Comparison between the short-term observed and long-term estimated wind power density using Artificial Neural Networks.

Transcription

1 Comparso betwee the short-term observed ad log-term estmated wd power desty usg Artfcal Neural Networks. A case study S Velázquez, JA. Carta 2 Departmet of Electrocs ad Automatcs Egeerg, Uversty of Las Palmas de Gra Caara, Campus de Tafra s/, 3507 Las Palmas de Gra Caara, Caary Islads (Spa). Tel.: , Fax: E-mal address: svelazquez@dea.ulpgc.es 2 Departmet of Mechacal Egeerg, Uversty of Las Palmas de Gra Caara, Campus de Tafra s/, 3507 Las Palmas de Gra Caara, Caary Islads ( Spa). Tel.: , Fax: E-mal address: jcarta@dm.ulpgc.es Abstract The ecoomc feasblty of a wd project s depedet o the wd regme sce t reles o the power output of the turbes over the stallato s workg lfe. Cosequetly, the teraual varablty of wd speed at a potetal wd eergy coverso ste s a ssue of captal mportace. Usually a wd data measuremet campag s lmted to a perod o loger tha oe year (.e. short-term). Therefore, the process of decso-makg for wd farm costructors must be based ths short-term data. Varous methods have bee proposed the scetfc lterature for estmato of the log-term wd speed characterstcs at such stes. These methods use smultaeous measuremets of the wd speed at the ste questo ad at oe or several earby referece stes wth a log hstory of wd data measuremets. I ths paper, log-term wd power destes whch have bee estmated through the use Artfcal Neural Networks (ANNs), wll be compared to those whch have bee calculated by meas of the short-term wd data (.e. cosdered to be represetatve of log-term wd performace). Mea hourly wd speeds ad drectos calculated a 0 year perod of tme at sx weather statos located o sx dfferet slads the Caara Archpelago (Spa) were used ths study. Amog the dfferet coclusos whch ths study revealed, we ca hghlght that the wd resource estmato based o ANNs s better tha that depedat o short-term wd data. Ths s true whe the correlato coeffcet betwee the referece ad caddate weather stato s of 0.6. Keywords: Wd Power, Log-Term, Artfcal Neural Network, Wd Speed, Wd Farm Itroducto Gve that there s a cubc relatoshp betwee wd speed ad wd power desty, t s uderstadable that the electrcal eergy obtaable wth a wd power turbe s very codtoed to the wd regme. As stated by Hester ad Peell [], the teraual varablty of wd speed at a potetal wd eergy coverso ste s a very mportat ssue. The frst cocer about a ste uder cosderato for a wd power stato s wth the log-term (may years) mea wd speed [2-8]. However, o may occasos there are o hstorcal wd data measuremet seres avalable for the caddate ste. Ths s a major obstacle for the assessmet of the ecoomc feasblty of a wd farm project o a tme horzo equvalet to the useful lfe of the stallato [9- ]. Oe opto that ca be used to get aroud ths lack of data for the caddate ste s to coduct a wd data measuremet campag that covers a suffcet umber of years. Accordg to Hester ad Peell [], accurate estmato of the mea values of the wd performace s dffcult wth less tha 0 years worth of data. Ths opto etals a evtable crease the costs of the measuremet campag ad, more mportatly, the postpoemet of ay fal decso-takg for a ormally uacceptably log perod of tme. I the partcular case of the Caary Islads, the stallato of wd farms s regulated through wd power teders stgated by the Caara Govermet through legslatve decrees [2]. Normally oly a short perod of tme s gve betwee publcato of these legslatve documets ad the deadles for the presetato of proposals. As a cosequece, a RE&PQJ, Vol., No.9, May 20

