Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 36 (2014) 623 – 628

Complex Adaptive Systems, Publication 4

Cihan H. Dagli, Editor in Chief
Conference Organized by Missouri University of Science and Technology
2014- Philadelphia, PA

Predicting Solar Irradiance Using Time Series Neural Networks

Alzahrani, A*. Kimball, J.W. Dagli, C
Missouri University of Science and Technology,

Abstract

Increasing the accuracy of prediction improves the performance of photovoltaic systems and alleviates the effects of intermittence on
the systems stability. A Nonlinear Autoregressive Network with Exogenous Inputs (NARX) approach was applied to the Vichy-Rolla
National Airport’s photovoltaic station. The proposed model uses several inputs (e.g. time, day of the year, sky cover, pressure, and
wind speed) to predict hourly solar irradiance. Data obtained from the National Solar Radiation Database (NSRDB) was used to
conduct simulation experiments. These simulations validate the use of the proposed model for short-term predictions. Results show
that the NARX neural network notably outperformed the other models and is better than the linear regression model. The use of
additional meteorological variables, particularly sky cover, can further improve the prediction performance.
© 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
ª

5IF
"VUIPST
1VCMJTIFE
CZ
&MTFWJFS
#7

(http://creativecommons.org/licenses/by-nc-nd/3.0/).
4FMFDUJPO
BOE
QFFSSFWJFX
VOEFS
SFTQPOTJCJMJUZ
PG
TDJFOUJGJD
DPNNJUUFF
PG
.JTTPVSJ
6OJWFSTJUZ
PG
4DJFODF
BOE
5FDIOPMPHZ
Peer-review under responsibility of scientific committee of Missouri University of Science and Technology
Keywords: Time Series Neural Network, NSRDB, NARX,

1. Introduction

Photovoltaic systems have been tremendously improved in the past few years. These improvements include those mad at the
operational level. These improvements have increased the amount of research focused on predicting solar energy. Predicting solar
radiation increases the economic benefit and reduces the renewable energy expansion constraints. Thus, models are needed that can
predict solar irradiation and improve the operations. A number of models are available that can estimate solar radiation. Both
physical models and stochastic models are most commonly used for these tasks. Physical models are good fit to sun radiation based
on its mean bias error (MBE) and root mean square error (RMSE) tests [1].

* Ahmad Alzahrani. Tel.: +001-970-430-7813.

E-mail address: [email protected]

1877-0509 © 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Peer-review under responsibility of scientific committee of Missouri University of Science and Technology
doi:10.1016/j.procs.2014.09.065
624 A. Alzahrani et al. / Procedia Computer Science 36 (2014) 623 – 628

The physical model used in this study estimates sunrise and sunset times without using extra equations (as other physical models do).
Starting by defining the declination δ and it was given for any day of the year as:

ª 360 (n  80) º
G 23.45o sin « »¼
(1)
¬ 365

where n is the day of the year, and 23.45° is the angle of the polar axis of the earth with the sun [10]. The indication took into account
that angles north of the equator are positive; they are negative when located south of the equator. Figure 1 illustrates the declination
at different times of the year. Time of solar noon can be calculated as:

M L  In
t 12  u 60 (2)
15
Where ɔ୐ is the longitude for the point of interest, ˗୬ is the longitude at which the solar noon occurs relative to the local time
zone, and (60 min/15°) is used because the earth needs 60 minutes to rotate 15 degrees. One essential connection exists between ф, δ,
and ω. This relationship is to determine the position of the sun in term of α at a given location, time and date was defined as in the
denominator of equation. Equation (3) was used to find the air mass coefficient (AM) to include the length of the path through the
atmosphere. This path length is typically compared to a vertical path located at sea level. Therefore, the air mass was thus calculated
as:
1
AM (3)
sin G sin I  cos G cos I cos Z

The sun radiation’s received energy of the sun radiation at a specific location, on a specific day and time, can be estimated by
I 1367 u (0.7) AM (4)

This equation is used to calculate unity AM. Nevertheless, according to the physical model [1], a better fit to the observed data is
given as:
0.678
I 1367 u (0.7) AM (5)

2. Prediction Using Neural Networks

There are many studies on how to predict solar irradiation using neural networks. In the literature, Raidal Bese Function Networks
was used to predict solar irradiance. However, in this paper, time series neural networks were used to perform the prediction. In this
study, a data set (containing time series measurements) was taken from the National Renewable Energy Laboratory in Golden,
Colorado. It was collected between January 1991 and December 2005, and contains 131,400 observations (year# = 8760
observations). All observation were made at the Vichy National Airport in Rolla, MO. Variables were chosen in the data were chosen
according to their variability. The data the information used are as in table 1.

