ÇÚËÕÌ ÛÈÒÏÉÇÚÏÕÔg ÛÍÛÙÚ;@e;9 ... · neural networks such multilayer perceptron...
TRANSCRIPT
USINGAMODELHYBRIDBASEDONANN‐MLPANDTHESPIINDEXFORDROUGHTPREDICTIONCASEOFINAOUENBASIN(NORTHERNMOROCCO)
1BOUDADB.,2SAHBIH.,3MANSSOURII.,1MANSSOURIT.
1Researchstudent,DepartmentofGeology,FacultyofSciencesMoulayIsmailUniversity,Meknes,Morocco,Email:[email protected]:[email protected]
2Dean,DepartmentofGeology,FacultyofSciencesMoulayIsmailUniversity,Meknes,Morocco,Email:[email protected]
3Professor,LaboratoryofMechanics,MechatronicsandControl,ENSAMMoulayIsmailUniversity,Meknes,Morocco,Email:[email protected]
ABSTRACT
Thispaperdescribesanapproach forpredictingdroughts inthebasin InaouenbyusingahybridmodelbasedonartificialneuralnetworkssuchMultilayerPerceptron(ANN‐MLP)andSPIindex(StandardizedPrecipitationIndex).
Duringthefirststep,thecalculationoftheSPIindexhasbeentaken.Thisisachievedbyadjustingthefrequencydistributionmonthlyrecordsprecipitation,toaprobabilitydensityfunction.
Inasecondstep, threemodelsofANNMLPwereconstructedusing, for inputsadatasetcontaining theSPIcalculated, thevaluesofmonthlyprecipitationandintroducingalsotheNAOindextoestimatetheeffectoftheNorthAtlanticOscillationonthedroughtintheregion.
The performance of the neural predictionmodel integrating the three variables as inputs (ANN‐MLP 3) are showing fargreaterthanthoseestablishedbyothermodelsconsideringonlyprecipitationand/orSPI.Thisoptimalmodel(ANN‐MLP3)wasapplied to thepredictionofdrought in the regionusingSPI3,SPI6,SPI9,SPI12andSPI24.TheseSPIvalueswerepredictedforonemonthahead.ThemodelshowsverygoodprecisionandwhichbecomegreaterwhenwemovefromSPI3toSPI24.
IndexTerms:PredictionofDrought,ArtificialNeuralNetwork,StandardizedPrecipitationIndex,modelhybrid
1. INTRODUCTION
Droughtisoneofthemajorphenomenaarisingfromthevariability and climate change in recent decades. Thewordmeteorologicaldroughtrefers toastateofariditycaused by lower than normal rainfall for an extendedperiod.Indirecteffectsofsuchmeteorologicaldroughtintime can have economic, agricultural, hydrological andsocial impacts.Tomitigate these effects, it is importantand interesting to have the ability to predict suchextremeweatherevents.
The Standardized Precipitation Index (SPI) originallyproposedbyMcKeeetal.(1993)[1]isusedasanindexofdroughtmonitoringby themajorityofrecentmodelsof drought prediction. Its use to predict droughts hasincreased the ability to compare results obtained indifferentregionsoftheworld.
In the literature, thereareseveralworksapplied to thepredictionprocess,whicharecitedasanexamplesomeworkbasedonneuralnetworks:
Mansouri et al. (2014) [2] initially develop a modelbased on neural of MLP (Multi‐Layer Perceptron) forpredictingindicatorsofgroundwaterqualityofthewaterSoussMassaMorocco.The choiceof the architectureoftheartificialneuralnetworkofMLP type isdetermined
bytheuseofdifferentstatisticaltestsofrobustnessandLevenbergMarquardtalgorithmsusedtofixweightsandbiases existing between the different layers of neuralnetwork. Secondly, a comparative study was madebetween theneuralpredictionmodelPMCtypeandtheclassical statistical models, namely, the total multiplelinear regression. The output of the neural modelexceedsbyfartheclassicalstatisticalmodels.