2 measuremet wd data campag s usually lmted to just oe year. The wd farm developers udertake the ecoomcfacal studes ad submt ther correspodg proposals wth the wd formato collected over ths perod of tme. The wd farms presetly stalled the Caara Archpelago are geerally foud the areas of hghest wd potetal. At the ed of 2009, the total stalled wd power the slads was 4 MW. The strategc target set by the Caara Govermet s to have 025MW stalled wd power by 205 [3]. The ew oshore wd farms that wll be stalled the future wll have to be sted areas wth less wd potetal tha the areas that are curretly beg exploted. Gve the shape of the partal load zoe of the powerwd speed curves of the wd turbes, the costeffectveess of these stallatos wll be more sestve to wd speed varatos. I order to get aroud the afore metoed coveece, the mea teraual wd performace at the caddate ste ca be estmated through statstcal methods. These methods [3,4] rely o the exstece of referece statos stalled at earby stes for whch logterm measuremets (0 or more years) of the wd resource are avalable. These methods also requre the results of a relatvely short-term (ormally oe year) wd data measuremet campag at the caddate ste. I addto, part of the wd data avalable for the referece stato must cocde legth of tme ad date wth the data measured at the caddate ste. Amog the varous methods used, some employ automatc learg techques [4-9] whch take ther sprato from statstcal learg algorthms, such as Bayesa Networks (BNs) [4] ad Artfcal Neural Networks (ANNs) [5-9], the latter from the bologcal scece feld. I the other had, there are methods that use tradtoal Measure-Correlate-Predct (MCP) algorthms [20-2]. I ths paper, ANNs have bee used as a tool to estmate the mea wd speed ad wd power at a caddate stato for whch oly complete wd data seres are avalable. The ANNs used ths paper were comprsed of three layer etworks wth feedforward coectos. More specfcally, multlayer perceptro topologes (MLPs) were used [22-23]. A sgle hdde layer wth 5 euros was employed so as ot to crease the trag tme. Ths archtecture has demostrated ts ablty to approxmate satsfactorly ay cotuous trasformato [22-23]. I geeral, oly the wd speeds recorded at referece statos are used as sgals of the put layer of multlayer perceptro (MLP) archtectures. The output layer represets the caddate stato wd speeds. I addto to the correspodg mea wd speeds, the mea hourly wd drectos are also used the put layers. I ths study, wd power destes estmated through the use of ANNs wll be compared wth those obtaed whe cosderg the short-term wd data perod (oe year) of the caddate ste to be represetatve of the log-term wd performace at the same ste. Dfferet metrcs wll be used to assess ths comparatve aalyss: the Mea Absolute Relatve Error (MARE ()) ad Pearso s correlato coeffcet betwee measured ad estmated data (CC (2)). CC O OE E 2 O O E E MARE O E O Where E are the estmated data ad O, are the observed or measured data ad s the umber of data 2 Meteorologcal data used The meteorologcal data used ths paper (mea hourly wd speeds ad drectos) were recorded over the years at sx weather statos stalled o sx dfferet slads the Caara Archpelago (Spa). Ths formato was provded by the State Meteorologcal Agecy (Spash tals: AEMET), lked to the Mstry of Evrometal, Rural ad Mare Evros of the Spash Govermet. I table I, the geeral data of the weather statos ca be studed. Table II shows the wd speed correlato coeffcets (3) betwee the dfferet weather statos used. It has bee calculated usg all the perod of tme avalable. 2 () (2) RE&PQJ, Vol., No.9, May 20

