
form a coupled multi-objective optimization methodology involv- [23] B. Kelly, D. Kearney, Parabolic Trough Solar System Piping Model, Golden, CO,
ing ANN and genetic algorithm to determine the maximum Technical Report No. NREL/SR-550-40165, 2006. <http://large.stanford.edu/
possible 4E (energy-exergy-economic-environment) benefits of publications/power/references/troughnet/solarfield/docs/40165.pdf> (last
accessed 07/04/2015).
such plants, with high energy and exergy efficiencies, as well as [24] T.E. Boukelia, M.S. Mecibah, Estimation of direct solar irradiance intercepted
annual power generation, in addition to low investment cost and by a solar concentrator in different modes of tracking (case study: Algeria), Int.
environmental impacts. J. Amb. Energy (2013) (in press). http://dx.doi.org/10.1080/01430750.2013.
864587.
[25] Solar Advisor Model SAM 2014.1.14, National Renewable Energy Laboratory
References (NREL), 2014. <sam.nrel.gov/> (last accessed 25/03/2015).
[26] N. Blair et al., System Advisor Model, SAM 2014.1.14, General Description,
[1] K.S. Reddy, K.R. Kumar, Solar collector field design and viability analysis of NREL Report No. TP-6A20-61019, National Renewable Energy Laboratory
stand-alone parabolic trough power plants for Indian conditions, Energy (NREL), USA, 19 pp.
Sustain. Dev. 16 (4) (2012) 456–470. [27] T. Larraȋn, R. Escobar, J. Vergara, Performance model to assist solar thermal
[2] N.B. Desai, S. Bandyopadhyay, Optimization of concentrating solar thermal power plant siting in northern Chile based on backup fuel consumption,
power plant based on parabolic trough collector, J. Clean. Prod. 89 (2015) 262– Renew. Energy 35 (8) (2010) 1632–1643.
271. [28] R.V. Padilla, Simplified Methodology for Designing Parabolic Trough Solar
[3] M.J. Montes, A. Abánades, J.M. Martinez-Val, M. Valdés, Solar multiple Power Plants PhD Thesis, University of South Florida, USA, 2012.
optimization for a solar-only thermal power plant, using oil as heat transfer [29] EBSILON Professional, Evonik Energy Services GmbH, 2011.
fluid in the parabolic trough collectors, Sol. Energy 83 (12) (2009) 2165–2176. [30] J. Dersch et al., Trough integration into power plants—a study on the
[4] J. Bonilla, L.J. Yebra, S. Dormido, E. Zarza, Parabolic-trough solar thermal power performance and economy of integrated solar combined cycle systems,
plant simulation scheme, multi-objective genetic algorithm calibration and Energy 29 (5) (2004) 947–959.
validation, Sol. Energy 86 (1) (2012) 531–540. [31] NREL, System Advisor Model (SAM) Case Study: Andasol-1. <sam.nrel.gov/
[5] A. Ramos, F. Ramos, Strategies in tower solar power plant optimization, Sol. sites/sam.nrel.gov/files/content/case_studies/
Energy 86 (9) (2012) 2536–2548. sam_case_csp_physical_trough_andasol-1_2013-1-15.pdf> (last accessed,
[6] A.L. Avila-Marin, J. Fernandez-Reche, F.M. Tellez, Evaluation of the potential of 2014, last accessed 17/07/2014).
central receiver solar power plants: configuration, optimization and trends, [32] L.M. Fu, Neural Network in Computer Intelligence, International Editions.,
Appl. Energy 112 (2013) 274–288. McGraw-Hill, New York, 1994.
[7] J. Spelling, D. Favrat, A. Martin, G. Augsburger, Thermoeconomic optimization [33] L.H. Tsoukalas, R.E. Uhrig, Fuzzy and Neural Approaches in Engineering, Wiley,
of a combined-cycle solar tower power plant, Energy 41 (1) (2012) 113–120. New York, 1997.
[8] J.M. Cabello, J.M. Cejudo, M. Luque, F. Ruiz, K. Deb, R. Tewari, Optimization of
the size of a solar thermal electricity plant by means of genetic algorithms,
Renew. Energy 36 (11) (2011) 3146–3153.

You might also like