Sie sind auf Seite 1von 4

ARTICLE IN PRESS

Renewable Energy 33 (2008) 23292332 www.elsevier.com/locate/renene

Technical Note

Predicting total solar irradiation values using articial neural networks


J. Mubiru
Department of Physics, Makerere University, P.O. Box 7062, Kampala, Uganda Received 7 August 2007; accepted 5 January 2008 Available online 21 February 2008

Abstract This study explores the possibility of developing an articial neural networks model that could be used to predict monthly average daily total solar irradiation on a horizontal surface for locations in Uganda based on geographical and meteorological data: latitude, longitude, altitude, sunshine duration, relative humidity and maximum temperature. Results have shown good agreement between the predicted and measured values of total solar irradiation. A correlation coefcient of 0.997 was obtained with mean bias error of 0.018 MJ/m2 and root mean square error of 0.131 MJ/m2. Overall, the articial neural networks model predicted with an accuracy of 0.1% of the mean absolute percentage error. r 2008 Elsevier Ltd. All rights reserved.
Keywords: Articial neural networks; Solar radiation; Prediction; Model

1. Introduction There are quite a number of models that require solar radiation data as input, such as ecological and crop models [1,2]. Williamson and Erell [3] has modeled a buildings thermal performance using solar radiation resource data. Solar radiation data can be provided through measurements but it is difcult to have measurements from all locations where these models can be applied; measuring instruments are expensive to purchase, install and maintain. An alternative to obtaining solar radiation data is estimating it using an appropriate solar radiation model. Several authors, such as in [46], have used empirical formulations to predict total solar radiation. Angstrom [7] pioneered the use of the empirical model that exploits empirical relationships between solar radiation and existing climatic parameters. Despite its wide use, the empirical model has not met the desirable accuracy required from a reliable prediction model. It does not wholly capture nonlinear characteristics exhibited by solar radiation. It is based on this background that several authors, such as in [813], have explored the use of articial neural networks (ANN) to predict total solar radiation.
Tel.: +256 772413517; fax: +256 41321401.

ANN models employ articial intelligence techniques and are data driven; they learn and memorize a data structure and subsequently simulate the structure. ANN are able to learn key information patterns within a multidimensional information domain [14]. In a way, ANN mimic the learning process of a human brain and therefore do not need characteristic information about the system; instead, they learn the relationship between input parameters and the output variables by studying previously recorded data. This makes ANN ideal for modeling non-linear, dynamic, noisy data and complex systems. Further, ANN are good for tasks involving incomplete data sets [15]. Fig. 1 shows a typical neural network, which consists of an input, a hidden and an output layer. An input xj is transmitted through a connection, which multiplies its strength by a weight wij to give a product xjwij. This product is an argument to a transfer function f, which yields an output yi represented as yi f
n X j1

! xj wij , (1)

E-mail address: jwm_mubiru@yahoo.com 0960-1481/$ - see front matter r 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.renene.2008.01.009

where i is an index of neurons in the hidden layer and j is an index of an input to the neural network.

ARTICLE IN PRESS
2330 J. Mubiru / Renewable Energy 33 (2008) 23292332

Neuron wij xj f yi

Table 1 Four study sites with their location parameters Station Mbarara Lira Tororo Kampala Latitude 0.62 2.28 0.68 0.32 Longitude (deg East) 30.65 32.93 34.17 32.58 Altitude (m) 1413 1189 1170 1220

Fig. 1. Typical neuron in a feedforward neural network.