3 Table I: Weather statos used the study WEATHER STATION YEARS OF DATA AVALIABLE HEIGHT Geographcal Coordates LONG-TERM ANNUAL MEAN WIND SPEED (m/s) Lattude (N) Logtude (W) Alttude (m) WS º55'44" 5º23'20" 6 7,4 WS º57'7" 3º36' 0 5,82 WS º27'0" 3º5'54" 24 5,83 WS º2'35" 6º34'6" 5 5,64 WS º36'47" 7º45'36" 85 4,82 WS º48'50" 7º53'0" 30 5,96 Table II: Lear correlato coeffcets, R, betwee the wd speeds of the dfferet aemometer weather statos. Log-Term Correlato Coeffcet (R). WS-,00 0,67 0,67 0,49 0,57 0,47 WS-2 0,67,00 0,65 0,49 0,49 0,44 WS-3 0,67 0,65,00 0,5 0,52 0,46 WS-4 0,49 0,49 0,5,00 0,38 0,26 WS-5 0,57 0,49 0,52 0,38,00 0,49 WS-6 0,47 0,44 0,46 0,26 0,49,00 data. The proporto of data selected for each of these processes was 80% ad 20%, respectvely. The trag data subset s used for estmato of the weghts of the ANNs. The valdato data subset s used to check the trag progress of the ANNs, ad thus allowg the optmzato of ther parameters. That s, they are used to measure the degree of geeralsato of the ANNs. Wd data o the remag years at the referece stato were used to estmate log-term data the caddate stato. The dfferet metrcs that wll be used the aalyss were calculated o comparg estmated ad observed data. R Vr VrVc Vc 2 2 Vr Vr Vc Vc (3) The put sgals of the etwork clude the seres of wd speeds ad wd drectos of a referece stato. Output sgals, o the other had, cluded the seres of wd speeds of the caddate stato. I these models, the wd drecto sgal s troduced as agular magtude (0º-360º). Notce that the agle correspodg to the ortherly drecto s take as agle 0º (Fgure ) Where Vr ad Vc are the measured wd speeds at the referece ad caddate weather stato, respectvely. Vr ad Vc are the mea wd speeds at the referece ad caddate weather stato.. 3 Methodology The dfferet metrcs used ths paper have bee assessed from the follow two hypotheses. Hypothess A: The log-term wd resource s estmated usg the Artfcal Neural Networks (ANNs). The Log-Term wd data estmato process oly used those weather statos from whch there were 0 years worth of wd data. I the creato of the etworks, oly oe of the years of data collecto was used ( ths case, t was year 2008). The weather formato was dvded to two radom subsets of dfferet data: trag data ad valdato Fg.. Schematc dagram of a ANN wth the wd speed (V) ad the wd drecto (D) of oe Referece weather stato as put sgals, ad the wd speed (V) of oe Caddate (target) stato as output sgal. Hypothess B: Short term wd data s cosdered as the log term wd resource performace RE&PQJ, Vol., No.9, May 20

4 I ths case there s ot a referece stato. The short term wd data observed of the caddate stato are cosdered as the log term wd resource performace. These oe wll be used to calculate log term wd power at the caddate ste. 4 Aalyss of Results Table III shows the results obtaed for the MARE of the wd speed (2) the case hypothess B. Data from each year have bee cosdered as estmated data, whlst logterm measured data have bee calculated from all avalable data (4). V LT j m WhereV m LT j V ST j, (4), s the measured log-term wd speed the ST stat of tme j; V j,, s the measured short-term wd speed the stat of tme j ad year ; ad m, s the avalable umber of years The last row table III shows the mea results for the MARE for every weather stato. Wth these results t s possble to assess the mea value for the MARE of the wd speed for the sx weather statos. Its value s Table III: MARE of the wd speed whe cosderg the short-term data to be represetatve of the log-term wd performace Caddate Weather stato year cosdered as the log Term MARE of the wd speed 999 0,40 0,38 0,34 0,45 0,35 0, ,54 0,48 0,43 0,50 0,39 0, ,52 0,5 0,44 0,59 0,45 0, ,49 0,47 0,43 0,60 0,46 0, ,49 0,5 0,45 0,55 0,47 0, ,52 0,45 0,45 0,52 0,47 0, ,45 0,44 0,45 0,45 0,47 0, ,44 0,44 0,37 0,52 0,52 0, ,44 0,42 0,37 0,48 0,45 