Available online at www.sciencedirect.corn MAT HE MAT I C AL
AND 8 C I n N C E ~ D I FI E! (~'11" " COMPUTER MODELLING Mathematical and Computer Modelling 40 (2004) 839-846 www.elsevier.com/locate/mcm Ra i n f a l l - Ru n o f f Mo d e l Us i n g an Ar t i f i c i al Ne u r a l Ne t wo r k Ap p r o a c h S. R I AD AND J . MANI A Universit4 des Sciences et Technologies de Li l l e- USTL- LML- EPUL. UMR CNRS 8107 Ecole Pol yt echni que Uni versi t ai re de Lille, D6par t ement de g~otechnique et g6nie civil Cit~ scientifique, Avenue Paul Langevi n, 59655 Villeneuve d' As cq Cedex, France <souad. riad><j acky. mania> polyt ech-lille, fr rsouad_2OO4yahoo, fr L. BOUCHAOU Universit6 I bn Zohr, Facult~ des sciences, D6par t ement de G6ologie Labor at oi r e de G~ologie Appliqu6e et G~oenvi ronnement ( G. A. G. E) Equi pe d' Hydrog~ol ogi e, BP 28/ S, 80000 Agadi r, Mar oc Bouchaoul car a~ai l . com Y . N A J J A R K a n s a s S t a t e U n i v e r s i t y , D e p a r t m e n t o f C i v i l E n g i n e e r i n g M a n h a t t a n , K S 6 6 5 0 5 , U . S . A . ea4146ksu, edu (Received March 2003; revised and accepted October 2003) Ab s t r a c t - - T h e use of artificial neural networks (ANNs) is becoming increasingly common in the analysis of hydrology and water resources problems. In this research, an ANN was developed and used to model the rainfall-runoff relationship, in a catchment located in a semiarid climate in Morocco. The multilayer perceptron (MLP) neural network was chosen for use in the current study. The results and comparative study indicate t hat the artificial neural network method is more suitable to predict river runoff t han classical regression model. (~) 2004 Elsevier Ltd. All rights reserved. Ke ywor ds - - Ra i nf a l l - r unof f , Catchment, Semiarid climate, MLP, Modelling, Artificial neural net- work, Multiple regression, Morocco. 1. I NTRODUCTI ON Th e ANNs mode l s ar e power f ul pr e di c t i on t ool s f or t he r el at i on be t we e n r ai nf al l a nd r unof f pa r a me t e r s . Th e r es ul t s will s u p p o r t deci si on ma k i n g i n t he a r e a o f wa t e r r es our ces p l a n n i n g a nd ma n a g e me n t . Besi des, t h e y assi st u r b a n pl a nne r s a n d ma n a g e r s u n d e r t a k e t h e ne c e s s a r y me a s ur e s t o face t h e b a d pr edi ct i ons . Thus , t h e y hel p avoi d l osses i n publ i c a nd pr i va t e pr oper t i es , a nd he a l t h a n d ecol ogi cal h a z a r d s t h a t ar e l i kel y t o oc c ur due t o f l oodi ng. Mor eover , t he ANN mode l s have been used i ncr eas i ngl y i n var i ous a s pe c t s of sci ence a nd en- gi neer i ng be c a us e of i t s abi l i t y t o mode l b o t h l i near a nd nonl i ne a r s y s t e ms wi t h o u t t h e need t o ma k e a n y a s s u mp t i o n s as ar e i mpl i ci t i n mo s t t r a di t i ona l s t a t i s t i c a l a ppr oa c he s . I n s ome o f t h e hydr ol ogi c pr obl e ms , ANNs have a l r e a dy be e n successf ul l y us ed for r i ver fl ow pr e di c t i on [1-8], 0895-7177/04/$ - see front matter (~) 2004 Elsevier Ltd. All rights reserved. Typeset by .Ah/tS-TE X doi:10.1016/j.mcm.2004.10.012 840 S. RIAD et al. for rainfall-runoff process [9-11], for the prediction of water quality parameters [12] and for char- acterization of soil pollution [13]. In addition, ANNs are applied for prediction of evaporation [14], for rainfall-runoff forecasting [15-19], for prediction of flood disaster [20], and for river flow time series prediction [21]. In these hydrological applications, a multilayer feed-forward backpropagation algorithm is used [22]. It usually is composed of a large number of interconnected nodes, arranged in an input layer, an output layer, and one or more hidden layers. The transfer function selected for the network was the sigmoid function. The aim of this paper is to model the rainfall-runoff relationship in the Ourika catchment located in semiarid climate in Morocco using a black box type model based on ANN methodology. 