Research Journal of Recent Sciences _________________________________________________ ISSN 2277-2502 Vol. 4(6), 7-12, June (2015) Res.J.Recent Sci. International Science Congress Association 7 Predicting Power Consumption using Algorithm of artificial Neural networks; Case Study: Golestan Province Seyed Ahmad Sheibatalhamdi and Fatemeh Keikha Islamic Azad University of Firouzkouh branch, Tehran, IRAN Available online at: www.isca.in , www.isca.me Received 20th November 2013, revised 15th February 2014, accepted 4th May 2014Abstract Today, using smart technologies for solving complex scientific problems in different industrial sectors have been significantly considered. The systems could achieve general facts through conducting calculations on empirical data. Hence, the systems can be called intelligent systems. Neural networks are kinds of these intelligent systems, which can transfer hidden knowledge beyond the data to structure of network through processing empirical data. The main objective of the present study is predicting power consumption using algorithm of artificial neural networks, which would be done as case study in Golestan province Iran. Generally, in order to predict future events, historical events and data would be considered. For this purpose, previous data would be processed, so that a generalized pattern for future can be achieved. In most methods of prediction, one can assume that relations among variables would be continued even in future. Information and data about demand for power consumption in each period is an essential issue in order to have exact planning and proper policy making. Hence, prediction of power consumption would be significant for different economic units. Obtained results from the study indicate that feed-forward neural network model has high validity level, comparing to routine prediction models for power consumption. Keywords: Neural network, artificial neural network, prediction, power consumption, GDP. IntroductionApplication of neural networks in economic issues has been started since late 1990s through study of White (1998) in financial markets and prediction of stock prices of IBM Company. Success of neural networks in relevant studies of financial domains has gained attention of many specialists in microeconomics and econometrics. There are also several studies in regard with using artificial neural networks in order to predict demand for power through applying mentioned models. Salama and George have considered prediction of short-term demand for power in Republic of Check. Through comparing artificial neural network model and ARIMA model, they have suggested that neural networks cam provide better prediction than ARIMA model because of non-linear nature of demand for power. Yaw et al have evaluated neural networks and wavelet transmission in short-term prediction of demand for power. They have suggested that wavelet transmission can be applied as a useful device for prediction. Steinrez et al have investigated application of artificial neural networks in prediction of demand for power. The mentioned study has been in fact an overview and general evaluation for all conducted studies in regard with prediction of demand for power from 1991 to 1999. The study has referred also problems, weakness and power points of artificial neural networks against other models. Conducted study by Jang and Dong has applied wavelet transmission model for short-term prediction in Australia. The study has indicated that it can present better results and outcomes than other relevant models. Kim et al have applied neural networks and wavelet transmission for short-term prediction of demand for power in South Korea. Obtained results from the mentioned study indicate that mentioned models are effective in short-term prediction. Hipret et al has applied neural network for prediction of daily power consumption in Brazil10. The mentioned study has been evaluated better than other relevant models through applying different 24-hour models. Gilermo et al has conducted a study in order to evaluate power consumption prediction method of new neural network based on structural application11. The study has indicated that because of high costs of energy, it would be essential to find some methods and solutions in order to optimizing new energy resources and enabling consumers to have a good perception about pressure curve. The perception would help improvement of consumer flexibility and ability to respond prices or other signs of power selling12. One of the most important stages of conducting current calculations would be predicted consumption curve. Through having an exact curve, consumers would be able to attend responding programs and confirming of relevant activities13. Kong Gi Li et al has conducted a comparative study for prediction of power consumption in buildings using neural network and hybrid fuzzy-neural system14. Obtained results Research Journal of Recent Sciences _____________________________________________________________ ISSN 2277-2502Vol. 4(6), 7-12, June (2015) Res.J.Recent Sci. International Science Congress Association 8 from the study indicate that Genetic Algorithm Model (GA) is able to predict energy consumption in buildings based on method of regular data on artificial neural networks15. In this model, regressive classification radiuses, which apply certain rules, have been optimized. Organizational hierarchy structure of ANFIS would regulate effective assumptions and parameters in order to optimize implementation of prediction. Liuki et al have conducted a study in order to investigate power energy consumption model based on neural network of GA-BP16. They have indicated that GA-BP