Comparison of the Efficiency of Input Determination Techniques with LM and BR Algorithms in ANN for Flood Forecasting, Mun Basin, Thailand

Transcription

1 Interntionl Journl Computer n Eletril Engineering, Vol., No., April 01 Comprison Effiieny Input Determintion Tehniques with n Algorithms in ANN for Foresting, Mun Bsin, Thiln Twee Chipimonplin n Thveesk Vngpisl ovious tht performne ANN moel epens on lerning lgorithms, input vriles n numer hien noes []. Lerning lgorithms in Mtl stwre for exmple, Sle Conjugte Grient (SCG), Levenerg-Mrqurt () n Byesin Regulriztion () hve ifferent pprohes, SCG is goo for pttern reognition, is fstest lgorithm n is n utomte regulriztion for improving generliztion []. Chipimonplin [] reviewe mny input etermintion tehniques tht hve een use for seleting input vrile in ANN moel for exmple, ross orreltion, stepwise regression, geneti lgorithm, PMI, et. He lso foun tht suitle tehniques for foresting wter level in Ping Bsin, Thiln, inlue ross orreltion, stepwise regression, ross orreltion plus stepwise regression n geneti lgorithm. Thus, this stuy ws onute to etermine effiieny two lerning lgorithms; n, for foresting wter level t speifie se stuy, Mun Bsin, Thiln. The inputs for moeling were otine y using four input etermintion tehniques, s well s supervision seletion, n y using ll input vriles. The tril n error numer hien noes ws lso tke into ount. Astrt The ims this stuy re to improve moel perforemne with input seletion n omprision fst lerning n utomte lerning lgorithms. Therefore, methoology ws, first, investigtion ross orreltion, stepwise regression, ross orreltion with stepwise regression, geneti lgorithm, supervise n ll input) n ompring lerning lgorithms: Levenerg Mrqurt- n Bysin Regulriztion-). The ANN ws use to forest wter level t guge sttion M.. The results showe seleting input from geneti lgorithm gives est result for foresting floo pek n provie etter results thn prtiully t floo pek. Inex Terms Neurl network, floo foresting, Levenerg-Mrqurt, Byesin Regulriztion, Mun Bsin. I. INTRODUCTION ing is serious prolem lmost every yer in some res Thiln. In 0, tropil storm NOCK-TEN psse Thiln, ringing with it torrentil rins, in whih over 0 provines were flooe, inluing Ching Mi n Uon Rththni. For effetive floo prevention, n erly wrning system is neessry. Hyrologil moel suh s MIKE [1], TANK [] n Artifiil Neurl Network (ANN) [], [], hve een evelope n pplie for floo foresting. Notly, ANN moel, whih is lk ox moel, mkes use t riven metho. The vntges ANN moel re tht it oes not require physil t or fiel t, n hs less omputtion time thn or pproh moels. It is lso esy to upte when new t eome ville. Therefore, Hyrology Division, Royl Irrigtion Deprtment, Thiln, hs hosen to pply ANN moel for floo foresting for over sins in Thiln []. For improving ANN moel for its etter n more effetive use in for floo foresting, re re severl methos, for exmple, seleting input vrile from input etermintion tehniques [], [], ing extr input vriles [], [], n seleting ifferent trnsfer funtions [], seleting ifferent lerning lgorithms []. However, it hs to e orne in min tht ifferent sins my require ifferent ANN moels euse ifferent runf ehviors in sin n ifferent t ville. In ition, it is II. STUDY AREA Mun Bsin is lrgest river sin in Thiln, overing n re pproximtely 1,000 squre kilometers or 1% ountry s ln re (Fig. 1). Inuntion prolem, whih ours frequently in Mun Bsin re, ten tkes ple in re long riversie Mun River in lower prt sin, espeilly in Uon Rththni provine. The mjor uses floo relte prolems in Mun Bsin inlue (1) inequte wter storge pity, preventing wter flow from retring from upper prt sin, () limite ringe pility ue to nturl ostles in lower setion Mun River, n () expnsion settlers into floo thretene res. The Uon Rththni provine is 0 kilometers est Bngkok. In pst 0 yers, floo events hve een reore for re. The highest level river runf is reore usully etween Septemer n Otoer. The floo prolem tkes ple in ommunity res long river nks in ity re Uon Rththni euse se res re lote in lower prt intereption point Mun River, ing to woes is high rinfll intensity re. It is well-known n epte ft tht strem wter level +1 meters ove men se level, mesure t M. in Uon Rththni, is use for onset floo Mnusript reeive My, 01; revise July 1, 01. This work ws supporte in prt y Ntionl Reserh Counil Thiln. Twee Chipimonplin is with Deprtment Geogrphy, Ching Mi University, Ching Mi, Thiln (e-mil: twee.@mu..th, hipimonplin@hotmil.om). Thveesk Vngpisl is with Civil Engineering Deprtment, Uon Rththni University, Uon Rththni, Thiln (e-mil: thveesk.v@u..th). DOI:./IJE.01.V.00 0