2. THE ARTI FI CI AL NEURAL NETWORKS APPROACH 2.1. Over vi ew of ANN The ANN technology is an alternate computational approach inspired by studies of the brain and nervous systems [23]. It is based on theories of the massive interconnection and parallel processing architecture of biological neural systems. The main theme of ANN research focuses on modelling of the brain as a parallel computational device for various computational tasks that were performed poorly by traditional serial computers. ANNs have a number of interconnected processing elements (PEs) that usually operate in parallel and are configured in regular architectures. The collective behavior of ANN, like a human brain, demonstrates the ability to learn, recall, and generalize from training patterns or data. The advantage of neural networks is they are capable of modelling linear and nonlinear systems. In the present study, we use an MLP trained with a backpropagation algorithm to predict the drainage basin runoff. The MLP consists of an input layer consisting of node(s) representing various input variable(s), the hidden layer consisting of many hidden nodes, and an output layer consisting of output variable(s). The input nodes pass on the input signal values to the nodes in the hidden layer unprocessed. The values are distributed to all the nodes in the hidden layer depending on the connection weights W~j and Wjk [24-26] between the input node and the hidden nodes. Connection weights are the interconnecting links between the neurons in successive layers. Each neuron in a certain layer is connected to every single neuron in the next layer by links having an appropriate and an adjustable connection weight. Weights A Wi ~ We i g h t s o t i 6 , Pt 1 ~ Input layer Hidden layer Output layer Input Neuron Output Xt ~ ) ~ X. S Figure 1. Architecture of t he neural network model used in this study. Rai nfal l -Runoff Model 841 The ar chi t ect ur e of t he neural net work used in t hi s s t udy and t he schemat i c represent at i on of a neuron are shown in Fi gure 1. Each node j receives i ncomi ng signals from every node i in t he previous layer. Associ at ed wi t h each i ncomi ng signal (Xi) is a weight (Wij). The effective i ncomi ng signal (Sj ) t o node j is t he weighted sum of all t he i ncomi ng signals and bj is t he neuron t hreshol d value. n s j = x i w q + b . (1) i =l The effective i ncomi ng signal, Sy, is passed t hr ough a nonl i near act i vat i on funct i on t o produce t he out goi ng signal (yj) of t he node. The most commonl y used in t hi s t ype of net works is t he logistic sigmoid funct i on. Thi s t ransfer funct i on is cont i nuousl y differentiable, monot oni c, symmet ri c, bounded bet ween 0 and 1 [27]. It is expressed mat hemat i cal l y as: 1 f (Sj) - 1 + e - s j ( 2 ) 2.].. ANN P e r f o r ma n c e In t hi s st udy, bot h st at i st i cal and graphi cal criteria were adopt ed t o select t he desired opt i mal net work model. The st at i st i cal