Tymvios et al. [16] has trained seven ANN models using daily values of measured sunshine duration, theoretical sunshine duration, maximum temperature and the month number. The period of data collection was 19861992 at a location at Athalassa, Cyprus, situated at latitude 351080 N, longitude 331230 E and altitude 161 m. Back-propagation method was used with tangent sigmoid as the transfer function. Two hidden layers with neurons varying between 23 and 46 were investigated. The best performing ANN model was one with all inputs except the month number and its result showed a mean absolute percentage error (MAPE) of 0.12%. Sozen et al. [17] created two datasets, using measured data from 17 stations in Turkey collected between 2000 and 2002. One set with data for 11 stations was used for training a neural network and the other dataset from six stations was used for testing. The back-propagation method was used with one and two hidden layers. The other settings were conjugate gradient and Levenberg Marquardt as the training algorithms, sigmoid as the transfer function and the number of neurons varied between four and nine. The inputs to the network were: latitude, longitude, altitude, month, average sunshine duration and average temperature. Results showed a MAPE of less than 6.7% in the prediction of the solar radiation. The present study explores the possibility of developing an appropriate model using the ANN method that could be used to predict total solar irradiation on a horizontal surface, based on geographical and meteorological data: latitude, longitude, altitude, sunshine duration, relative humidity, and maximum temperature. Four stations were selected for the study and the data collected from these stations was split into two sets. The rst dataset includes data from three stations, which is used for building the ANN model and the other dataset has data from one station, used for validating the ANN model. 2. Test area and data Table 1 shows four locations that have been selected and used for the study. Kampala and Mbarara stations exhibit wetter and more humid conditions than the other two stations. Lira and Tororo stations register longer sunshine durations. The wetter stations are located at higher altitudes than the other study stations. However, all the four stations are generally characterized by scattered bushes and short grass. Total solar irradiation data was measured on an hourly basis and a daily value recorded at the end of the day. The

Positive north latitude.

measurements were carried out from April 2003 to December 2005, on a horizontal surface, using Kipp and Zonen CM6B pyranometers installed at the four locations. The measured total solar irradiation data was validated using a service at the World Wide Web: http://www.helioclim. net [18]. Sunshine duration (hours) was measured using Kipp and Zonen CSD 1 sunshine duration sensors and the data covers the same period as the total irradiation data. The relative humidity (percentage) and maximum temperature (degrees celsius) data was obtained from the Uganda Meteorological Department and covers the period 19932005. Monthly average daily values of these parameters were computed and used in the present study.

3. Experimental procedure The data from the four sites were split into two, such that the dataset from three stations, that is, Mbarara, Lira and Tororo, was used for training a neural network. The dataset from the Kampala station was reserved for validating the ANN model. A feedforward back-propagation neural network was used in the present study, with six input variables, which included: latitude Lat, longitude Lon, altitude Alt, sunshine hours S, relative humidity R and maximum temperature T max . Three transfer functions were investigated, including the tangent sigmoid, log sigmoid and linear functions. Further, one-, two- and three-hidden layer architectures were tested, in which the number of neurons was varied. LevenbergMarquardt was used as the training algorithm. According to Lourakis [19], the Levenberg Marquardt training algorithm has become a standard technique for non-linear least-squares problems, widely adopted in a broad spectrum of disciplines. Predicted values of total solar irradiation were compared with measured values through correlation and error analysis. The latter was carried out through computation of mean bias error (MBE), root mean square error (RMSE) and mean absolute percentage error (MAPE), represented, respectively, as !, N X MBE y i xi N, (2)
i1

ARTICLE IN PRESS
J. Mubiru / Renewable Energy 33 (2008) 23292332 2331

v !, u N u X 2 t RMSE y i xi N,
i1

(3)

variation can be attributed, partly, to the differences in composition of the inputs into the ANN models. 5. Conclusions

MAPE

 , !! X N N X    y xi  xi 100,   i1 i  i1

(4)

where yi is a predicted value, xi is a measured value and N is equal to the number of observations.

4. Results and discussion A linear transfer function was xed at the output layer, while the tangent sigmoid and log sigmoid functions were tested in the hidden layer. Results from either sigmoid transfer functions in the hidden layer did not show signicant differences. Nevertheless, the tangent sigmoid transfer function was chosen because of its wider output range. Further, there was no signicant difference in the use of one, two and three-hidden layer architectures. Onehidden layer was used in order to minimize the complexity of the proposed ANN model. After several trials, 15 neurons in the hidden layer were found to be appropriate for the training process. Fig. 2 shows the nal neural network architecture, where H is the total solar irradiation. Predictions of total solar irradiation obtained for the validation site (Kampala) were correlated with the measured values, giving a correlation coefcient of 0.997. The corresponding MBE was 0.018 MJ/m2 and the RMSE was 0.131 MJ/m2. Further, the proposed ANN model, which accepts six geographical and meteorological input variables, predicts with a MAPE of 0.1%. This percentage error is lower for the proposed ANN model than for the models corresponding to the two studies by [16,17]. This
Lat