2 Interntionl Journl Computer n Eletril Engineering, Vol., No., April 01 prolems in re. The highest wter level +. meters ove men se level (. meters) ws reore in yer 1, n this flooing use extensive mge to ommunities, environment, n eonomy. It ws stte in wter-resoures mngement pln for Mun Bsin [] tht trying to solve floo prolems in Mun Bsin y resorting to onstrutionl mesures ws uneonomil, n tht promulgtion non-onstrutionl mesures suh s floo forest n wrning systems, ln utiliztion ontrol, n evution floo thretene res re wht must e implemente. () () Fig. 1. Mun sin (eite from []) III. METHODOLOGY Inputs for moel were reore wter levels t guge sttions in stuy re. Four upper sttions (M., M.1, M.1 n M.1) were use to forest wter level t M. guge sttion (Fig. 1). The M. n M.1 re upstrem guge sttions M. in min river, while M.1 n M.1 re guge sttions in triutry min river. The longest istne kilometers is from M.1 to M., followe y kilometers, whih is from M.1 to M.. The ville t in stuy were hourly wter levels. The ville t from five sttions overe five yer perio from 00 to 0, in whih four floo events h ourre inluing most evstting floo in 0 (Fig. ). At M., wter level meters enotes onset floo in stuy re. In orer to explore effiieny lerning lgorithms, experiment ws ivie into five for ifferent input vriles. The moel performnes, etween n, were ompre. Fig.. Hyrogrphs t M. sttion floos etween 00-0 A. Input Determintion Tehniques For this stuy, four ifferent pprohes input etermintion tehnique (ross orreltion-ac, stepwise regression-as, ross orreltion n stepwise regression-acs, geneti lgorithm-ag), supervise seletion-asp n ll input vriles- were explore to inite ifferene in input vrile seletion. Stepwise regression, whih is multiple regression metho, removes less orreltion vrile AC, n is metho use to lulte reltionship etween input vriles, for this stuy only those input vriles with orreltion greter thn 0. were selete. AS is multiple regression metho, y whih less orreltion vriles were remove n input vriles tht remine fter removl were selete. ACS selete input vriles from those vriles with orreltion greter thn 0., AG is se on iologil evolution n nturl seletion n ws evelope y Holln [1] n ASp selete only input vrile time t eh sttion. The ross orreltion n stepwise regression were lulte from SPSS stwre n The WEKA 1

3 Interntionl Journl Computer n Eletril Engineering, Vol., No., April 01 stwre ws use to lulte geneti lgorithm. Three input vriles were use for eh sttion, whih is vrile time t, time step k hours (t-) n hours (t-). The totl numer input vriles ws 1. Tle I presents selete input vriles for eh tehnique. AC, AS n ASC, whih i not selete vriles from M.1 showe similr input vriles, lso ross orreltion n stepwise regression provie sme input vriles ( inputs). IV. RESULTS AND DISCUSSION The stuy results re presente seprtely orresponing to input etermintion tehniques s followe. A. Cross Correltion (AC) n Stepwise Regression (AS) Both tehniques; ross orreltion n stepwise regression, selete ll input vriles from ll sttions exept sttion M.1. It my e euse this sttion s istne to M., whih is longest istne, lso it is not lote t min river. The hyrogrph (Fig. ) shows tht forest results wter levels gree well with oserve t. The n moels show very similr performnes. Noneless, moel (re line) seems to e etter thn moel s with inresing hien noe, it ereses ury pek foresting (Fig. ), vlues slightly erese n errors rise up when numer hien noes inreses (Fig., ). TABLE I: INPUT VARIABLES Input Input Determintion Tehniques AC AS ACS AG ASp M. M._ M._ M. M._ M._ M.1 M.1_ M.1_ M.1 M.1_ LevelM.1_ (m) AC/AS M.1 M.1_ M.1_ Totl AC/AS Network Moel B. Artifiil Neurl Moels were evelope to ompre two lerning lgorithms; n. The numer hien noes for moel epene on numer input noes, whih vrie oring to five tehniques input etermintion. The hien noes were set from 1 to n+1[] (n ws numer input vriles), refore, numers hien noe AC, AS, ACS, AG, Asp n moels rnge from 1-, 1-1, 1-, 1- n 1-1, respetively. The result this stuy ws wter level t M t hours in vne. For ville tset, t in perio ws use for moel lerning, n t in yer 0 ws use for moel testing. The finl results were otine from verge 0 loop lultions. To ssess moel performnes, Pek Differene (1), Root Men Squre Error- n Coeffiient Effiieny (Nsh-Sutliffe effiieny)- were pplie [1]. = mx Q' i - mx (Qi) Fig.. Results moels with inputs from AC n AS Therefore, est hien noe is hien noe whih is inite s :1:1. The performnes n moels re similr ut tht moel (in whih 0.0 meters error () n vlue 0. were otine) gives etter pek wter level foresting thn moel. All sme vlues n moels seem to e similr s re exists ifferene only. B. Cross Correltion n Stepwise Regression (ACS) The omintion tehniques (ross orreltion n stepwise regression) require only two input vriles less thn ross orreltion. The etween two lerning lgorithms hs ifferene 0.0 n vlues n re 0.0 n 0. respetively. However, moel still provies etter result t pek wter level thn moel (0.0/0.01). In ition, oth moels hyrogrphs present goo results (Fig. ). The effet numer hien noes re still similr with AC s is etter thn with more numers hien noe result in poorer moel performnes (Fig.,, ). (1) ' where Q i is moele vlue t time i, n Qi is oserve vlue t time i. If result is positive vlue, it mens moel foresting is over tul pek, while, negtive vlue mens moel foresting is uner tul pek.