cri t eri a consist of average squared of error (ASE), coefficient of det er mi nat i on (R 2) and t he mean absol ut e relative error (MARE). They are given by N 2 ASE = 4=1 Y ' (3) N( Q ~ ) 2 E t~ - Qt~ i =l R 2 = 1 - N , ( 4 ) E ( Qt i - Qt 0 2 i=1 N MARE = ~=1 x l o o , ( 5 ) N where Qti and Qti are respectively, t he act ual and predi ct ed value of flow (normalized between 0 and 1), (~ti is t he mean of Qti values and N is t he t ot al number of dat a sets. The R 2 st at i st i c measures t he linear correlation between t he act ual and predi ct ed flows values. The ASE and MARE st at i st i c measures are used t o quant i f y t he error bet ween observed and predi ct ed values. The opt i mal value for R 2 is equal t o 1.0 and for ASE and MARE is equal t o 0.0. The graphi cal performance i ndi cat or gives bet t er results when t he da t a pairs are closing t o 45 line and t he good superposi t i on bet ween t he desired and cal cul at ed flow values in t he t rai ni ng and t est i ng phases. For t he da t a set considered in t he present st udy, t he i nput variables as well as t he t ar get variables are first normal i zed l i nearl y in t he range of 0 and 1. Thi s range is selected because of t he use of t he logistic funct i on (which is bounded between 0.0 and 1.0) as t he act i vat i on funct i on for t he out put layer, i.e., equat i on (2). The normal i zat i on is done using t he following equation. .,~ = X -- Xmi n . Xma x -- Xmi n' (6) where . ~ is t he st andar di zed value of t he i nput , Xmi n and Xmax are respectively, t he mi ni mum and maxi mum of t he act ual values, in all observations and X is t he original dat a set. The mai n reason for st andar di zi ng t he dat a mat r i x is t hat t he variables are usual l y measured in different uni t s. By st andar di zi ng t he variables and recasting t hem in dimensionless uni t s, t he ar bi t r ar y effect of si mi l ari t y bet ween obj ect s is removed. 842 S. RIAD et al. 3. T H E S T U D Y C A T C H ME N T A N D D A T A B A S E In t he present st udy, t he flow and rainfall series observed in Ouri ka basin at Aghbal ou st at i on in Morocco is anal yzed usi ng t he ANN model. The Ouri ka basi n is t he most i mpor t ant subcat chment of Tensift basin drai nage l ocat ed in semi-arid region of Marrakech, whi ch is drai ni ng an area of about 503 km 2. (See Fi gure 2.) Tensift drainage basin \ t ) / / / j / t . _. ~o CIH(~[AOUA / /I 0 1 0 2 0 3 0 K m I I I I k \ \ , t / / _, _%/ / ~ Ga u g e s t at i on Fi gure 2. Locat i on of Ouri ka Wadi in t he Tensift basin. The Rai nfal l and Runoff dai l y dat a at t he Aghbal ou st at i on was used for model investigation. The dat a cont ai ns i nformat i on for a peri od of seven years (1990 t o 1996). The ent i re dat abase is represent ed by 2550 dai l y values of rainfall and runoff pairs. The ANN model was t r ai ned using t he resul t i ng runoff and rainfall dai l y dat a. The dat abase was collected by t he Rabat hydraul i c admi ni st r at i on. The i nput vect or is represent ed by rainfall and runoff values for t he preceding seven days, (i.e., t - 1, t - 2, t - 3, t - 4, t - 5, t - 6, t - 7) as well as t he rainfall value expect ed for day t. Accordingly, t he out put vect or represents t he expect ed runoff value for day t (Qt). 