Mansouri et al. (2013) [3] give twomodelingmethodsused for thepredictionofmeteorologicalparameters ingeneral, and the humidity in particular. Initially, themethods are based on the study of artificial neuralnetworks type MLP applied for the prediction ofmoisture, region of Chefchaouen in Morocco. Secondlythe new architecture of neural networks type MLPproposed was compared with that of the model ofmultiple linear regression (MLR). The modelsestablished by the MLP neural network predictivemodelsaremoreefficientcomparedtothoseestablishedby multiple linear regression, the fact that goodcorrelation was obtained by the parameters from aneuralapproachwithaMeanSquareErrorof5%.
El‐ Shafie et al. (2011) [4] developed two models forpredicting annual and monthly precipitation forAlexandria in Egypt. The first is basedonANNand thesecond on themultiple linear regressionmodel (MLR).
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1301
The MLR model showed a modest performanceprediction. The linear character of estimators MLRmodelmakes it inadequate toprovidegoodpredictionsfor a highly nonlinear variable. On the other hand, theANN model developed as a tool for non‐linearrelationship, which is potentially more suitable forpredicting precipitation (nonlinear physical variable),hasbeensuccessfulprediction.
Somvanshi et al. (2006) [5] have applied twofundamentally different approaches, the statisticalmethodbasedon theAutoregressive IntegratedMovingAverage (ARIMA) and new computational techniquesbased on RNA. The approaches ANN and ARIMA areapplied to the average annual rainfall data to calculaterespectivelytheweightsandregressioncoefficients.ThestudyshowsthattheperformanceofANNmodelismoreappropriatetopredictprecipitationcomparedtoARIMAmodel.
Inthiscontext,ourmainobjectiveinthisworkistobuilda hybrid model based on ANN Multilayer Perceptrontype and SPI developed byMcKee for the prediction ofdrought at different time scales on the basis of dataselectedstationslocatedwithinthewatershedInaouen.
2. APPLIEDMETHODS
2.1 The Standardized Precipitation Index(SPI)
The Standardized Precipitation Index (SPI) wasdevelopedbyMcKeeetal.(1993)[1]asawaytodefineandmonitordroughtevents. It isa simple index,basedsolely on rainfall data, and allows equally checking thewet periods / cycle’s and the dry periods / cycles. TheSPI compares precipitation over a certain period(normally 1‐24 months) to the average long‐termprecipitationobservedatthesamesite.[6]
TheSPIwasdesignedtoquantifytheprecipitationdeficitfor multiple time scales. These time scales reflect theimpactofdroughtontheavailabilityofdifferenttypesofwater resources. Ground moisture reacts relativelyquickly to rainfall anomalies in the short term, whilegroundwater, stream flowandwater volumes stored inreservoirs are sensitive to rainfall anomalies in thelongerterm.
TheSPIisdeterminedbyanormalizationofrainfallforagiven station, afterwhich there is a probability densityadjusted. The distribution that best represents theevolutionof sets of rain is theGammadistribution ([7]and [8]). This Gamma distribution is defined by itsprobabilitydensityrepresented,forx>0,by:
1Γ
Where:
αistheshapeparameter βisthescaleparameter xistheheightofthemonthlyprecipitation Γ represents the Gamma mathematical functiondefinedas:
Γ∞
The adjustment of the gamma distribution to datarequires therefore to determine α and β. They can beestimatedinthisway[7]:
α 1 1 ,βαwithA ln x
Σ
Where N is the size of the sample and is the mean ofobservations.
Theequationfortheprobabilitydensityfunctioncanbeintegrated to provide the cumulative function F(X) ofprobabilityforx>0:
F X f x dx1
βαΓ αxα e β dx
Butwenotethat,forthisdefinition,x>0,howeverthereare periods where x = 0. To take into account of theprobability of zero, the function of the cumulativeprobabilitydistributionofgammaismodifiedasfollows:
H X q 1 q F X
With q, the probability of each station to have a zeroprecipitationovertheentireperiod.
