CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Energy Procedía 74 (2015) 1205 - 1212

International Conference on Technologies and Materials for Renewable Energy, Environment and

Sustainability, TMREES15

Using Artificial Neural Networks for Prediction of Global Solar Radiation in Tehran Considering Particulate Matter Air Pollution

Masoud Vakilia, Saeed-Reza Sabbagh-Yazdib*, Koosha Kalhorb, Soheila Khosrojerdic

aDepartment of Mechanical Engineering, Iran University of Science and Technology, Tehran,1311416846, Iran bFaculty of Civil Engineering, K. N. Toosi University of Technology, 1996715433, Tehran, Iran cDepartment of Mechanical Engineering, Central Tehran Branch-Islamic Azad University, Tehran, Iran

Abstract

Long term measurements of the amount of solar energy at ground level are not easily possible in many locations. Therefore, using empirical relations and recently applying Artificial Neural Networks (ANN) are common means for prediction of the available solar energy at desired areas. Recent studies indicate that the performance of ANN provides better prediction than empirical relations. In former researches about ANN modeling of solar energy for some geographical locations, the parameters such as maximum and minimum daily temperature, relative humidity and wind speed were considered as the input of the soft computing. In present Multilayer Perceptron (MLP) ANN modeling, the amount of suspended Particulate Matters (PM10 and PM25) in the atmosphere is also added to the soft computation input. This ANN modeling strategy is used for estimating the amount of daily absorption of global solar radiation (both beam and diffuse radiation) on the land surface of Tehran (Longitude 51.23N and Latitude 35.44E) during a year. Furthermore, Indexes of Root Mean Square Error (RMSE), Absolute Fraction of Variance (R2) and Mean Absolute Percentage Error (MAPE) are used for accuracy evaluation of modeling results.

© 2015TheAuthors. PublishedbyElsevierLtd.This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of the Euro-Mediterranean Institute for Sustainable Development (EUMISD) Keywords:Global solar radiation; Artificial neural network; Particulate matter air pollution

Corresponding author. Tel.: +98-21-8876-3733; fax: +98-21-8877-9476. E-mail address: syazdi@kntu.ac.ir

1876-6102 © 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4 .0/).

Peer-review under responsibility of the Euro-Mediterranean Institute for Sustainable Development (EUMISD) doi: 10.1016/j.egypro.2015.07.764

1. Introduction

The amount of solar radiation as one of the most significant parameters in design and optimum using of different solar systems is critical [1]. Moreover, one of the most crucial parameters for installation and initiation of solar collectors and panels is to estimate and measure the amount of solar radiation in each place [2, 3].

Due to the fact that paying the cost of installation and initiation of facilities and equipment for measuring the amount of solar radiation is required and in some areas this cost cannot be afforded, using empirical relations and modeling with neural networks has always been considered since the past [4]. Some empirical relations such as relations of Angstrom [5], Hargreaves-Samani [6], Ogelman et al. [7], Akinoglu and Ecevit [8] and Louche et al. [9] are considered more significantly.

One of the limitations of empirical relations is existence of constant coefficients which are measured based on experience and considering the location. Another limitation of empirical relations is that the prediction is made based on only a few specific parameters such as the number of sunshine duration [5], maximum and minimum daily temperature [10] and cloud observations [11]. Thus, the empirical relations are not efficient and applicable if other parameters are used. Furthermore, using Artificial Neural Networks will not cause limitation for assessing and choosing effective parameters such as sunshine hours, maximum and minimum air temperature, mean relative humidity, mean wind speed, the amount of radiation in sunny and cloudy sky, air pressure etc. In the past decade, many researchers have predicted and estimated the amount of global solar radiation using different models of artificial neural network and meteorological data. Some of their research results will be discussed below.

The diffused and global amount of solar radiation has been predicted by Moustris et al. in Greece using some input parameters such as air temperature, relative humidity, sunshine duration and cloud observation with regard to Multilayer Perceptron (MLP) and it has been proven that the modeling result with neural networks is more accurate in comparison with that of empirical relations [13]. Mellit et al. have presented an adaptive model "a" with a correlation coefficient of 97%. This model has estimated and predicted the amount of beam, diffused and global solar radiation using input temperature of the air and with regard to ANN [14].

In addition, Behrang et al. have compared the performance of two types of neural networks-Multilayer Perceptron (MLP) and Radial Basis Function (RBF)-in order to predict the amount of daily global solar radiation of Dezful city in Iran. Using average daily temperature inputs, relative humidity, sunshine hours, evaporation rate and wind speed, they have come into result that multilayer network with Mean Absolute Percentage Error of 5.21% and Absolute Fraction of Variance of 99.57% produces a better result compared with Radial Basis Function (RBF) [15]. Benghanem and Mellit have predicted the daily global solar radiation with correlation coefficient of 98.8% using Radial Basis Function (RBF) neural network and also some input parameters such as air temperature, the sunshine duration and relative humidity [16].