4. R E S U L T S A N D D I S C U S S I O N S The dat abase compiled represents seven years dai l y sets of rai nfai l -runoff values for t he Ouri ka Wadi basin. In t hi s paper, we used t he dat a for t he l ast year (1996) for model t est i ng, while t he ot her remai ni ng da t a (1990 t o 1995) was used for model t r ai ni ng/ cal i br at i on. The t r ai ni ng phase of ANN model was t er mi nat ed when t he average squared error (ASE) on t he t est i ng dat abases was mi ni mal . The goal of t he t r ai ni ng process is t o reach an opt i mal sol ut i on based on some performance measur ement s such as ASE, coefficient of det er mi nat i on known as R-square value (R2), and t he MARE. Therefore, requi red ANN model was developed in two phases: t r ai ni ng (cal i brat i on) phase, and t est i ng (generalization or val i dat i on) phase. In t he t r ai ni ng phase, a larger part for dat abase (six years) was used t o t r ai n t he net work and t he remai ni ng par t of t he dat abase (one year) is used in t he t est i ng phase. Test i ng sets are Rai nfal l -Runoff Model 843 Tabl e 1. St at i st i cal accuracy measures of t hi s net work model at t est i ng and t r ai ni ng phases. ASE R 2 MARE Trai ni ng Phase 0.000076 0.948 1.029% Test i ng Phase 0.000007 0.917 1.524% 180 160 120 ~ 8o ,- 60 40 20 0 180 160 . , ~ 14o ~ 120 ,~ 100 ~ 60 40 T r a in in g p h a i ~ , , i t . / . Equa l i t y l i ne , p . ~ , ~ . . - , .~. 5 ~ - . . r . / ' . ~ i l r a " . : " IR-- 0,948 u . o , i , , , , , , , 0 2 0 4 0 6 0 80 I00 120 140 160 180 Actual flow (l~/S) ~ r a i n i n g p h a s e l . . . . . . . Predicted flow Actual flow Act ual fl ow (m3/ s) Fi gur e 3. Compari son bet ween t he act ual and ANN predi ct ed flow values. 30 ~ 25 20 q~ 15 ~ ~ lO [ T e s t l n g p h a s e l . I s e ~ . j / Equa l i t y l i ne le ~ m I I ) mm ~ ' * [R=0,917 I I l I I I 5 I0 15 2 0 2 5 3 0 A c tu a l f l o w (m 3 /s) 844 S. RIAD et aL 30 ' 25 20 ' ~ 15 I T e s t i n g p h a s e ] . . . . . . . Predicted flow Actual flow I i ,..: ( ~'. b Table 2. testing phases. (a) Training Phase A c t u a l f l o w ( m ~ / s ) Figure 3. (cont.) Statistical parameters of the predicted and actual flow at training and Statistical Actual Flow Predicted Flow Parameters (ma/s) (ma/s) Average 9.79 9.70 Standard of 18.93 17.81 Deviation Minimum 0 1.37 Maximum 184 187.17 Coefficient of 1.93 1.84 Variation (b) Testing Phase Statistical Actual Flow Predicted Flow Parameters (ma/s) ( ma/ s ) Average 2,71 3.25 Standard of 4.33 3.96 Deviation Minimum 0.10 1.38 Maximum 31.90 31.16 Coefficient of 1.60 1.22 Variation usual l y used t o select t he best per f or mi ng net work model . I n t hi s research, t he ANN was opt i mal at 600 i t er at i ons wi t h 12 hi dden nodes. The cor r espondi ng accur acy measur es of t hi s net wor k model on t est i ng and t r ai ni ng dat a are gi ven i n t he following t abl e (Tabl e 1). General l y, accuracy measur es on t r ai ni ng dat a are bet t er t ha n t hose on t est i ng dat a. 180 160 " . , ~ 140 120 l o o . ~. . . 8 0 ..,_ ~ ***.;.,._ ~ . . 2 * . * * ~ * ~ - 60 . . . * X ~ . . * * 1 4 0 1 - * * R * 0 . , 1 1 * 20 .'* ~. I R = 0 , 9 2 4 F - - - I I I I I I I ! 0 20 40 60 80 100 120 140 160 180 Actual flow (m~/s) Figure 4. Comparison between the a~tual and predicted flow values by multiple linear regression (MLR). Rainfall-Runoff Model 845 3O ,,,~, 25 ~ 2o 15 5 E q u a l i t y l i n e
