Professional Documents
Culture Documents
Renewable Energy
journal homepage: www.elsevier.com/locate/renene
Department of Mechanical Engineering and Materials Science and Engineering, Cyprus University of Technology, P.O. Box 50329, 3603 Limassol, Cyprus
Solar & Other Energy Systems Laboratory, DEMOKRITOS National Center for Scientic Research, 15310 Agia Paraskevi, Attikis, Greece
a r t i c l e i n f o
a b s t r a c t
Article history:
Received 15 May 2013
Accepted 30 August 2013
Available online 27 September 2013
In this paper, articial neural networks (ANNs) are used for the performance prediction of large solar
systems. The ANN method is used to predict the expected daily energy output for typical operating
conditions, as well as the temperature level the storage tank can reach by the end of the daily operation
cycle. These are considered as the most important parameters for the user. Experimental measurements
from almost one year (226 days) have been used to investigate the ability of ANN to model the energy
behavior of a typical large solar system. From the results, it can be concluded that the ANN effectively
predicts the daily energy performance of the system; the statistical R2-value obtained for the training and
validation data sets was better than 0.95 and 0.96 for the two performance parameters respectively. The
data used in the validation were completely unknown to the ANN, which proves the ability of the ANN to
give good predictions on completely unknown data. The results obtained from the method were also
compared to the inputeoutput model predictions with good accuracy whereas multiple linear regression
could not give as accurate results. Additionally, the network was used with various combinations of input
parameters and gave results of the same order of magnitude as the suggested method, which prove the
robustness of the method. The advantages of the proposed approach include the simplicity in the
implementation, even when the characteristics of the system components are not known, as well as the
potential to improve the capability of the ANN to predict the performance of the solar system, through
the continuous addition of new data collected during the operation of the system.
2013 Elsevier Ltd. All rights reserved.
Keywords:
Articial neural networks
Large solar thermal systems
Daily energy output
Maximum temperature of storage tank
1. Introduction
The development of large capacity installations for the exploitation of solar thermal energy can be characterized as critical for
the broadening of solar thermal energy applications [1]. Such an
expansion though, seems to come up against the suspicion of the
potential users regarding the expected energy benet, and consequently the reciprocity of the required investment. From this point
of view, the realistic evaluation of the energy performance of large
solar thermal systems (LSTS) constitutes a critical parameter, not
only for the candidate investor, but also for any type of supporting
actions provided by the state.
In the case of small, standardized, factory-made domestic
systems, the demand for reliable information regarding the expected energy output is satised by national or regional
* Corresponding author. Tel.: 357 2500 2621; fax: 357 2500 2637.
E-mail addresses: Soteris.kalogirou@cut.ac.cy, skalogir@spidernet.com.cy
(S.A. Kalogirou).
0960-1481/$ e see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.renene.2013.08.049
user to shape a realistic view for the reciprocity of the investment required, and, eventually, to redesign some elements of
the energy system (e.g. changing of the capacity of the auxiliary
sources).
- The analysis of potential signicant deviations between the
predictions and the actual performance, may lead to the
detection of failures related to the implementation phase, and to
the elaboration of corrective actions.
- The continuous comparison of actual energy performance to the
predicted one, or the one which had been initially veried, can
be proved to be valuable for the on time detection and correction of potential malfunctions which have appeared in due
course. Such a comparison could be implemented on an automated basis by the control system, which would also undertake
the activation of the proper alarms in case of a problem.
Moreover, the need for extending the certication of solar
thermal systems on large capacity custom-made systems has
turned the interest on methodological approaches which allow the
reliable assessment of the expected energy benet [4,5]. Within
this context, the discussion of the methodologies suitable for the
verication of energy performance of LSTS has been intensied in
recent years [6,7].
Thus, the comparison between expected and actual performance presupposes the existence of a simulation tool, which would
allow the reliable estimation of the expected energy benet of a
LSTS, as this system has been nally implemented. Such a tool can
belong in one of the following two main categories:
- Models which allow the dynamic simulation of the system
behavior on the basis of information related to its structure and
the characteristics of the individual subcomponents, e.g.
TRNSYS [2,3]. The quality of the predictions by such models is a
function of the reliability of the available information, noting the
usual signicant deviations between the nal actual conguration of a LSTS and the working hypotheses of the initial study.
