Simulation of rainfall-runoff process by artificial neural networks and HEC-HMS model (case study Zard river basin)

Transcription

1 Proceedgs of The Fourth Iteratoal Ira & Russa Coferece 43 Smulato of rafall-ruoff process by artfcal eural etworks ad HEC-HMS model (case study Zard rver bas Mehrdad Akbarpour MSc. Graguate, Water Structures Egeer, Shahd Bahoar Uversty, Kerma, Ira, Phoe: +98( , E-mal: Abstract The rafall-ruoff modelg s olear process accordg to the temporal ad spatal dstrbuto of the rafall, ad t s ot possble to expla the respose catchmet system wth the smple models. I the preset research smulato of the rafall-ruoff process were carred out by Artfcal Neural Networks (ANNs ad HEC-HMS model. The ANN models of, Mult Layer Perceptro (MLP wth two structures of oe ad two hdde layer, ad Radal Bass Fucto (RBF, was used for smulato of ths process. It has bee appled to the Zard rver bas Khuzesta provce usg daly rafall ad ruoff data, durg the perod of 99-. I ths perod, 4 flood evets were selected for smulato of the HEC-HMS model. The obtaed results of the above models were compared wth the observed data from Zard rver bas. Ths comparso shows that RBF model has much more power tha MLP ad HEC-HMS models for smulatg of the rafall-ruoff process Zard rver bas. Key Words: Artfcal eural etwork, HEC-HMS, Rafall-ruoff process Itroducto Smulato of rafall-ruoff process s very mportat water resources maagemet, rver egeerg, flood cotrol ad surface water ad groudwater utlzato. Due to exstece of varous hydrologc factors bas s respose s very complex to the rafall. Ruoff depeds o geomorphologc propertes of a bas such as geometry, vegetatve coverg, sol type ad clmate characterstcs such as rafall, temperature, etc. The effects of these factors are ot uform ruoff geerato. Up to ow may physcally used models have bee suggested to smulate ths process such as HEC-HMS model. These models have requred may catchmets characterstcs for smulato such as rafall depth, evapotrasprato, fltrato, tal losses, tme of cocetrato, etc. Recetly wth ew advacemet may world scetfc commutes become terested dfferet braches of artfcal tellgece, such as eural etworks. Wth the same logc hydrologsts also were persuaded to smulate the hydrologc volute processes by these techques. Artfcal eural etworks (ANNs Artfcal eural etworks (ANNs are smple model of huma s bra. A ANN s olear mathematcal structure that has ablty to show the olearty process for commucatg betwee puts ad outputs of ay system. Ths etwork s trag wth preset data learg process ad ca be used for future predcto. Geerally, each ANN s formed wth a umber of layers whch bult-up some euros. Neuros are the smallest ut of ANN costructor ad are lke to huma bra cells. Each etwork has bee formed from oe put layer, oe output layer ad oe or more hdde layer. Neuros of each layer are coected to ext layer by weghts. Durg the etwork trag process, weghts

2 Proceedgs of The Fourth Iteratoal Ira & Russa Coferece 44 ad values that called bas s frequetly chagg utl obectve fuctos wll reach to desre value. For trasferrg outputs of each layer to the others, actvato fuctos are used for olear amplfer to euros. The techque s adopted for access of weghts ad bases to deal values called Learg Rule. That almost t s complex mathematcal algorthm. Each ANN eeds two data sets for to create ad to acceptable: tra set ad test set. About % of data are used for tra set ad resdue for test set. Durg the trag process, etwork learg rate s regularly measured by obectve fuctos ad fally a etwork wll be acceptable whch has less value of error ad maxmal correlato coeffcet. The obectve fuctos that are more used tha others clude the root mea squared error (RMSE, the sum squared error (SSE ad the correlato coeffcet (R that equatos as follows: RMSE = SSE = R = = ( Q Qˆ = = Where = ( Q Qˆ ( Q Qˆ ( Q Q Q s the ; Qˆ s the forecastg data; Q s the average of observed data. Mult Layer Perceptro (MLP eural etwork The MLP eural etwork structure s show Fgure. I ths etwork, euros outputs of each layer have bee etrace to ext layer euros ad t s cotued to mport the etwork s output. Learg algorthms of MLP eural etwork are bass o back propagato algorthm (BP. Fgure. Mult layer perceptro eural etwork Learg process MLP eural etworks were doe three steps: forward pass, backward pass ad computato terace pass. The put value to the each euro s gve by: et m = w =. o