0 e . e ~ O . . o~r " ' 0 ~ [ R= 0,888 ! u a u , , 0 5 10 15 20 25 30 Actual flow (m3/s) Figure 4. (cont.) Table 3. Comparison of correlation coefficients between actual and predicted flow by ANN and MLR models. ANN t MLR R2 Training Phase 0.948 I 0.924 Testing Phase 0. 917 i 0.888 The comparison between the predicted and actual flow values at training and testing phases show excellent agreement with the R 2 are respectively 0,948 and 0,917 (see Figure 3). Note that, data pairs closer to the 45 line represent better prediction cases. The good performance and convergence of the model are illustrated in Figure 3. The statistical parameters of the predicted and actual values of flow for the entire database are practically identical (see Table 2). ]In order to evaluate the performance of the ANN, the multiple linear regression (MLR) tech- nique was applied with the same data sets used in the ANN model. Figure 4 shows the compara- tive results obtained by MLR technique. The R 2 values for MLR and ANN models are presented in Table 3. Apparently, the ANN approach gives much better prediction t han the traditional method (MLR). 5. CONCLUSI ON The art i fi ci al neur al net wor k ( ANN) model s show good capabi l i t y t o model hydrol ogi cal pro- cess. Th e y are useful and powerful t ool s t o handl e compl ex pr obl ems c ompa r e d wi t h t he ot her t r adi t i onal model s. I n t hi s st udy, t he resul t s obt ai ned show cl earl y t ha t t he art i fi ci al neur al net - works are capabl e of model r ai nf al l - r unof f r el at i onshi p in t he ari d and semi ar i d regi ons in whi ch t he rai nfal l and r unof f are ver y i rregul ar, t hus, conf i r mi ng t he general e nha nc e me nt achi eved by usi ng neur al net wor ks in ma n y ot her hydrol ogi cal fields. The r esul t s and c ompa r a t i ve s t udy i ndi cat e t h a t t he art i fi ci al neur al net wor k me t hod is mor e sui t abl e t o pr edi ct ri ver r unof f t ha n classical r egr essi on model . The ANN appr oach coul d pr ovi de a ver y useful and accur at e t ool t o solve pr obl ems in wat er resources st udi es and ma na ge me nt . REFERENCES 1. A.H. Halff, H.M. Halff and M. Azmoodeh, Predicting runoff from rainfall using neural networks, In Engineer- ing Hydrology, Proceedings of the Symposium sponsored by the Hydraulics Division of ASCE, San Francisco, CA, July 25-30 (Edited by Kuo Cy), pp. 760-765, New York, (1993). 2. N. Karunanithi, W.J. Grenney, D. Whitley and K. Bovee, Neural networks for river flow prediction, Journal of Computing in Civil Engineering 8 (2), 201-220, (1994). 846 S. RIAD et al. 3. S. Lek, I. Di mopoul os, M. Derraz and Y. El Ghacht oul , Mod@lisation de la rel at i on pluie-d@bit ~ l ' ai de des r~seaux de neurones artificiels, Rev. Sei. Eau 9 (3), 319-331, (1996). 4. C.E. Imrie, S. Dur ucan and A. Korre, Ri ver flow predi ct i on using artificial neural networks: General i sat i on beyond t he cal i br at i on range, Journal of Hydrology 233 (1-4), 138-153, (2000). 5. T. S. Hu, K.C. Lam and S. T. Ng, Ri ver flow t i me series predi ct i on wi t h a r ange- dependent neural network, Hydrological Science Journal 46 (5), 729-745, (2001). 6. B. Si vakumar, A. W. Jayawar dena and T. M. K. G. Fernando, Ri ver flow forecasting: Use of phase-space recon- st ruct i on and artificial neural networks approaches, Journal of Hydrology 265 (1-4), 225-245, (2002). 7. N.G. Wri ght , M. T. Dast orani , P. Goodwi n and C. W. Slaughter, A combi nat i on of neural networks and hydr odynami c models for river flow predi ct i on, In Proceedings of the Fifth International Conference on Hydroinformatics, Cardiff, U. K. , (2002). 8. S. Riad, J. Mania, L. Bouchaou and Y. Naj j ar, Pr edi ct i ng cat chment flow in semi ari d region vi a artificial neural net work t echni que, Hydrological Process (t o appear). 