Wavelet and Neural Network Approach to Demand Forecasting based on Whole and Electric Sub-Control Center Area

Transcription

1 Internatonal Journal of Soft Computng and Engneerng (IJSCE) ISSN: , Volume-1, Issue-6, January 2012 Wavelet and Neural Networ Approach to Demand Forecastng based on Whole and Electrc Sub-Control Center Area Ptu Bunnoon, Kusumal Chalermyanont, Chusa Lmsaul Abstract Whole and electrc sub-control area load demand forecastng based on a wavelet transform and a neural networ method that are very sgnfcant technque for a load predcton. The research used wavelet transform method n preprocessng stage; furthermore, a neural networ s used to predct n forecastng stage for whole and sub-control areas predcton. The comparson results show that sub-control area forecastng has a good predcton than that the whole area forecastng based on two levels of wavelet transform. An accuracy of forecast s an essental actvty for fuel reserve plannng n a power system. Index Terms Whole area, electrc sub-control area, wavelet transform, neural networ, forecastng. I. INTRODUCTION Load forecastng, can be classfed nto four dfferental types n the power system plannng: very short term, a mnute to an hour; short term, an hour to three months; md term, three months to three years; long term, three years to twenty fve years. Several methods have been developed for forecast that s statstcal and artfcal ntellgent methods. In conventonal, many papers have used statstcal methods such as regresson, multple lnear regressons (MLR), and autoregressve ntegrated movng average (ARIMA). These methods focus on estmatng the coeffcents of varables that are lnear. There are prospected functonal forms descrbng quanttatve relatonshp between load demands and nfluencng factors by usng the hstorcal data. Recently, wth the emergence of artfcal ntellgent (AI) technologes, the artfcal neural networ (ANN), the fuzzy logc (FZ), and the genetc algorthm (GA) have been wdely mplemented to mprove the accuracy of load forecastng, hstorcal data n tme seres are both non-lnear and lnear. In the research, proposes the md term load forecastng, plays an mportant role n the relablty of a power system. It s vtal for plannng of an adequate fuel reserves to generate the electrcty to the consumpton demand for the Manuscrpt receved November 16, Ptu Bunnoon, Electrcal engneerng department, Engneerng faculty, Prnce of Songla Unversty, Hadya, Songhla, Thaland, Moble No , (e-mal: add2002@hotmal.com). Kusumal Chalermyanont, Electrcal engneerng department, Engneerng faculty, Prnce of Songla Unversty, Hadya, Songhla, Thaland, (e-mal: usumal.c@psu.ac.th). Chusa Lmsaul, Electrcal engneerng department, Engneerng faculty, Prnce of Songla Unversty, Hadya, Songhla, Thaland, (e-mal: chusa.l@psu.ac.th). future. Ths forecast s complcated an effect on load demand by dependng two factors: the complexty of an economy and a weather factor of the whole area and each regon of the country. Several researchers preferred dfferental applcatons. The paper ref. [1] shows the hybrd model based on wavelet support vector machne and modfed genetc algorthm penalzng Gaussan noses for power load forecasts, furthermore a short-term load forecastng by usng smlar day-based wavelet neural networ [2]. In ref. [3], the ntellgent hybrd wavelet models for predcton. In 2009, the combnaton of wavelet transform and neuro-evolutonary algorthm approach to demand forecastng [4]. In 2008, the research shows an adaptve wavelet neural networ-based energy prce forecastng n electrcty marets [5] and n the year 2006 the wavelet based nonlnear mult-scale decomposton model [6]. In ref. [7], n the same year, the researcher proposes technques of applyng wavelet transform nto combned model for demand forecastng n electrcty. In research ref. [8], presents an adaptve neural wavelet model and n ref. [9] proposes a hybrd wavelet- Kalman flter method for load forecastng. Lastly, the wavelets transform and neural networs for short-term electrc load forecastng are proposed [10]. All of the researches above were proposed load and prce forecastng by usng wavelet transform and neural networ algorthm but dd not present n sub-control center area forecastng. In ths research, proposes the whole area and electrc sub-control center area forecastng, are mplemented by usng wavelet transform n preprocessng stage of all areas and neural networ for forecastng n the last process. In preprocessng stage, wavelet transform s used to decompose the orgnal sgnal of demand nto one to four levels and after that wll tae t to fnd the relatonshp between factors and demand before choosng the sutable factors for feature nput for neural networ to predcton. Fnally, ths paper presents the comparson between the whole and sub-control center area forecastng based on mean absolute percentage error (MAPE). Ths artcle proposes the fve major sectons. The second secton presents an electrcty demand n a whole and sub-control area of Thaland. The thrd secton offers an mplementaton of the research stages. The fourth secton shows the results and comparson of the research. Fnally, concluson s drawn n the ffth secton. 81