Moreover, Using Multilayer Perceptron (MLP) neural network and considering different inputs, Asl et al. have predicted the amount of daily global solar radiation of Dezful city in Iran with Mean Absolute Percentage Error of 6.08%. [17].Ramedani et al. have assessed the amount of global solar radiation of Tehran city using Multilayer Perceptron (MLP) neural network and also three layers with neuron number of 6-37-1 and some minimum input parameters such as maximum daily temperature, relative humidity, sunshine duration and the amount of precipitation. They have reached the optimum model with root-mean-square error of 3.09 [18]. Using Multilayer Perceptron and Radial Basis Function neural networks with different input parameters, Maitha et al. have predicted monthly solar radiation. In this study, they have demonstrated that Radial Basis Function neural network with Absolute Fraction of Variance of 92% gives a better performance in comparison with Multilayer Perceptron neural network [19].

Furthermore, Waewsaket al. predicted the monthly amount of solar radiation of Bangkok city in Thailand regarding various inputs and functions of neural networks and finally presented a model with root-mean-square error of 0.0031 to 0.0035 [20]. The daily amount of solar radiation has been predicted by Assasa et al. using four-layer neural network and considering numerous parameters. Consequently, they have presented a new model with root-mean-square error of 0.1169 [21].

Regarding the previous conducted researches and made predictions by neural networks, the effect of the parameter of air particulate matter has not been considered carefully and only the importance of these factors in empirical relations has been regarded which is mentioned below.

According to the researches done by Wang et al. in different cities of China, the air pollution has a substantial effect on the amount of solar radiation [22]. Zhao et al. started predicting the amount of solar radiation regarding the index of air pollution using multifarious experimental equations. According to the result of their study, it has been proven that the air pollution has a major influence on prediction and its error [23].

Moreover, applying empirical relations and by considering the amount of solar radiation factors and air pollution index (API), Suthar et al. have presented an exponential quadratic equation with Root-Mean-Square Error of 3.08 and Mean Absolute Percentage Error of 0.1342 for India [24].

Hence, in this research, the amount of global solar radiation is predicted using neural network considering the importance of the parameter of Particulate Matter in the air.

2. Artificial Neural Network

The artificial neural networks are one of the concepts of artificial intelligence which is inspired by human brain performance in identification of phenomena. In a multilayer artificial neural network, neurons are placed in different layers. The first layer is called input layer, which receives the input information and unto its communication ability with other neurons, transfers the input signal to next layer. The communication ability of each neuron with another neuron is called neuron weight that the number of neurons in each layer is depended on the weight and the number of neurons of previous layer. In addition to input layer, neural networks are consists of hidden layers and also output layer. It is worth mentioning that the number of hidden layers and also the number of the neurons of each layer can be arbitrary. However, it should be noticed that although adding each neuron to the hidden layer will reduce the calculation error, it will cause more time consuming for calculation. Thus, reaching a logical proportion in choosing the number of neurons is obligatory [25]. In neural network, neuron is considered as the major processor. A schematic view of the neural network performance is depicted in Figure 1.

3. Statistical Analysis

In this research, several statistical indexes such as Root-Mean-Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and Absolute Fraction of Variance have been used for evaluation of accuracy level and the performance of model and network.

Activation Function

Output

Weights

Fig. 1. Schematic view of neuron diagram regarding the performance state and the value of effective parameters in modeling

Root-Mean-Square Error and Mean Absolute Percentage Error indexes are considered as appropriate indexes for determining the modeling accuracy. Closeness of these two indexes to zero leads to high modeling accuracy. The Absolute Fraction of Variance expresses the probability of correlation between two types of data in the future. Likewise, the closeness of this amount to one hundred, leads to a better modeling performance. Equations used for calculating the indexes are mentioned below.

RMSE =

- £(Gp - G

MAPE =

Gp - Ga

2 i i=1

I (Ga " Gp )

R2 = 1 -

I (Ga " Ga )2

Where:

Gp is the predicted daily global solar radiation and Ga is real daily global radiation on a horizontal surface. Gais the daily average global radiation in the measured period and n is the number of observations and measured days.

4. Case Study Region and Data

In this study, the information of Tehran Meteorological Synoptic Station in a region with longitude of 51.23N and latitude of 35.44E with a height of 1419 meters from the sea level has been used. Furthermore, the modeling is based on information about sunny days of the years 2012 and 2013 for network learning and of the years 2013 and 2014 in order to network testing. The information about minimum and maximum of daily temperature, relative humidity, wind speed and the amount of particulate matter in the air (PM10 and PM2.5) as the input and also the amount of daily global solar radiation as the output have been used and measured in this station.