method has been presented based on neural network in order to predict quantitative changes of investment in power industry. Basis of the method has been Sliding Window, which has applied power industry structure in order to form a part of successive constant linear capital structure. Ping Jang and Hui Wang has conducted a study in international conference of 2012 on future electric energy and presented energy system, in order to investigate fuzzy violet neural networks for prediction of urban power consumption17. Obtained results from the mentioned study have indicated that just in fuzzy violet neural network method, prediction would be based on non-linear stages. In non-linear method, focusing on local strong stimulants and slow process of training would have the most effects on this method. Fuzzy violet neural network method would be presented for cities with high level of energy consumption18. Methodology The present study has applied feed forward neural networks for prediction and post-distribution algorithm has been also applied for training networks. In order to investigate effects of social and economic criteria on demand for power energy consumption in Gorgan Iran, related data to GDP, population, and price of KWH from 2001 to 2011 have been considered19. Then, using artificial neural networks and considering social and economic criteria, power consumption of the city has been predicted between 2012 and 2021. Inputs have included neural network, GDP, population, and price per KWH; while output of neural network has been demand for power consumption in Gorgan province. For input data of the model from 2012 to 2021, neural network has been applied in order to predict the years. Finally, the data has been applied as main data to train the model20. Data collection method: Related data to GDP, population, and power consumption have been collected from Department of Energy. Results and Discussion Structure of artificial neural networks: Artificial neural networks have been created through inspiration from neural system and its components. The models have less capabilities and expansion than natural neural networks. However, it should be noted that, the network has calculative ability in conducting some activities such as approximation of a non-linear function21. Artificial equivalents for neural networks are some units, which a schematic of them has been presented in figure-1.In artificial neural networks, synapses would be equaled with a unit weight, so that each input can be affected before entering body of processor unit. At the next stage, weighted inputs would be combined and finally, output amount of a neuron would be determined through a stimulant function based on its inputs. The first applied function in an artificial neuron has been stepwise function. As it was indicated in fig.1, while using the function and after combining weighted inputs, the values would be compared with a threshold value. If output of neuron is smaller than its input, would be equal to 1; otherwise, it would be equal to 0. The term “network” would be applied for any system including artificial neurons. The network can be constructed from a neuron or a set of linked neurons. Table-1 Input and output data of the model Year Input1 Input 2 Input 3 output Population (per 1000) Price per KWh (Rials) GD(billion Rials) Power consumption (MWh) 2001 372,081 10.5 330,565 9433,160 2002 379,462 10.8 355,554 5,901,187 2003 387,053 11.2 379,838 6,302,271 2004 394,844 11.4 420,928 4,752,357 2005 402,845 11.4 446,880 7,469,357 2006 411,066 12.6 499,071 5,452,783 2007 419,514 20.2 501,000 2,574,367 2008 420,059 31.6 501,892 5,692,309 2009 428,189 37.6 502,897 4,712,835 2010 437,099 43.2 503,259 3,112,334 2011 441,075 43.2 504,009 6,455,718 Research Journal of Recent Sciences ______ ______________________________ Vol. 4(6), 7-12, June (2015) International Science Congress Association Schematic of a processor unit (artificial neuron) In biological neurons, stability of synapses may be changed under special conditions in order to regulate and adjust behavior of a neuron against its input stimulus. In artificial neurons, equivalent of the process would be conducted through changing amoun t of weights in learning process of network. In data processing, process of weight regulation and desirable weight achievement would be known as network learning. Saved weights are in fact knowledge of the network. Process of weight changes would be contin ued until achieving sufficient in network’s response. Every specific instruction and method for weight regulation would create a learning algorithm. The process, which includes in general presentation style of teaching patterns, criteria for end of training process, and learning methods, would form training algorithm of a network22 . In general, artificial neural networks include three main components as follows23: Topology: structural properties of a neural network method of saving information in the network method of recovering saved data in the network At the present study, relevant criteria of GDP, population, and price per KWH have been considered as input data. A significant issue in this regard can be normalizing Normalization of data in a range between 0 and 10 would be essential, so that greater data would not have higher value than smaller data as well as avoiding early satisfaction of hidden neurons, which can prevent learning of neural network. T no standardized method to normalize input and output data. A method for normalization would be as follows: Where; x is normalized value of z; zmax and z and minimum values of z. Hence, before predicting input data, first normalized and then other stages should be