2 Wavelet and Neural Networ Approach to Demand Forecastng based on Whole and Electrc Sub-Control Center Area II. ELECTRICITY DEMAND IN WHOLE AND SUB-CONTROL CENTER AREA areas s llustrated n Fg.2 (a)-(d). A. Electrcty demand n whole area The energy consumpton load demand of the Electrcty Generatng Authorty of Thaland (EGAT) n a whole area llustrates n Fg.2 (e). It can be seen that load characters from January to December and from 1997 to 2006 are ale; load shapes for each months of a year are qute dfferent. The graph shows the load demand (Wh), as llustrates n Fg.2 (e) that s a behavor of an electrc trend component ncreasng the demand every year. The demand nterval year 1997 to 2006, case study of Thaland are used n the research. The trend n 1997 to 1999 was qute stable because the world economc problems had occurred and affected to an ndustral sectors of the country. In 2000 to 2006, the trend of electrcty demand grows at rate about 5 percent per year. The energy load demand on May s 8,749,926,336 Wh n 2000, 11,171,134,811 Wh n 2004, and 12,444,389,521 Wh n 2006 beng the maxmum demand n each year. Fg.2 Electrc energy consumpton of a) Central b) Northern c) Northeast d) Southern e) whole area of the country. Fg.1 Map of each electrc sub-control center area n Thaland. B. Electrcty demand n sub-control center area Sub-control center area locaton, as llustrates n Fg.1, whch shows a map of each sub-control center area through the country. Fg.2 (a)-(d) shows energy consumpton demand n each sub-control center area of the country. Usually, the large energy consumpton wll be occurred n an ndustral area such as central area, Bango and metropoltan can be seen n Fg.2 (a). Normally, the pea of energy consumpton demand occurred between March to May n each year. In Fg. 1 llustrates the map of electrc sub-control center located area of the country. The Northern, Northeast, Southern, and Central sub-control center area located at Ptsanulo, Khonan, Krab, and Nontabur provnces respectvely. The energy consumpton load demand n these C. Basc theory C.1 Wavelet decomposton and reconstructon Electrcty load demand s generally complex and conssts of multple frequences. Furthermore, features of the demand cannot be fully captured by a sngle neural networ. Consequently, wavelet decomposton method n ths research s developed and combned wth the neural networ for ncreasng the relable forecast. The research s done by decomposng an orgnal demand nto one level, two levels, three levels, and four levels of wavelet. The feature of neural networ s selected based on the correlaton between affectng factors and demand. Wavelet theory s an applcable to several subjects. It s a powerful mplement whch can be used for a wde range of applcatons, specfcally; sgnal processng, data compresson, mage de-nosng, speech recognton, computer graphcs, and many areas of physcs and engneerng. All of the wavelet transform may be consdered forms of tme-frequency representaton for contnuous tme (analog) sgnals and so are related to harmonc nvestgaton. Almost all practcally useful dscrete wavelet transforms use dscrete-tme flter bans. These flter bans are called the wavelet and scalng coeffcents n wavelet nomenclature. Ths secton provdes a bref summary of wavelet transform method whch can be dvded nto two categores: contnuous 82

3 wavelets transform (CWT) and dscrete wavelet transforms (DWT). In ths research, dscrete wavelet transform s used. The DWT algorthm s capable of producng coeffcents of fne scales for capturng hgh frequency nformaton, and coeffcents of coarse scales for capturng low frequency nformaton. For a mother wavelet functon ψ and for a gven sgnal f() t, a DWT can be expressed as follows [3]: ( ) j 2 j j0, j0, ( ) j, 2 (2 ) j j0 (1) f t c t t where j s the dlaton or level ndex, s a translaton or scalng ndex, j0, s a scalng functon of coarse scale coeffcents, c j0,, j, are the scalng functon of detal coeffcents, and all functon of (2 j t ) are orthonormal. Wavelet processng has two stages: decomposton and reconstructon. The decomposton computes the convoluton between the load demand and hgh pass/ low pass flter, whle the reconstructon calculates the convoluton between the load and nverse flter. A mother wavelet based on