5. Results

The measured parameters in sunny days in 2012 and 2013, which are used for network learning, have been regarded for 226 days. At first, data has been normalized between -1 and 1 and afterwards has been used in the network. In this research, modeling was considered using Multilayer Perceptron (MLP) neural network.

The optimum result is achieved using trial and error method and also by changing the types of functions, the number of hidden layers and the number of neurons of each layer. In the modeling process, by using Feed Forward Neural Network with Levenberg-Marquardt function, the network has been learned. The result of evaluations with different number of neurons and hidden layers and also by using tangent sigmoid transfer function and Log-sigmoid transfer function in hidden layers and by using linear transfer function (purelin) in the final layer of hidden layer is elaborated in Table 1.

Table 1. The results of measuring indexes of modeling accuracy in learning, credit rating and network testing regarding different situations

Function Neurons MAPE Train RMSE R2 MAPE Validation RMSE R2 MAPE Test RMSE R2

Tansig-Purelin 22-1 1.66 0.07 0.99 7.2 0.13 0.92 5.34 0.18 0.92

Tansig-Purelin 24-1 0.22 0.06 0.99 4.14 0.15 0.87 3.7 0.08 0.95

Tansig-Purelin 32-1 0.80 0.074 0.99 8.5 0.14 0.92 0.005 0.14 0.95

Tansig-Purelin 46-1 4.96 0.11 0.97 5.44 0.15 0.94 6.61 0.19 0.93

Logsig-Purelin 14-1 3.84 0.097 0.98 6.71 0.11 0.97 2.42 0.098 0.96

Logsig-Purelin 29-1 2.53 0.042 0.99 4.5 0.15 0.93 4.05 0.12 0.97

Logsig-Purelin 9-1 2.5 0.07 0.99 0.35 0.08 0.98 4.04 0.10 0.98

Tansig- Logsig-Purelin 12-24-1 0.95 0.04 0.99 2.4 0.065 0.98 3.13 0.077 0.97

Tansig- Logsig-Purelin 8-32-1 0.26 0.017 0.99 3.16 0.13 0.98 0.46 0.10 0.94

Tansig-Tansig-Purelin 8-24-1 0.46 0.054 0.99 4.01 0.09 0.98 1.6 0.14 0.93

Logsig- Tansig-Purelin 9-18-1 3.7 0.092 0.98 3.83 0.096 0.98 8.5 0.12 0.97

The most optimum state, which is the one with lowest error, is achieved by using two hidden layers with Tansig function and with 12 neurons in the first layer, with Logsig layer and with 24 neurons in the second layer and also by using linear transfer function in the output layer.

In the following, regarding the assessed information and by choosing the best and most optimum model, the real amount of global solar radiation has been scaled by presented model and its result has been demonstrated in Figure 2.

Day of year

Fig. 2. The amount of measured global solar radiation in comparison with the modeling result

According to Figure 3, which is the result of optimum modelling and the amount of measured global solar radiation in drawn correlation diagram, the correlation coefficient based on the results is approximately 0.992. Due to the fact that the correlation coefficient is roughly close to 1, it can be inferred that the predicted result is exact enough.

Finally, considering the statistical parameters which have introduced before, the network has been presented with Root-Mean-Square Error of 0.05 j/cm2 .day, Mean Absolute Percentage Error of 1.5 per cent and Absolute Fraction of Variance of 99 per cent. Additionally, the existent error between each of the days of the year has been shown in Figure 4.

1300 1600 1900 2200 2500 2800 3100 3400 3700 ANN(j/Cm2)

Fig. 3. The total amount of measured radiation in comparison with the modeling result

Day of year

Fig. 4. Diagram of the deviation between the measured and predicted data regarding the days of the year

6. Conclusion

Application of the neural network modeling can be used for prediction of the amount of solar radiation in regions in which measured solar radiation data is not available.

In present neural network computations, the total absorption of global solar radiation energy (summation of both beam and diffused radiations) is modeled and effects of adding the suspended particulate matters to the atmospheric parameters as input parameters are studied. For this purpose, the total solar radiation is modeled on the land surface during a year in Tehran using atmospheric parameters such as minimum and maximum daily temperature, relative humidity and wind speed, which are measured in some meteorological stations, are considered for modeling input.

Comparison of the present neural network modeling results with previous measurements shows that considering the effect of particulate matter in the atmosphere provides considerable accuracy improvements on predicted global solar radiation.