These concern mainly the weather data used in the simulation,
usually in the form of a Typical Meteorological Year (TMY), and
the load prole assumed, which strongly depends on the habits
of the user/s and the type of application (hotel, industry, etc.). An
additional difculty is related to the requirement for specialized
knowledge and adequate experience by the personnel responsible for the operation of the specic software.
- Models, which treat the LSTS as a black box, and their application, can be performed on two distinct phases. In the rst phase,
i.e., training, experimental measurements taken through operation of the system are exploited, in order to shape an adequate
picture for the behavior of the system (mapping). The calculation procedure uses experimental information in order to
elaborate a kind of energy identity for the system, usually under
the form of coefcients of one or more characteristic equations.
Advantages of this kind of models include the relatively simple
use and the realistic assessment of the performance of the
system as it has been nally congured. As a disadvantage, one
may consider the inability to perform simulation for congurations or conditions of use signicantly different from the ones
used during the training phase [6].
Another factor, which inuences the quality of results of the
discussed models, is the operating conditions assumed. Given the
stochastic variation of meteorological conditions and their large
inuence on the energy behavior of the solar system, in order the
discussed comparison between expected and actual performance
to be realistic, this comparison has to be performed for certain
reference conditions, the same as the ones used during the
91
Value
Units
2.59
0.732
4.455
29
5.8
48
13,300
32,000
150
m2
[e]
Wm2 K1
kJK1
kg
[e]
W K1
l
W K1
92
Fig. 1. The solar system considered. (a) Schematic diagram of the solar system. (b) Position of the various thermocouples in the storage tank.
93
Such work is never done before, as these systems were not studied
in the appropriate depth.
3. The experimental setup
Fig. 3. Comparison between actual (measured) and ANN predicted values for Tsmax for the training data set.
94
Fig. 4. Comparison between actual (measured) and ANN predicted values for Q for the training data set.
Up to 10%
10e20%
20e30%
More than 30%
95.45%
4.55%
0%
0%
59.10%
27.25%
9.10%
4.55%
95
Fig. 5. Comparison between actual (measured) and ANN predicted values for Tsmax for the validation data set.
through the network. Three element inputs have been used corresponding to the values of the input parameters listed above and
the output parameters are two, Q and Tsmax. The learning procedure
was implemented by using the back-propagation algorithm. For the
training of the network, a learning rate and a momentum factor
needs to be specied by the user. Both of these constant terms are
specied at the start of the training cycle and determine the speed
and stability of the network. For this purpose, the learning rate was
set to a constant value of 0.1 and the momentum factor to 0.3. The
weights were initialized to a value of 0.3. The back-propagation
learning algorithm was used, described in Refs. [12,13]. The originators of back-propagation are Rumelhart [20] and Werbos [21].
As indicated before a total of 226 test patterns were available.
These concern actual measurements carried out as described in
Section 3 of this paper. From these patterns 184 were used for the
training of the ANN, and the rest 42 (20%) were randomly selected
to be used as test (20) and validation (22) patterns. Test patterns are
used during training to evaluate the training accuracy whereas the
validation patterns are used after the training phase to evaluate the
nal model accuracy.
The training was stopped when the average error, obtained by
comparing the actual and the ANN modeled data, remained constant for 200,000 events; i.e., about 1087 iterations through all data
(epochs) in the training dataset. This is considered a good value,
enabling the network to learn the input patterns satisfactorily and
to give good predictions while avoiding overtraining, which will
Fig. 6. Comparison between actual (measured) and ANN predicted values for Q for the validation data set.
96
Training
1800
Qsim
1600
1400
1200
1000
800
600
400
200
0
1
21
41
61
81
101
121
141
161
181
201
Patterns
Fig. 7. Comparison between actual (measured), ANN predicted values and inputeoutput method (Qsim) for Q for the training data set.
enable the ANN to map the data more accurately but lose the ability
of the ANN to generalize, i.e., to give good predictions for unknown
data sets.
The correlation coefcient obtained between the predicted and
training data set is 0.9546 and 0.9681 for the maximum water
temperature in the storage tank at the end of the day (Tsmax) and
the daily energy output of the system (Q) respectively. Both values
are very close to 1, indicating a very good mapping of the data.