4 Interntionl Journl Computer n Eletril Engineering, Vol., No., April 01 ACS respetively. input vriles were This is euse only five selete n tht my not e enough informtion for ANN to lern n forest. Thisinite influene numer inputs on moel performne, tht is fewer numer input vriles les to reution in moel s pility. Fig. presents hyrogrphs 1investigtive ASp oth n lgorithms unerestimte with oservtion hyrogrph. Performnes moels with ifferent numers hien noes rop rmtilly t greter numers hien noes (Fig.,, ) Fig.. Results moels with inputs from ACS C. Geneti Algorithm (AG) -1.0 The moel with input selete from geneti lgorithm tehnique est performne ASp seems to provie for foresting wter level t pek. The moeling only 1 n results with n hve errors meters, whih is est sore for eh lerning lgorithms. On or hn, - (0.0/0.0) n (0./0.) vlues AG_ n AG_ re worst when ompre with or three input etermintion tehniques. The reson might e, first, AG Fig.. Results moels with inputs from ASpAll Inputs () hooses M.1_, whih hs less orreltion ut my hve gret effet on pek wter level M. guge sttion, All inputs onsiste 1 hien noes, whih ws n seon, it ignores M._ wheres or three mximum numer onsiere for this stuy. The tehniques o not ignore it (TleI) s M. guge hyrogrphs in Fig. present moe performne sttion is lote t min river, n so ury trining with n lgorithms using ll 1 input foresting wter level t M. my epen on it (Fig. 1). vriles, it seems to e goo results ut when looks t Agin, it n e seen tht performnes oth grphs t, n (Fig.,, ), re re moels re similr (Fig. ) n tht n moels lrge errors ourring when numer hien noes provie less ury t greter numers hien noes inreses. Moreover, etter results re otine when trining (Fig.,, ). with lgorithm. The prtie using ll input vriles gives etter results t pek thn or input AG 1.0 etermintion tehniques exept geneti lgorithms s 1.0 inlues n input vrile t M.1_. However, too 1 mny input vriles, whih some re unsuitle reue overll moel performne s / n 1 - / re 0./0., 0.00/0.0 respetively, t moel nees 1lso euse lrge size set -.00 more time prtiulrly for lerning proess D. Supervise Seletion (ASp) 1 Fig.. Results moels with inputs from AG ASp The ie ASp is to reue numer input vriles 1 1 to 1 1selet 1-1 trie s muh s possile, hene this tehnique only 1 - t t time t fivesttions, Unfortuntely, -1.0 foresting results oth n moels were -.00 n were poorest, in whih vlues, 0.1 meters lower thn pek, 0.1 n 0., Fig.. Results moels with inputs from 1 1

5 Interntionl Journl Computer n Eletril Engineering, Vol., No., April 01 [] V. CONCLUSION AND RECOMMENDATION To sum up, moel trining with n shows similr performnes in foresting wter level t M. sttion ut for pek wter levels, it is ovious tht foresting with lgorithm provie etter results thn with. However, mjor isvntge lerning lgorithm is tht it tkes long time to finish lerning proess, prtiulrly with lrger numers hien noes. Aitionlly, ANN moels forest wter levels t M. hours in vne from selete input vriles from six tehniques. The overll results re quite similr ut AG seems to e est tehnique for seleting input vrile for pek foresting. In ontrst, it is ovious tht insuffiient input vrile (ASp) oul le to worst foresting performne. These tehniques lso show tht M., M. n M.1 re importnt input vriles for floo foresting t M.. As for fining est numer hien noes for this stuy re, it n e pointe out tht only one hien noe is perfet numer so it seems esy or not omplex for foresting wter levels hours he t M.. Therefore, reommention for future stuy is to exten forest perio to more thn hours or to use testing moels with ifferent input etermintion