Table 1: Different Inputs of Neural Networks

Experiment 1 Experiment 2 Experiment 3
Day (from 1 to 365) Day (from 1 to 365) Day (from 1 to 365)
Hour of the day (1 to 24) Hour of the day (1 to 24) Hour of the day (1 to 24)
Azimuth angle (degrees) Azimuth angle (degrees) Azimuth angle (degrees)
Zenith angle (degrees) Zenith angle (degrees) Zenith angle (degrees)
Wind speed (mph) Wind speed (mph)
Wind direction Wind direction
Temperature (C)
 A. Alzahrani et al. / Procedia Computer Science 36 (2014) 623 – 628 625

2.1 Visualizing Data and Preprocessing

Solar irradiation was visualized in one year to determine if there is a correlation between the inputs and the output. Solar irradiations
that were equal to zero were eliminated first. Doing so makes training data easier and aids in more readily identifying the neural
network’s pattern. This data presented in Fig. 4. Here, large portion of the year had a solar irradiation of zero.

Figure 1: Solar Irradiation in One Year

Figure 2 depicts the data after the zero irradiation was removed. This data illustrates that days are longer during the summer than
during the winter. After eliminating zero solar irradiation, all data points were normalized between -1 and 1 using the Matlab built-in
function Mapminmax. This normalized data will use Mapminmax to de-normalize the data and present them in the output.

Figure 2: Data after Zero Irradiation is removed

2.2 Neural Network Design

The structural design of the selected back propagation neural network is depicted in Fig, 3. This design uses time delay to predict
solar irradiation. A time delay neural network (TDNN) is a type of neural network. The primary aim of this type of networks is to
figure on serial information. The TDNN units recognize options freelance of time-shift (i.e., sequence position) and are typically
kind a part of a bigger pattern recognition system. They are most often used to change continuous audio into a stream of
classified sound labels for speech recognition. A sign is increased with delayed copies as different inputs. The neural network is time-
shift invariant since it has no internal state [5]. This network is referred to as NARX (nonlinear autoregressive with external input).
It is given previous values of a time series so that it can learn to predict future values. The input received from the feedback is known
as the external (or exogenous) time series [2].
626 A. Alzahrani et al. / Procedia Computer Science 36 (2014) 623 – 628

Figure 3. NARX Neural Network Architecture [10]

3. Case Studies

3.1 Case 1: Predicting Solar Irradiation Using Three Variables

Three variables were used in case 1 to predict irradiation: hours, zenith angle, and Azimuth angle. Feedback from the output was
also used.

Figure 4: MSE Error Figure 5: The predicted output vs. the actual output

This case gives very good result because the predicted output is matching the physical model. However, more cases were needed to
make sure that this networks performance under the noise (shading conditions) was very accepted. The regression in Fig. 5 illustrates
that the relationship between the predicted output and the actual output is approximately 1. Thus, the predicted output matches the
actual output. Figure 4 illustrates that there is a small error comparing. It is oscillates around zero.

3.2 Case 2: Predicting Solar Irradiation Using Five Variables

Five variables were used in Case 2 to predict solar Irradiation: hours, zenith angle, Azimuth angle, wind speed, and wind direction.
Feedback received from the output was also used.
 A. Alzahrani et al. / Procedia Computer Science 36 (2014) 623 – 628 627

Figure 6: MSE Error Figure 7: The predicted output vs. the actual output

3.3 Case 3: Predicting Solar Irradiance Using Six Variables

In these section different windows sizes which I used one back propagation training algorithm, as summarized in Table 2. A
maximum number of epochs (1,0000) was used; the training goal value was 0. The Levenberg-Marquardt algorithm was used to
reduce the error function. Overall, a window size of 3 is the best case where the smallest mean square error (MSE). Thus, this
windows size used was to evaluate additional test values and cases.