Daubeche2 (Db2) s used for the flter s coeffcents. It used to decompose an nput load demand nto low frequency and hgh frequency components. The decomposton s mplemented by usng multchannel flter ban: one, two, three, and four channels. The reconstructed detals and approxmatons are true parameters of the orgnal sgnal as follow [3]: S = A1 + D1 (Level 1 ) = A2 + D2 + D1 (Level 2 ) = A3 + D3 + D2 + D1 (Level 3 ) = A4 + D4 + D3 + D2 + D1 (Level 4 ) For example, n Level 1, the coeffcent vectors A1 and D1 are produced by down samplng and only half a length of the orgnal sgnal. Thus, they cannot drectly combne to reproduce the sgnal. It s necessary to reconstruct the approxmatons and detals before combnng wth each other. Internatonal Journal of Soft Computng and Engneerng (IJSCE) ISSN: , Volume-1, Issue-6, January 2012 frequency components of the load demand, and weather factors such as maxmum temperature, mnmum temperature, mean temperature, humdty, and ranfall. Subsequently, the hgh frequency component, see n sub-secton C.3, feature nputs of neural networ are also selected based on correlaton testng results; nclude the low frequency components of the load demand and economc factors le consumer prce ndex (CPI) and ndustral ndex (IDI). Hence, the artfcal neural networ s one of a good choce to apply for the load demand forecastng problem because ths technque s not requrng explct models to represent the complex relatonshp between the load demand and factors. The neural networ algorthm presented n ths paper composes of three layers: the nput layer, the hdden layer, and the output layer based on feed-forward bac propagaton algorthm (FFBP). The nput varables come from hstorcal and present data of factors affectng the load demand. The fundamental structure of ths algorthm can be presented n Fg.4. C.3 Correlaton method Equaton (2) s the Pearson s coeffcent correlaton equaton. It s used to evaluate correlaton between varables and s one of the most famlar measures of dependence between two quanttes to show how good a lnear relatonshp among varables or factors s. In the case that the Pearson correlaton s +1, t wll sgnfy a perfect lnearly postve ncreasng correlaton trend. If the Pearson correlaton s -1, t wll ndcate a perfect lnearly negatve declnng correlaton trend. If the Pearson correlaton s between 1 and -1, t wll ndcate a degree of lnear dependence between the two varables. Lastly, f the Pearson correlaton s zero, t shows that there s no relatonshp between the varables. Before a load forecastng, we must correlate all varables wth the energy consumpton load demand to choose the approprate varables for the best feature nputs for the research model. r XY ( X X )( Y Y) ( X X ). ( Y Y) 2 2 Where X and Y are varables and r s the Pearson product moment of correlaton. (2) III. IMPLEMENTATION Fg.3 Wavelet decomposton and reconstructon. C.2 Artfcal ntellgence-neural networ algorthm In ths artcle, the neural networ models are proposed and used for separatng the low frequency and hgh frequency components of demand n each level. The low frequency components, feature nputs of neural networ are selected based on the correlaton testng results; nclude hgh A. Case study To demonstrate the low and hgh frequency demand features employs the hstorcal electrc demand data, whch were recorded as monthly from the Electrcty Generatng Authorty of Thaland (EGAT) from January 1997 to December 2007 are decomposed. Weather factor and an economc factor are used n ths research. The hstorcal load and approxmate (A) demand, the weather factor, and the economc factor for tranng and testng n the forecastng model were normalzed nto the nterval 0.00 to 1.00 by usng equaton (3), as follows: 83

4 Wavelet and Neural Networ Approach to Demand Forecastng based on Whole and Electrc Sub-Control Center Area ( y mn( y)) z (max( y) mn( y)) Where z s normalzed value and y s data nformaton. In developng model the cyclcal component (detal) was normalzed nto nterval to 1.00 by usng equaton (4), as follows: 2( y mn( y)) z 1 (max( y) mn( y)) B. Preprocessng and forecastng stages In Fg.4 descrbes the overall structure of the electrc load demand forecastng for ths research. The wavelet transform and neural networ algorthms are used n the research. The man steps proposed for the load demand forecastng model are as follows: 1) The Northern, Northeast, Southern, and Central sub-control center area forecastng wll also use sx stages below: 1.1) The frst, an orgnal sgnal of load demand s decomposed to hgh and low frequency by usng db2 mother wavelet (db2) for calculatng the coeffcent of the detals (D) and approxmate (A) components. The level of wavelet transform vares from one level, two