The present study was prompted by the lack of ground measurements of total solar irradiation in some locations of interest. Modelers and designers of solar energy systems may need resource data on total solar irradiation to perform their respective tasks. Predicted data can be used in the absence of measurements. This study has attempted to develop a model that can be used to predict monthly average daily total solar irradiation using the articial neural networks (ANN) method. The proposed ANN model, which accepts six geographical and meteorological input variables, predicts with an accuracy of 0.1% for the MAPE. This is within an acceptable range of accuracy. Acknowledgements The authors are grateful to The Norwegian Council for Higher Educations Programme for Development Research and Education (NUFU) for providing the measuring equipment and software used in data analysis. Special thanks go to the Uganda National Meteorological Department for the provision of meteorological data. References
[1] Grant RH, Hollinger SE, Hubbard KG, Hoogenboom G, Vanderlip RL. Ability to predict daily solar radiation values from interpolated climate records for use in crop simulation models. Agric For Meteorol 2004;127(12):6575. [2] Suzaki T, Kume A, Ino Y. Evaluation of direct and diffuse radiation densities under forest canopies and validation of the light diffusion effect. J For Res 2003;8:28390. [3] Williamson TJ, Erell E. The thermal performance simulation and the urban microclimate: measurements and prediction. In: Proceedings of the seventh international IBPSA conference, Rio de Janeiro, Brazil, August 1315, 2001. [4] Mosalam Shalout MA, Hassan AH, Fathy AM. Study of the solar radiation over Menia. Renew Energy 2001;23:62139. [5] Paulescu M, Schlett Z. Performance assessment of global solar irradiation models under Romanian climate. Renew Energy 2004;29:76777. [6] Mubiru J, Banda EJKB, DUjanga F, Senyonga T. Assessing the performance of global solar radiation empirical formulations in Kampala, Uganda. Theor Appl Climatol 2007;87(14):17984. [7] Angstrom A. Solar and terrestrial radiation. Q J R Meteorol Soc 1924;50:1215. [8] Elminir HK, Areed FF, Elsayed TS. Estimation of solar radiation components incident on Helwan site using neural networks. Sol Energy 2005;79:2709. [9] Mellit A, Benghanem M, Arab AH, Guessoum A. A simplied model for generating sequences of global solar radiation data for isolated sites: using articial neural network and library of Markov transition matrices approach. Sol Energy 2005;79:46982. [10] Mihalakakou G, Santamouris M, Asimakopoulos DN. The total solar radiation time series simulation in Athens using neural networks. Theor Appl Climatol 2000;66:18597.

Lon

Alt H S

Tmax

Input layer

Hidden layer

Output layer

Fig. 2. Final neural network architecture.

ARTICLE IN PRESS
2332 J. Mubiru / Renewable Energy 33 (2008) 23292332 [16] Tymvios FS, Jacovides CP, Michaelides SC, Scouteli C. Comparative study of Angstroms and articial networks methodologies in estimating global solar radiation. Sol Energy 2005;78:75262. [17] Sozen A, Arcaklioglu E, Ozalp M. Estimation of solar potential in Turkey by articial neural networks using meteorological and geographical data. Energy Convers Manage 2004;45:303352. [18] Geiger M, Diabate L, Menard L, Wald L. A web service for controlling the quality of measurements of global solar irradiation. Sol Energy 2002;73(6):47580. [19] Lourakis MIA. A brief description of the LevenbergMarquardt algorithm implemented by Levmar. /http://www.ics.forth.gr/lourakis/ levmar/levmar.pdfS; accessed on 4th May 2006. [11] Mohandes M, Rehman S, Halawani TO. Estimation of global solar radiation using articial neural networks. Renew Energy 1998;14(14): 17984. [12] Reddy KS, Ranjan M. Solar resource estimation using articial neural networks and comparison with other correlation models. Energy Convers Manage 2003;44:251930. [13] Alawi S, Hinai H. An ANN based approach for predicting global solar radiation in locations with no direct measurement instrumentation. Renew Energy 1998;14:199204. [14] Kalogirou SA. Articial neural networks in renewable energy systems applications: a review. Renew Sustain Energy Rev 2001;5:373401. [15] Kalogirou SA. Applications of articial neural-networks for energy systems. Appl Energy 2000;67:1735.

Das könnte Ihnen auch gefallen