Figure 8: MSE Error Figure 9: The predicted output vs. the actual output

The predicted output was not better than that for either Case 1 or Case 2 because the temperature (which fluctuates) was included.

4. Evaluating Different Window Sizes and Performance

The network was evaluated under different window sizes (see Table 2). The mean square error was used to compare the experiments
to one another. The best performance occurred when three previous values were used and the convergence point was reached faster
than other cases.
628 A. Alzahrani et al. / Procedia Computer Science 36 (2014) 623 – 628

Table 2. Summary of Five different Windows Sizes

Algorithm Number of Neurons Hidden Layer Windows Size Output MSE

Levenberg-Marquardt (trainlm) 21 2 5.7%
Levenberg-Marquardt (trainlm) 21 3 2.2%
Levenberg-Marquardt (trainlm) 21 4 7.6 %
Levenberg-Marquardt (trainlm) 21 5 11.02 %
Levenberg-Marquardt (trainlm) 21 6 15.33 %

The MSE was used to evaluate all three cases (see Table 3). Again, the use of three variables was found to yield the best
performance because no strong correlation between temperature and irradiation exists (as in Case 3).
Table 3, Performance Evaluation of the cases
Case Number of Neurons Hidden Layer Window Size Output MSE
Case 1 21 3 5.7 %
Case 2 21 3 5.5 %
Case 3 21 3 8.2 %

5. Conclusions

The completed back propagation neural network was successfully predicted the solar irradiation. Time series neural networks
provided good predications as well. Three cases were conducted and tested. The output produced promising results. Different
window sizes were tested. The best case occurred when three window sizes were used. Future studies should be conducted in which
three networks are evaluated meticulously and compares them to other linear regression approaches, and other types of neural
networks.

References
1. Cococcioni, M.; D'Andrea, E.; Lazzerini, B., "24-hour-ahead forecasting of energy production in solar PV systems," Intelligent Systems Design and Applications
(ISDA), 2011 11th International Conference on , vol., no., pp.1276,1281, 22-24 Nov. 2011

2. NARX, Matlab http://www.mathworks.com/help/nnet/ref/narxnet.html

3. Mori, H.; Takahashi, M. "A prediction method for photovoltaic power generation with advanced Radial Basis Function Network", Innovative Smart Grid
Technologies - Asia (ISGT Asia), 2012
4. Illalva, M.G.; Gazoli, J.R.; Filho, E.R. "Comprehensive Approach to Modeling and Simulation of Photovoltaic Arrays", Power Electronics, IEEE Transactions
on, On page(s): 1198 - 1208 Volume: 24, Issue: 5, May 2009
5. Mori, H.; Takahashi, M. "Development of GRBFN with global structure for PV generation output forecasting", 2012 IEEE,
6. Hao Huang; Jin Xu; Zhenzhou Peng; Shinjae Yoo; Dantong Yu; Dong Huang; Hong Qin, "Cloud motion estimation for short term solar irradiation
prediction," Smart Grid Communications (SmartGridComm), 2013 IEEE International Conference on , vol., no., pp.696,701, 21-24 Oct. 2013
7. Mekki, H.; Mellit, A.; Salhi, H.; Khaled, B., "Modeling and simulation of photovoltaic panel based on artificial neural networks and VHDL-
language," Electronics, Circuits and Systems, 2007. ICECS 2007. 14th IEEE International Conference on , vol., no., pp.58,61, 11-14 Dec. 2007
doi: 10.1109/ICECS.2007.4510930

8. Alexander Waibel et al, Phoneme Recognition Using Time-Delay Neural Networks IEEE Transactions on Acoustics, Speech and Signal Processing, Volume 37,
No. 3, pp. 328. - 339 March 1989.
9. Nguyen, D.D.; Lehman, B.; Kamarthi, S., "Solar Photovoltaic Array's Shadow Evaluation Using Neural Network with On-Site Measurement," Electrical Power
Conference, 2007. EPC 2007. IEEE Canada , vol., no., pp.44,49, 25-26 Oct. 2007
10. Singh, V.P.; Vaibhav, K.; Chaturvedi, D.K., "Solar power forecasting modeling using soft computing approach," Engineering (NUiCONE), 2012 Nirma
University International Conference on , vol., no., pp.1,5, 6-8 Dec. 2012

11. National Renewable Energy Laboratory http://www.nrel.gov/midc/srrl_bms/