levels, three levels, and four levels based on wavelet dscrete transform (DWT). From each level decomposton by wavelet method wll obtan a detal and a approxmate as follows: there are a detal and a approxmate components for a level, these are D1 and A1; there are the two detals and a approxmate components for the two levels, these are D1, D2, and A2 respectvely; there are the three detals and a approxmate for the three levels, these are D1, D2, D3, and A3 respectvely ; lastly, there are the four detals and a approxmate for the four levels of wavelet transform, these are D1, D2, D3, D4, and A1 respectvely. All data above after decomposng are recorded and taen them to step ) Coeffcent components from step 1.1 are reconstructed to the actual components usng smlar mother wavelet (db2); these are an actual detal and approxmate components of each level. 1.3) Actual detal and approxmate components are taen them to fnd the relatonshp between each component n each level and factors: temperature, humdty, ranfall, consumer prce ndex, and ndustral ndex. The correlated method by usng equaton (2) s used. 1.4) The factor that s related wth the component more than that 40 percent (up) wll be chosen t beng as qualty nputs for a neural networ model. Note that, the factor s chosen n each component of wavelet dfferng followng the related value. 1.5) The feature nputs of a model NND forecast conssts of the detal component and factors are selected from step 1.4 and the feature nputs of a model NNA forecast conssts of an approxmate (A) and factors are also selected from ths step. 1.6) Subsequent to partal forecasts n each sub-model: for example, the 4 levels, sub-model conssts of A4, D4, D3, D2, and D1 models; the output for all sub-model are ntegrated to (3) (4) the fnal forecasted value. 2) The output of each sub-control area wll ntegrate together for the whole area forecastng of the country. 3) Analyses and summary of all forecast results are carred out. Northern area Wh hstorcal data Northeast area Central & metropoltan, Bango area Southern area Wh hstorcal data Decomposton 1 f HIGH 1 f LOW 2 f HIGH 2 f LOW D1 D2 A2 r xy Reconstructon ( X X ).( Y Y ) ( X X ).( Y Y ) 2 2 Forecastng Feature selecton Input NN D 1 Feature selecton Feature selecton Input Input Southern area sub-model NN D 2 NN A 2 Northeast area Forecastng Northern area Forecastng +12 Months + Central & metropoltan, Bango area Forecastng Fg.4 Archtecture model of forecastng. IV. RESULTS AND COMPARISON Southern area Forecastng The Mean Absolute Percentage Error (MAPE) was calculated by the use of equaton (5). MAPE ndcates the forecastng accuracy and the dentfable varable n the model, MAPE s gven by the followng: N 1 WhA WhF MAPE 100 N (5) Wh 1 Where WhA s the actual energy consumpton load demand, WhF s the forecast load demand, N s the number of the month forecasted. Table I shows the results of forecastng n ths research. The column number one, two, three, four, fve, and sx present the months, actual load demand, one level, two levels, three levels and four levels respectvely. The demand s forecasted nterval January to December n 2007 unt n lo-watt hour (Wh). Other columns le L-no.1, 2, 3, and 4 show the percentage error n each month n year 2007 after forecastng. Hence, n ths table shows the forecasted results obtaned from the four models by varyng the level of wavelet transform n preprocessng stage. The mean absolute percentage error (MAPE) of one level, two levels, three levels and four levels were 3.58, 2.25, 2.83, and 3.45 percent respectvely, were presented n table I, n whole area forecastng. The sutable of level of wavelet has been nvestgated usng the MAPE computaton for fndng the best level that s, two levels. Whereas table II presents sub-control center area forecastng based on same method that s used wavelet and neural networ approaches to forecasts. In ths table demonstrates the forecasted results obtaned from the A + 84

5 four models by varyng the level of wavelet transform n pre-processng stage. The mean absolute percentage error (MAPE) of one level, two levels, three levels and four levels were 2.36, 2.07, 2.93, and 3.48 percent respectvely, were presented n Table II. Furthermore, the sutable of level of wavelet n ths nvestgaton has been demonstrated usng the MAPE calculaton for fndng the best level that also s, two levels smlarly. Therefore, the comparson between the whole area forecastng and sub-control center area forecastng can demonstrate n table I and II. The best level of wavelet transform for two methods s two levels whereas when we forecast by usng sub-control center area, t can show better result than whole area forecastng. Table I: Mean absolute percent