0, ,40 0,38 0,35 0,44 0,39 0,36 Mea Results 0,47 0,45 0,4 0,5 0,44 0,49 To estmate the log term wd resource wth the ANNs method, the speed ad drecto of the wd from oly oe referece stato has bee used as the put layer parameter. Each log term wd performace has bee calculated takg the rest of the weather stato (oe by oe). Therefore, there wll be fve dfferet estmatos for each weather stato. I fgure 2 s compared the results obtaed wth those of hypothess B (Table III) MARE 0,65 0,6 0,55 0,5 0,45 0,4 0,35 0,3 0,25 0,2 Caddate Weather Stato Referece Stato, WS- Referece Stato, WS-2 Referece Stato, WS-3 Referece Stato, WS-4 Referece Stato, WS-5 Referece Stato, WS-6 Hypotess B Fg. 2. Comparso of the MARE of the wd speed for each weather stato whe t s used the ANNs wth oe referece stato to estmate log-term wd speed (Hyp. A), wth the case where s used the short-term data to be represetatve of the log-term wd performace (Hyp. B) RE&PQJ, Vol., No.9, May 20

5 It s observed that the log term estmato wth ANNs, the best results are obtaed whe the correlato coeffcet, R, betwee the wd speed of the caddate ad referece weather stato s largest. The worst results are obtaed opposte cases. I four of the total sx cases, where R s betwee 0.5 ad 0.6 (see fgure 3 ad table II), the MARE of the wd speed for hypothess A s better tha that for hypothess B. Whe R (3) s hghest (up to 0.6), the results for hypothess A are better all cases. From the sx cases where R are betwee 0.5 ad 0.6 (see fgure 3 ad the table II), four of them, the MARE of the wd speed the hypothess A are better tha the hypothess B. Whe R (3) s greater tha 0.6, the results for the hypothess A are better all of the cases. As metoed by S. Velázquez et al[8], whe more weather statos are added to the put layer the ANNs method, the estmato are always better whatever be the correlato coeffcet betwee the secod ad follows weather stato added wth the caddate weather stato. As metoed by S. Velázquez et al [8], the larger the umber of weather statos added to the put layer usg the ANNs method, the better performace of the estmato. Ths s true, regardless of the correlato coeffcet, R, for the secod ad cosecutve statos added, as well as for the caddate weather stato. It s therefore possble for ths cocluso to be geeralzed for ay correlato coeffcet smaller tha 0.6. Fgure 3 shows the comparso for the correlato coeffcet exstet betwee measured ad estmated wd speed, CC (), for hypotheses A ad B. CC 0,9 0,8 0,7 0,6 0,5 0,4 0,3 0,2 Caddate Weather Stato Referece Stato, WS- Referece Stato, WS-2 Referece Stato, WS-3 Referece Stato, WS-4 Referece Stato, WS-5 Referece Stato, WS-6 Hypotess B Fg. 3. Comparso of the metrc CC whe t s used the ANNs to estmate log-term wd speed (Hyp. A), wth the case where s used the short-term data to be represetatve of the log-term wd performace (Hyp. B) For the CC, the coclusos reached (fgure 3) are the same as those for the MARE of the wd speed (fgure 2). That s, results for the CC are better the greater the value of R. For correlato coeffcets, R, greater tha 0.6, better results have bee obtaed for hypothess A tha for hypothess B. If more weather statos are cluded the put layer of the ANNs model, results obtaed wll eve be better [8]. For the MARE of the wd power (fgure 4) results for hypothess A as good or better that for hypothess B for all of the cases, cludg whe R s betwee 0.5 ad 0.6. MARE 8,2 7,2 6,2 5,2 4,2 3,2 2,2,2 0,2 Caddate Weather Stato Referece Stato, WS- Referece Stato, WS-2 Referece Stato, WS-3 Referece Stato, WS-4 Referece Stato, WS-5 Referece Stato, WS-6 Hypotess B Fg. 4. Comparso of the MARE of the wd power for each weather stato whe t s used the ANNs wth oe referece stato to estmate log-term wd speed (Hyp. A), wth the case where s used the short-term data to be represetatve of the log-term wd performace (Hyp. B) RE&PQJ, Vol., No.9, May 20