References

[1] Bakirci K. Correlations for estimation of daily global solar radiation with hours of bright sunshine in Turkey, J Energy, 2009;34: 485-501.

[2] Bany J, Appelbaum J. Effect of shading on the design of a field of solar collectors, J Solar Cells, 1987;20: 201-228.

[3] Bany J, Appelbaum J. Shadow effect of adjacent solar collectors in large scale systems, J Solar Energy, 1979;23:497-507.

[4] Rehman S, Mohandes S. Artificial neural network estimation of global solar radiation using air temperature and relative humidity, J Energy

Policy, 2008;36:571-576.

[5] Angstrom A. Solar and terrestrial radiation, Q. J. R. Meteorol.Soc,1924;50:121- 126

[6] Hargreaves GH, Samani ZA. Estimation Potential Evapotranspiration, Journal of the Irrigation and Drainage Division, 1982;108. 3: 225-230.

[7] Ogelman H, Ecevit A,Tasdemiroglu E. A new method for estimating solar radiation from bright sunshine data, J Solar Energy,1984;33:619-

[8] Akinoglu BG, Ecevit. A further comparison and discussion of sunshine based models to estimate global solar radiation, J Energy, 1990;

15:865-872.

[9] Louche A, Notton G, Poggi P, Simonnot G. Correlations for direct normal and global horizontal irradiation on a French Mediterranean site, J

Solar Energy, 1991;46: 261-266.

[10] Bristow KL, Campbell GS. On the relationship between incoming solar radiation and daily maximum and minimum temperature, J Agricultural and Forest Meteorology, 1984;31:159-166.

[11] Ehnberg JSG, Bolle MHJ. Simulation of global solar radiation based on cloud observations, J Solar Energy, 2005; 78: 157-162.

[12] Yadav AK, Chandel SS. Solar radiation prediction using Artificial Neural Network techniques: A review, J Renewable and Sustainable

Energy Reviews, 2014;33:772-781.

[13] Moustris K, Paliatsos AG, Bloutsos A, Nikolaidis K, Koronaki I, Kavadias K. Use of neural networks for the creation of hourly global and diffuse solar irradiance data at representative locations in Greece, J Renewable Energy, 2008;33: 928-932.

[14] Mellit A, Eleuch H, Benghanem M, Elaoun C, Pavan AM. An adaptive model for predicting of global, direct and diffuse hourly solar irradiance, J Energy Conversion and Management, 2010;51:771-782.

[15] Behrang MA, Assareh E, Ghanbarzadeh A, Noghrehabadi AR. The potential of different artificial neural network (ANN) techniques in daily global solar radiation modelling based on meteorological data, J Solar Energy, 2010;84:1468-1480.

[16] Benghanem M, Mellit A. Radial basis function network-based prediction of global solar radiation data: application for sizing of a standalone photovoltaic system at Al-Madinah, Saudi Arabia, J Energy, 2010;35:3751-3762.

[17] Asl SFZ, Karami A, Ashari G, Assareh A, Hedayat N. Daily global solar radiation modelling using multi-layer perceptron (MLP) neural networks, World Academy of Science, 2010;5.

[18] Ramedani Z, Omid M, Keyhani A. Modelling Solar Energy Potential in a Tehran Province Using Artificial Neural Networks, International Journal of Green Energy, 2013;10:427-441.

[19] Al-Shamisi MH, Assi AH, Hejase HAN. Artificial neural networks for predicting global solar radiation in Al Ain city - UAE, International Journal of Green Energy, 2013;10:443-456.

[20] Waewsak J, Chancham C, Mani M, Gagnon Y. Estimation of monthly mean daily global solar radiation over Bangkok, Thailand using artificial neural networks, Energy Procedia, 2014;57:1160-1168.

[21] Assasa O, Bouzgou H, Fetah S, Salmi M, Boursas A. Use of the artificial neural network and meteorological data for predicting daily global solar radiation in Djelfa, Algeria, International Conference on Composite Materials & Renewable Energy Applications (ICCMREA), Sousse, Tunisia, 2014.

[22] Wang Y, Yang Y, Zhao N, Liu C, Wang Q. The magnitude of the effect of air pollution on sunshine hours in China, J Geophysical Research, 2012;117:D00V14.

[23] Zhao N, Zeng X, Han SH. Solar radiation estimation using sunshine hour and air pollution index in China, J Energy Conversion and Management, 2013;76:846-851.

[24] Suthar M, Singh GK, Saini RP. Effects of air pollution for estimating global solar radiation in India, International Journal of Sustainable Energy, 2014.

[25] Haykin S. Neural networks, A Comprehensive foundation. 2nd Ed. pp. 9.32-9.43, New York: McMillan; 1999.