tehniques for smll floo events. [] [] [] [] [] [] [1] [1] Twee Chipimonplin grute in B.S (soil siene), KhonKen University, KhonKen, Thiln in 1, M.Ap.S (Geosptil Informtion), RMIT University, Melourne, Austrli in 00, Certifite Proessing Rr Dt Erth Surfe Remote Sensing, NPO Mshinostroyeni, Mosow, Russin Feertion in 00, n Ph.D. (physil geogrphy), The University Lees, Lees, UK in 0. He is memer Climte n Environmentl Reserh Group (RG) n leturer t Deprtment Geogrphy, Ching Mi University, Thiln. During working t Ching Mi University, his trining experiene re GIS n Remote Sensing for Nturl Hzr n Risk Assessment t ITC, The Nerlns, CASITA finl workshop t IIRS, Ini n Integrtion Geo-Informtis for Environment n Disster Mngement t GISTDA, Thiln. His re interest is using GIS, Remote Sensing n Artifiil Neurl Network for nturl hzr; floo foresting n lnslie mpping ACKNOWLEDGMENT Thnks to Ntionl Reserh Counil Thiln for finnil support, Hyrology n Wter Mngement Center for Lower Norstern Region, Thiln, for wter level t n to Mrs. Phithy Chipimonplin for eiting stuy re mp. REFERENS [1] [] [] [] [] T. Chipimonplin, L. M. See, n P. E. Knele, Improving neurl network for floo foresting using rr t on Upper Ping River, in Pro. 1th Interntionl Congress on Moelling n Simultion, Perth, 0, pp. 0-. T. Chipimonplin, L. M. See, n P. E. Knele, Using rr t to exten le time neurl network foresting on River Ping, Disster Avnes, vol., no., pp. -, July 0. H. Yon, F. Antil, n V. Fortin, Compring sigmoi trnsfer funtion for neurl network multistep he stremflow foresting, Journl Hyrologi Engineering, vol. 1, no., pp. -, April 0. A. P. Piotrowski n J. J. Npiorkowski, Optimizing neurl network for river flow foresting-evolutionry Computtion methos versus Levenerg-Mrqurt pproh, Journl Hyrology, vol. 0, no. 1, pp. -, Sep 0. T. Chipimonplin, An explortion neurl network moelling options for upper River Ping, Ph.D. isserttion, Deprtment Geogrphy, The University Lees, UK, 0. M. H. Bele, M. T. Hgn, n H. B. Demuth, Neurl Network TooloxTM User s Guie, The Mth Works, In., Ntik, MA, 0, h, pp. :1-:. DWR, Intergrte wter resoures mngement pln for Mun Bsin: Exutive summry, Deprtment Wter Resoures, Ministry Nturl Resoures n Environment, 00. [Thi] J. Holln, Apttion in Ntul n Artifiil Systems, University Mihign Pres, Ann Aror, 1. C. W. Dwson. (Mrh 01). HyroTest: sttistil ssessment hyrologil forests. [Online]. Aville: S. Kure n T. Tekri, Hyrologil impt regionl limte hnge in Cho Phry River Bsin, Thiln, Hyrologil Reserh Letters, vol., pp. -, My 0. T. Tingsnhli n M. R. Gutm, Applition TANK, NAM, ARMA n neurl network moel to floo foresting, Hyrologil Proesses, vol. 1, no. 1, pp. -, Nov 000. S. Suphrti, Applition neurl network moel in estlishing stge-ishrge reltionship for til river, Hyrologil Proesses, vol. 1, no. 1, pp. 0-0, Aug 00. Hyrology Division. (August 0). Preition wter ishrge 1- ys in vnes in mjor thments y ANN, Royl Irrigtion Deprtment [Online]. Aville: (Thi). G. J. Bowen, H. R. Mier, n G. C. Dny, Input etermintion for neurl network moels in wter resoures pplition. Prt. Cse stuy: foresting slinity in river, Journl Hyrology, vol. 01, no. 1-, pp. -, Jn 00. Thveesk Vngpisl reeive B. Eng (ivil engineering) from KhonKen University, KhonKen, Thiln in 1, M. Eng.S. (wter engineering) from University New South Wles, Syney, Austrli in 1 n Ph.D. (ivil engineering) from Monsh University, Melourne, Austrli in 00. He is n Assistnt Pressor t Deprtment Civil Engineering, Fulty Engineering, Uon Rththni University, Thiln. His reserh interest overs smll wtershe mngement, wter qulity, grounwter hyrulis n groun wter ontmintion.