implemented. Prediction of each input criterion using neural networks order to predict demand for energy using statistic methods such as regression model, auto regression, and moving average, a first functional relation should be determined between dependent variable (energy consumption) and independent ______________________________ __________ _______________ International Science Congress Association Figure-1 Schematic of a processor unit (artificial neuron) In biological neurons, stability of synapses may be changed under special conditions in order to regulate and adjust behavior of a neuron against its input stimulus. In artificial neurons, equivalent of the process would be conducted through changing t of weights in learning process of network. In data processing, process of weight regulation and desirable weight achievement would be known as network learning. Saved weights are in fact knowledge of the network. Process of weight ued until achieving sufficient accuracy Every specific instruction and method for weight regulation would create a learning algorithm. The process, which includes in general presentation style of teaching patterns, criteria for end of training process, and learning methods, would form training . In general, artificial neural networks Topology: structural properties of a neural network . Learning: information in the network . Anamnesis: method of recovering saved data in the network . At the present study, relevant criteria of GDP, population, and price per KWH have been considered as input data. A significant issue in this regard can be normalizing applied data. Normalization of data in a range between 0 and 10 would be essential, so that greater data would not have higher value than smaller data as well as avoiding early satisfaction of hidden neurons, which can prevent learning of neural network. T here is no standardized method to normalize input and output data. A method for normalization would be as follows: (1) and z min are maximum Hence, before predicting input data, first data should be normalized and then other stages should be implemented. Prediction of each input criterion using neural networks : In order to predict demand for energy using statistic methods such as regression model, auto regression, and moving average, a t the first functional relation should be determined between dependent variable (energy consumption) and independent variables (population of state, GDP, and number of automobiles)24 . Most of the time and for easiness, second degree or logarithmic linear relations would be assumed and the simplification may lead to incorrect results functional relation between energy consumption and effective factors on it can be a complicated issue and would not be possible easily. Using intelligent systems networks, which has gained attention of many researchers recently, seems logical. In this section, each input criterion has been predicted for years from 2011 to 2025. Table- 2 Statistic of normalized population Normalized values of population Population (per 1000) 0. 000000000000000000 0. 085454526417188400 0. 131348815781207500 0. 247661926999350800 0. 364414801164324800 0. 481628379368835500 0. 599323602705589200 0. 617479530081880000 0. 736138043683119000 0. 863696521684501000 0. 992343936506607000 Table- 3 Statistic of normalized price per KWh Normalized values of population 0. 000000000000000000 0. 185454526412684500 0. 231348815597864100 0. 447545568542560800 0. 466598421598625800 0. 588955542855569500 0. 698974656987452200 0. 718954128552521000 0. 834569875632854500 0. 993895452785265000 0. 992985645285887000 _______________ ISSN 2277-2502 Res.J.Recent Sci. 9 variables (population of state, GDP, and number of . Most of the time and for easiness, second degree relations would be assumed and the simplification may lead to incorrect results 25. Determining functional relation between energy consumption and effective factors on it can be a complicated issue and would not be possible easily. Using intelligent systems such as neural networks, which has gained attention of many researchers recently, seems logical. In this section, each input criterion has been predicted for years from 2011 to 2025. 2 Statistic of normalized population Population (per 1000) Year 372,081 2001 379,462 2002 387,053 2003 394,844 2004 402,845 2005 411,066 2006 419,514 2007 420,059 2008 428,189 2009 437,099 2010 441,075 2011 3 Statistic of normalized price per KWh Normalized values of Price per KWh (Rials) Year 10.5 2001 10.8 2002 11.2 2003 11.4 2004 11.4 2005 12.6 2006 20.2 2007 31.6 2008 37.6 2009 43.2 2010 43.2 2011 Research Journal of Recent Sciences _____________________________________________________________ ISSN 2277-2502Vol. 4(6), 7-12, June (2015) Res.J.Recent Sci. International Science Congress Association 10 Table-4 Statistic of normalized GDP Normalized values of population GDP ( billion Rials) Year 0. 000000000000000000 330,565 2001 0. 096422365426852500 355,554 2002 0. 138965423545654100 379,838 2003 0. 235648985458525800 420,928 2004 0. 356987456955233200 446,880 2005 0. 469875645965456500 499,071 2006 0. 589564556525222200 501,000 2007 0. 623698756456581000 501,892 2008 0. 712369854855865600 502,897 2009 0. 889654896542655500 503,259 2010 0. 975695547565255500 504,009 2011 A network with 3 hidden layers and mentioned inputs has been trained and its results have been presented in table5. Table-5 Predicted outputs Year Predicting normalized data of power consumption Data prediction in real scale of power consumption (MWh) 2012 1. 1459629317785100 4,856,987 2013 1. 2325533044703100 5,012,236 2014 1. 3175237575406000 5,232,253 2015 1. 3992369769881800 5,356,275 2016 1. 4761757579092700 5,402,221 2017 1. 