6 5 Coclusos Whe short-term data s take to be represetatve of the log-term wd speed, the Mea Absolute Relatve Error (MARE) of the wd speed ad wd power (for all the studed cases wth the dfferet weather stato), were 0.46 ad 4.47, respectvely. For all of the aalyzed cases, f the correlato coeffcet R (3) betwee the wd speeds of the caddate ad referece weather stato, s greater tha 0.6, the results the estmato of log term wd resource wth the Artfcal Neural Networks methods (ANNs) s better tha whe usg short-term data to be represetatve of the log term wd resource. For the MARE of the wd power, the results obtaed the estmatg usg the ANNs method also are better whe R s the rage [7] Mabel C, Feradez E. Aalyss of wd power geerato ad predcto usg ANN: A case study. Reewable Eergy 2008, 33, pp [8] S. Velázquez, JA. Carta, JM. Matías, Ifluece of the put layer sgals of ANNs o wd power estmato for a target ste: A case study, Reewable ad Sustaable Eergy Revews, 5, 20, pp [9] DA. Fadare, The applcato of artfcal eural etworks to mappg of wd speed profle for eergy applcato Ngera, Appled Eergy, 87, 200, pp [20] Rogers AL, Rogers JW, Mawell JF. Comparso of the performace of four Measure-Correlate-Predct algorthms. Joural of Wd Egeerg ad Idustral Aerodyamcs 93, 2005, pp [2] Clve JM. No-learty MCP wth Webull dstrbuted wd speeds. Wd Egeerg 2008, 32, [22] JC. Prcpe, NR. Eulao, WC. Lefebvre, Neural ad Adaptve Systems. Fudametals Through Smulatos, frst ed. New York: Joh Wley & Sos, Ic.; [23] T. Masters, Practcal Neural Network Recpes C++, frst ed. Calfora: Morga Kaufma Publshers; 993. If more tha oe weather stato s used the put layer of the ANNs, t s possble to obta a better result the log term estmato tha usg a sgle stato. Refereces [] TR. Hester, WT. Peell, The stg hadbook for large wd eergy systems. st ed. New York: WdBook, 98. [2] V. Corad, LW Pollack, Methods clmatology. Secod ed. Cambrdge-Massachusetts: Harvard Uversty Press, 962. [3] JA. Carta, J. Gozález, Self-suffcet eergy supply for solated commutes: wd-desel systems the Caary Islads. The Eergy Joural 200, 22, pp [4] CI. Asplde, DL Ellott, LL Wedell, Resources assessmet method, stg, ad performace evaluato. I: Guzz R, Justus CG. Physcal clmatology for solar ad wd eergy, New Jersey: World Scetfc; 988, pp [5] PC. Putam, Power from the wd. st ed. New York: Va Nostrad Rehold Compay; 948. [6] GW Koeppl, Putam s power from the wd. Secod ed. New York: Va Nostrad Rehold Compay, 982. [7] CG. Justus, K. Ma ad AS. Mkhal, Iteraual ad moth-to-moth varatos of wd speed. Joural of Appled Meteorology 8, 979, pp [8] PA. Daels ad TA. Schroede, Stg large wd turbes Hawa. Wd Eg 988; 2, pp [9] Nelso V. Wd eergy.st. ed. FL: CRC Press; [0] Hau E. Wd turbes. 2d ed. New York: Sprger; 2005 [] Swft-Hook DT. Secod wd. The Egeer 988, 267, pp 30-3 [2] 32/2006 of 27 March Decree, whch regulates the stallato ad operato of wd farms the area of the Caara Archpelago. BOC 6, 2006, pp [3] Eergetc Pla of Caary Islads, [4] JA. Carta, S. Velázquez, JM. Matías, Use of Bayesa Networks classfers for log-term mea wd turbe eergy output estmato at a potetal wd eergy coverso ste, Eergy Coverso ad Maagemet 52, 20, pp [5] M. Blgl, B. Sah, A. Yasar, Applcato of artfcal eural etworks for the wd speed predcto of target stato usg referece statos data, Reewable Eergy 32, 2007, pp [6] Kalogrou SA. Artfcal eural etworks reewable eergy systems applcatos: a revew. Reewable ad Sustaable Eergy Revews 5, 200, pp RE&PQJ, Vol., No.9, May 20