error (MAPE) n each level of whole area forecastng. Month Actual L-no. Wh January February March Aprl May June July August September October November December MAPE Table II: Mean absolute percent error (MAPE) n each level of electrcty sub-control center area forecastng. Month Actual L-no. Wh January February March Aprl May June July August September October November December MAPE Internatonal Journal of Soft Computng and Engneerng (IJSCE) ISSN: , Volume-1, Issue-6, January 2012 (MAPE) better than that other level and sub-control center area can show better result than whole area forecastng. The objectve of load forecastng has sgnfcantly forecast for fuel reserve plannng n the power system. ACKNOWLEDGMENT Ths wor was funded by the Offce of the Hgher Educaton Commsson. Ptu Bunnoon was supported by CHE Ph.D. Scholarshp. The authors would le also to than the Tha Meteorologcal Department, Mnstry of Transport and Communcatons; the Mnstry of Commerce; the Offce of the Natonal Economc and Socal Development Board; and lastly the Electrcty Generatng Authorty of Thaland (EGAT) n provdng her valuable data and nformaton. REFERENCES [1] Q Wu, Hybrd model based on wavelet support vector machne and modfed genetc algorthm penalzng Gaussan noses for power load forecasts, Internatonal journal of expert systems wth applcatons pp , [2] Yng Chen, Peter B. Luh, Che Guan, Yge Zhao, Laurent D. Mchel et.al., Short-term load forecastng: Smlar day-based wavelet neural networs, IEEE Trans.on power syst., vol.25, pp , [3] Ajay Shehar Pandey, Devender Sngh, and Sunl Kumar Snha, Intellgent hybrd wavelet models for short-term load forecastng, IEEE Trans.on power syst., vol. 25, pp , [4] N. Amjady, and F. Keyna, Short-term load forecastng of power systems by combnaton of wavelet transform and neuro-evolutonary algorthm, Internatonal journal of energy, vol.34, pp.46-57, [5] N. M. Pndorya, S. N. Sngh, and S. K. Sngh, An adaptve wavelet neural networ based energy prce forecastng n electrcty marets, IEEE Trans.on power syst., vol.23, pp , [6] D. Benaouda, F. Murtagh, J.-L. Starc, and O. Renaud, Wavelet-based nonlnear mult-scale decomposton model for electrcty load forecastng, Internatonal journal of neuro computng, vol.70, pp , [7] Ta Nenglng, Jurgen Stenzel, and Wu Hongxao, Technques of applyng wavelet transform nto combned model for short-term load forecastng, Internatonal journal of electrc power systems research, vol.76, pp , [8] Ba-Lng Zhang, and Zhao-Yang Dong, An adaptve neural-wavelet model for short term load forecastng, Internatonal journal of electrc power systems research, vol.59, pp , [9] Tongxn Zheng, Adly A. Grgs, and Elham B. Maram, A hybrd wavelet-kalman flter method for load forecastng, Internatonal journal of electrc power systems research, vol.54, pp.11-17, [10] S. J. Yao, Y. H. Song, L. Z. Zhang, and X. Y. Cheng, Wavelet transform and neural networ for short-term electrcal load forecastng, Internatonal journal of energy converson and management, vol.41, pp , V. CONCLUSION The load demand forecastng based on the sutable level of wavelet transform, and a neural networ method, approaches to the whole and sub-control center area forecastng, s proposed. The load demand data whch employs the hstorcal data from Electrcty Generatng Authorty of Thaland (EGAT) s decomposed nto one, two, three, and four levels of dfferent frequences. The correlaton technque s used to choose the sutably affectng factors for each frequency component of load demand. The factors whch affected n each level are chosen for neural networ nputs. As a result, two levels s gven the Mean Absolute Percentage Error Ptu Bunnoon (Member IEEE), receved the B.S. degree from Kng Mongut's Insttute of Technology Ladrabang, Thaland, n 1998, and the M.S. degree n electrcal engneerng from Prnce of Songla Unversty, Thaland, n Hs research nterest s an applcaton of artfcal ntellgence to power system plannng and operaton. 85

6 Wavelet and Neural Networ Approach to Demand Forecastng based on Whole and Electrc Sub-Control Center Area Kusumal Chalermyanont receved the B.S. degree from Prnce of Songla Unversty, Thaland, n 1993, the M.S. degree n electrcal engneerng from Unversty of Colorado at Boulder n 1999, and the Ph.D. n electrcal engneerng from the Unversty of Colorado at Boulder n Her research nterests are power electroncs, magnetc desgns for power electroncs, renewable energy system/management. Chusa Lmsaul receved the B.S. degree from Kng Mongut's Insttute of Technology Ladrabang, Thaland, n 1978, and the D.E.A. degree from INSAT France, n 1982, and Docteur Ingeneur form INSAT France, n Hs research nterests are dgtal sgnal processng, sensors and nstrumentatons, and automaton. 86