5471201702524100 5,457,894 2018 1. 6112490069019100 5,501,435 2019 1. 6681634750786800 5,522,569 2020 1. 7178487175182100 5,856,254 2021 1. 7605971159156800 6,056,469 Conclusion Despite the fact that Islamic Republic of Iran includes much potential for developing energy, development of power energy industry has slow process. According to available literature, the main barrier in regard with development of the power industry would be lack of efficient policies and imposing international sanctions. In fact, lack of a clear framework of effective factors in power consumption in Iran and its components has forced specialists to make decisions and policies based on unprofessional ideas. The present study, which has been designed based on a proper approach for relations among determination indices of power consumption, has tried to use systematic knowledge and identify effective factors in power consumption. For this purpose, the study has presented a model for prediction of power consumption in Iran, in order to making plans for the future. The pattern would determine effective factors in power consumption and the most common accelerative or limiting relations for it. Through completing and validating the model, it can be significantly effective in making production and export policies of power; social significance of power energy; its relation with other departments; and awareness of amount of power consumption in daily life in order to make required policies. The present study has considered comparative and non-linear study of artificial neural network for prediction of feed-forward neural network. Obtained results from the study indicate that feed-forward neural network includes low level of prediction error. References 1.Patterson D.W, Artificial Neural Networks, Prentice Hall, Singapore, (1996) 2.Engelbretcht Andries, Computational Intelligence South Africa WILEY Ltd., 668- 683 (2007) 3.Menhaj Mohammad Bagher, KazemiAliyeh, Shakoori Ganjavi Hamed, Mehregan Mohammad-Reza, and TaghiZadeh Mohammad-Reza, prediction of demand for energy in transportation sector using neural networks: case study in Iran, Journal of Management Researches,2, 203-219 (2010) Figure-2 Schematic of general model’s performance with MSE=7.85×10-5 Research Journal of Recent Sciences _____________________________________________________________ ISSN 2277-2502Vol. 4(6), 7-12, June (2015) Res.J.Recent Sci. International Science Congress Association 11 Figure-3 Diagram of regression Figure4 Results of normalized prediction of power consumption until 2021 4.Shadnia Human, Data analysis using artificial neural networks, Engineering group of Ahoo, (2008)5.Reza Khodaie Mahmoodi, Sedigheh Sarabi Nejad and Mehdi Ershadisis, Expert Systems and Artificial Intelligence Capabilities Empower Strategic Decisions: A Case study, Res. J. Recent Sci.,3(1), 116-121, (2014)6.Menhaj Mohammad Bagher, Basis of neural networks, publications of Industrial University of Amir KabirEnserink, B and J.F.M. Koppenjan and W.A.H. Thissen, (2009), Policy Analysis Of Multi-actor Systems, 149-150 (2014) 7.Hosseini Reza, Presenting proposed structural pattern in 1,000,0002,000,0003,000,0004,000,0005,000,0006,000,0007,000,0001391139213931394139513961397139813991400consumptive MWyear diagram of predicted consumptive power until … 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 Research Journal of Recent Sciences _____________________________________________________________ ISSN 2277-2502Vol. 4(6), 7-12, June (2015) Res.J.Recent Sci. International Science Congress Association 12 order to predict amount of demand in supply chain of exporting companies of car pieces Bahman group, Azad University of Semnan Motaghi Hayedeh, Production and operation management, Pub. Avaye Patris, (2014)8.Momen Nejad Aida, Artificial intelligence and intellectual awareness, Journal of Data Processing, 4, 49-82 (2008) 9.Sharma Kalpa, Health IT in Indian Healthcare System: A New Initiative, Res. J. Recent Sci.,1(6) 83-86 (2012) 10.Pathak C., Mandalia H.C. and Rupala Y.M., Biofuels: Indian Energy Scenario, Res. J. Recent Sci., 1(4), 88-90 (2012) 11.Lawrance, forecasting in supply chain managementadvance in business and forecasting, volume 4, 3-12 (2001) 12.Abdul Basit Shaikh, Muhammad Sarim, Sheikh Kashif Raffat, Kamran Ahsan, Adnan Nadeem and Muhammad Siddiq, Artificial Neural Network:A Tool for Diagnosing Osteoporosis, Res. J. Recent Sci.,3(2), 87-91 (2014) 13.Goldberg D.E. and Kuo C.H., Genetic algorithms in pipeline optimization, Journal of Computing in Civil Engineering, 1(2), 128-141 (1987) 14.Jaderian Morteza, Moradzadeh Hossein, Madadipouya Kasra, Firoozinia Mohammad and Shamshirband Shahaboddin, A Method in Security of Wireless Sensor Network based on Optimized Artificial immune system in Multi-Agent Environments, Res. J. Recent Sci.,2(10), 99 (2013) 15.Kasrayi Esrafil, Fussy systems and optimization, commercial research, 2, 87-124 (1997) 16.Maria Anu, Introduction To Modeling and Simulation, State University Of New York at Binghamton, 7-10 (1997) 17.Kazemi Abbas, information of transportation and energy of Iran, National Oil Company of Iran, research center of basic sciences of Shahid Beheshti University, 48-68(2007) 18.Safari Mahdi, Conceptual modeling in knowledge reviewing, Journal of Data Processing,4, 73-104 (2004) 19.Gurney K, An Introduction to Neural Networks ,UCL Press, London, (1997) 20.Kohonen T, An Introduction to Neural Computing, Neural Networks,1(10), 3-16 (1988) 21.Mitchell T., Machine Learning, McGraw-Hill, New York, (1997) 22.Monsef Hamed, Introducing MATLAB software, Internet journal of Parvaz, (2006) 23.Demuth H and M Beale, Matlab Help / Neural Network Toolbox Version 4.0.1, (2006)