Comparative plots of the actual (measured) and ANN predicted
values for the two output parameters for the training data set are
shown in Figs. 3 and 4 respectively. It should be noted that multiple
linear regression tried on the same data set give correlations of
0.9232 and 0.9382 for the above two parameters respectively,
which cannot be considered adequate.
Once a satisfactory degree of inputeoutput mapping was achieved, the network training was frozen and a set of completely
unknown test data was applied for verication. The validation of
the network was performed by using the unknown data for 22
cases of the validation data set. The correlation coefcients for the
unknown cases were of the same order of magnitude as in the
training phase and were equal to 0.9625 and 0.9560 for the two
output parameters respectively. The range of prediction accuracy
for the various patterns is shown in Table 2.
Comparative plots of the actual (measured) and ANN predicted
values for the two output parameters for the validation data set,
which includes data that are completely unknown to the ANN, are
shown in Figs. 5 and 6 respectively.
As can be seen from Figs. 5 and 6, the performance for both
output parameters is very satisfactory.
It should be noted that two different neural networks, one for
each output parameter, of the same architecture as shown in Fig. 2
but with one output neuron only give very comparable results, so
for the sake of compactness of the model the one ANN model with
two outputs as described before is preferred.
Various other tests have been performed to prove the
robustness of the method. The rst method concern the inpute
output modeling method described in Belessiotis et al. [7], which
is used both for the training and validation data sets to predict
the daily energy output of the system, Q. The results are shown in
Figs. 7 and 8 respectively. As can be seen from the graphs the
ANN model can give as accurate results as a simulation program
(Qsim) used for the modeling of the data. In fact the inpute
output method give slightly inferior t of the data with R2-values
equal to 0.9273 and 0.9327 for the training and validation data
sets respectively as opposed to values of about 0.96 for the ANN
(as indicated before).
Finally, various other combinations of input data were tried to
investigate the accuracy of daily energy output predictions. These
are shown in Table 3 together with the input data used in each case
and the correlation obtained. For all these models, the same
Validation
1600
1400
Qsim
1200
1000
800
600
400
200
0
1
10 11 12 13
14 15 16
17 18 19
20 21 22
Patterns
Fig. 8. Comparison between actual (measured), ANN predicted values and the inputeoutput method (Qsim) for Q for the validation data set.
Input data
R2-value
1
2
3
4
Taav, H, Tsin
H, DT
Tsm, H, DT
Tsmax, H, DT
0.9324
0.9310
0.9331
0.9325
Where: DT is temperature difference (Ta Tsin); Tsm is the mean storage tank
temperature; Other symbols are as described before.
architecture shown in Fig. 2, was used and the output in all cases
was Q, the daily energy output of the solar system.
As shown, all cases investigated give the same order to accuracy as the method suggested in this paper, which is based on
easily obtainable parameters without the need to carry out extra
estimations, like for DT or Tsm, and also gives slightly better accuracy with R2-values of about 0.96. This exercise however
proves the robustness of the method to provide reliable
predictions.
5. Conclusions
In the present paper, the use of ANN for the performance prediction of large solar systems is discussed. The analysis is concentrated on the two efciency indices, which are considered as the
most important for the user. These indices are the expected daily
energy output for typical operation conditions, as well as the
temperature level the storage tank can reach by the end of the daily
operation cycle.
Experimental measurements from almost one year (226 days)
have been used to investigate the ability of ANN to model the energy behavior of a typical large solar system. From the results
presented, it can be concluded that the ANN effectively predicts the
daily energy performance of the system; the statistical R2-value
obtained for the training data set was better than 0.95 and 0.96 for
the two performance parameters respectively. The validation also
included investigation of the sufciency of the trained ANN when
having to cope with unknown data, i.e., data the network has not
seen before. The results obtained from the method were also
compared to the inputeoutput model predictions with good accuracy. Additionally, the network was used with various combinations of input parameters and gave results of the same order of
magnitude as the suggested method, which prove the robustness of
the method.
The advantages of the proposed approach include the simplicity
in the implementation, even when the characteristics of the system
components are not known, as well as the potential to improve the
capability of the ANN to predict the performance of the solar system, through the continuous integration of new data.
97