9. K.L. Hsu, H.V. Gupt a and S. Sorooshian, Artificial neural net work model i ng of t he rainfall-runoff process, Water Resources Research 31 (10), 2517-2530, (1995). 10. J. Smi t h and R. N. Eli, Neural -net work models of rainfall-runoff process, Journal of Water Resources Planning and Management 121 (6), 499-508, (1995). 11. F. J. Chang and J. P. Suen, A st udy of t he artificial neural net work for rainfall-runoff process, Journal of Chinese Agricultural Engineering (In Chinese) 43 (1), 9-25, (1997). 12. H.R. Mai er and G.C. Dandy, The use of artificial neural net work for t he predi ct i on of qual i t y wat er param- eters, Water Resources Research 32 (4), 1013-1022, (1996). 13. A. Dan, J. Oost er baan and P. J amet , Cont r i but i on des r~seaux de neurones artificiels (RNA) ~ la car- act ~ri sat i on des pol l ut i ons de sol. Exempl es des pol l ut i ons en hydrocarbures ar omat i ques polycycliques ( HAP) , C. R. Geoscience 334, 957-965, (2002). 14. K.P. Sudheer, A. K. Gosain, D. Mohana Rangan and S.M. Saheb, Model l i ng evapor at i on using an artificial neural net work al gori t hm, Hydrol. Process. l f i , 3189-3202, (2002). 15. A. W. Mi nns and M. J. Hall, Artificial neural networks as rainfall-runoff models, Hydrological Sciences Journal 41 (3), 399-418, (1996). 16. C. W. Dawson and R.L. Wi l by, An artificial neural net work approach to rai nfal l -runoff modelling, Hydrological Sciences Journal 48 (1), 47-66, (1998). 17. C. W. Dawson and R.L. Wi l by, A compari son of artificial neural net works used for rainfall-runoff modelling, Hydrology and Earth Systems Sciences 3, 529-540, (2000). 18. C.L. Kin, J. E. Ball and A. Sharma, An appl i cat i on of artificial neural networks for rainfall forecasting, Mathl. Comput. Modelling 33 (6/ 7), 683-693, (2001). 19. K. W. Kang, C.Y. Par k and J.H. Kim, Neural net work and its appl i cat i on t o rai nfal l -runoff forecasting, Korean Journal of Hydroscienee 4, 1-9, (1993). 20. Y. Wei, W. Xu, Y. Fan and H. T. Tasi, Artificial neural net work based pr edi ct i ve met hod for flood disaster, Computers ~ Industrial Engineering 42, 383-390, (2002). 21. T. S. Hu, K. C. Lam and S.T. Ng, Ri ver flow t i me series predi ct i on wi t h a range-dependent neural network, Hydrological Science Journal 46 (5), 729-745, (2001). 22. R. P. Li ppmann, An i nt roduct i on t o comput i ng wi t h neural nets, IEEE ASSP Magazine, 4-22, (1987). 23. W. S. McCulloch, and W. Pi t t s, A logical calculus of t he ideas i mmi nent in nervous act i vi t y, Bulletin of Mathematical Biophysics 5, 115-133, (1943). 24. Y. Naj j ar and H. Ali, On t he use of BPNN in liquefaction pot ent i al assessment tasks, In Artificial Intelligence and Mathematical Methods in Pavement and Geomechanical Systems, ( Edi t ed by At t oh- Oki ne) , pp. 55-63, (1998). 25. Y. Naj j ar and H. Ali, CPT- bas ed l i quefact i on pot ent i al assessment: A neuronet approach, In ASCE Geotech- nical Special Publication Number 75, ( Edi t ed by P. Dakoulas et al.), pp. 542-553, (1998). 26. Y. Naj j ar and X. Zhang, Char act er i zi ng t he 3D st ress-st rai n behavi or of sandy soils: A neuro-mechani st i c approach, In ASCE Geotechnical Special Publication Number 96, ( Edi t ed by G. Filz and D. Griffiths), pp. 43- 57, (2000). 27. R. Hecht-Nielsen, Sol ut i on for a di st r i but ed hydrological model and appl i cat i ons, In Neurocomputing, pp. 89- 93, Addison-Wesley, Reading, MA, (1990).