3 Proceedgs of The Fourth Iteratoal Ira & Russa Coferece 45 Where et s the put value of th euro th layer; w s the coecto weghts betwee th euro th layer ad th euro the (-th layer; o s the output of th euro the (-th layer; m s the umber of euros the (-th layer. The output of each euro wll be obtaed after to apply the actvato fucto. The commo actvato fucto for use back propagato algorthm s sgmod fucto. The output value of each euro was calculated by: Sg = + exp( et ( et Radal Bass Fucto (RBF eural etwork The RBF etwork structure s show Fgure. The ma dffereces betwee ths etwork ad MLP etwork are as follows: The RBF etwork has oe hdde layer ad actvato fuctos of euros are Gussa fucto wth partcular ceter ad spread. There are ot weghts betwee put layer ad hdde layer ad the dstace betwee each patter ad ceter vector of each euro hdde layer s used as a put of Gussa actvato fucto. I ths etwork, actvato fuctos of output euros are smple lear fuctos ad because of ths reaso we ca use of lear optmum algorthms. They have bee caused to mprove the processg rate ad prevet to fall local mmums that deal wth there at learg process MLP etwork. u, σ w u, σ u, σ Fgure. Radal bass fucto eural etwork Output of th hdde euro gve by: h exp Where X U U the ceter ad = σ Iput Hdde Output σ the spread of gussa fucto; X the th put vector.

4 Proceedgs of The Fourth Iteratoal Ira & Russa Coferece 46 HEC-HMS Program The HEC-HMS program was developed at the Hydrologc Egeerg Ceter (HEC of the US Army corps of egeers. Of ths program advatages s ablty to parameters optmzato. It used for smulato of rafall-ruoff process by losses, drect ruoff ad baseflow that each of those calculated by dfferet methods. I ths research has bee used of SCS curve umber to losses calculato, SCS ut hydrograph to ruoff ad expoetal recesso model to baseflow. Case study Case study ths research s Zard rver bas Khuzesta provce southwest of Ira. The area of the bas s 875 Km, the legth of ma chael s about 7 Km ad average slope s about 3%. There are sx precptato gages ad oe hydrometrc stato the bas, ad ther locatos are show Fgure 3. Rafall ad ruoff data were used for perod of 99-. Fgure 3. Zard rver bas Results ad dscusso The target of ths research was estmato of rver dscharge a umber of evets. For trag of MLP ad RBF, etworks were cosdered the rafall data for put ad dscharge data for target. Also, t was used two dfferet structures by oe ad two hdde layer for MLP etwork. All of the smulato stages of etworks were used MATLAB software. Results are show below:

5 Proceedgs of The Fourth Iteratoal Ira & Russa Coferece 47 Dscharge MLP layer Fgure 4. Comparso betwee observed data ad MLP oe hdde layer estmato data Dscharge MLP layers Fgure 5. Comparso betwee observed data ad MLP two hdde layer estmato data Dscharge RBF Fgure 6. Comparso betwee observed data ad RBF estmato data

6 Proceedgs of The Fourth Iteratoal Ira & Russa Coferece 48 The results showed that the RBF etwork has a better ablty tha MLP etwork. The 4 flood evets were smulated ad parameters were optmzed wth use of HEC-HMS program. For compared betwee MLP ad RBF etworks ad HEC-HMS model oe flood evets that were t used to prevous smulatos as show: Dscharge MLP layer MLP layer RBF HMS 95// 95// 95//3 95//4 95//5 Tme Fgure 7. Comparso betwee MLP, RBF ad HEC-HMS Cocluso I geeral ANN s a techque, whch solve the olear ad complex ature of catchmet s system requred the log tme ad accurate seres of put, ad output data ad f o-adequate data s gve, the accurate results wll be obtaed. Both ANN ad HEC-HMS models have ther ow advatages. The HEC-HMS model f calbrated for a bas the ca be used to estmate the flood dscharges ugaged catchmet. The advatage of ANN model s that the predcto s based o data prevous tme terval. Ackowledgemet The author wshes to thak the water egeerg researches departmet of Khuzesta Water ad Power Agecy (KWPA for to support of ths research. Refereces Dawso CW, Wlby RL ( Hydrologcal modelg usg artfcal eural etworks. Progress Physcal Geography 5(: -8 Hsu K, Gupta HV, Soroosha S, (995 Artfcal eural etwork modelg of the rafall-ruoff process. Water Resources Research (3: Jayawardea AW, Achela D, Ferado K, (998 Use of radal bass fucto type artfcal eural etworks for ruoff smulato. Computer-Aded Cvl ad Ifrastructure Egeerg 3: 9-99 Shamseld AY, O Coor KM, Lag GC, (997 Methods for combg the outputs of dfferet rafall-ruoff models. Joural of Hydrology 97: 3-9 Imre CE, Duruca S, Korre A, ( Rver flow predcto usg artfcal eural etworks: geeralzato beyod the calbrato rage. Joural of Hydrology 33: Sakumar N, Thadaveswara BS, (999 A No-lear rafall-ruoff model usg a artfcal eural etwork. Joural of Hydrology 6: 3-55