IfmisthenumberofzerosinaseriesofprecipitationofNvariable,qcanbeestimatedby[6]:
qmN
We therefore obtain for each x, a corresponding valueH(x). This finally allows to calculate the SPI, as follows[9]:
SPI Wc c W c W
1 d W d W d W
withW ln for0 0.5
SPI Wc c W c W
1 d W d W d W
withW ln for0.5 1
The constants are: c0 = 2.515517, c1 = 0.802853, c2 =0.010328,d1=1.432788,d2=0.189269,d3=0.001308.
NegativevaluescorrespondtoaSPIdeficitrainfallwhile,conversely positive values indicate higher than normalrains.Table1detailsaclassificationsystemthatdefinestheintensityofdroughtsbasedonthevalueoftheSPIforatimescalewhatsoever[1].
Table1:ClassificationofcategoriesofdroughtaccordingtothevaluesoftheSPI[1]
SPIValues Category2.0andmore Extremelyhumid
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1302
1.5to1.99 Highly humid1.00to1.49 Moderatelyhumid‐0.99to0.99 Closetonormal‐1.0to‐1.49 Moderatelydry‐1.5to‐1.99 Highlydry‐2.0andless Extremelydry
2.2 Artificial Neural Networks (ANN):MultilayerPerceptron(MLP)
An artificial neural network is a computational modelwhose original inspiration was a biological model. ThemainstrengthofANNistheirabilitytoidentifycomplexand nonlinear relationship between input and outputdata setswithout theneed to understand thenatureofphenomena [10]. Multilayer Perceptron Type (MLP) isthe simplest andmost commonly neural network used.Themathematical representationofanartificialneuroncalledartificialneuronisshownschematically inFigure1[11].
Figure1:Structureofartificialneuralnetwork
Wecanthencharacterizeaformalneuronby:
A combination function or an added thatperformstheweightedsum.Theweightedsumisequalto:
A W X
WhereWijisthesynapticweightandXiistheinputvaluerelative to the variable i. This is the weighted sum ofactivation,whichconvergestotheneuronj;
An activation function (or transfer) f thatanimatestheneuronbydeterminingitsactivation; Anactivation,theequivalentofneuronoutput.Itisequalto:
Whereθ isthebiasoftheneuronj.
There are several activation functions (hyperbolictangent, Gaussian, sigmoid ...), but themost used is thesigmoid function ([12], [13], [14]). It is written asfollows:
11 exp
The configuration of the best MLP model and itsimplementation amounts to choosing the transferfunctions, to identifytherelevant inputs, thenumberofneurons in the hidden layer, to choose the learningalgorithmandthenoptimizeandtestthenetwork.
3. APPLICATION
3.1 Studyregionanddataused
This study focuses on the basin of Inaouen (Figure 2).Locatedbetween latitudes34°35’24’’Nand33°50’24’’N,and longitudes 4°49’48’’W and 3°48’36’’W, it covers anarea of 3648 km2 with an average altitude of 694 m.Located in the western part of the basin of Sebou, theInaouenregion profits from a Mediterranean climatewithoceanicinfluence[15].Thisclimateischaracterizedby strong seasonal contrasts and very irregular rainfall[16]. Geological domainof Inaouenbasin is divided intotwomainareas:thenorthernpartofthebasinbelongstothe Pre‐Rif area and the southern part of the basin isrelatedtotheAtlasarea.
Thedatausedtopresenttheresultsinthisarticlearethemonthly rainfall in the Bab Marzouka station locatedupstream of the basin and Idrissfirst station located atthe downstream (Figure 2). These valueswere used tocalculatetheSPI.AlsothevaluesoftheNAOindexwereused to estimate the influence of global climate indexNorthAtlantic.
ThefollowingTable2showssomecharacteristicsofthetwostationsused:
Table2:Descriptionofthestations
stations BabMarzouka Idriss1st
Latitude/longitude34°12'2''N/4°8'27''W
34°9'47''N/ 4° 45' 2''W
Altitude(m) 368 200
Periodofdataused 1971‐2011 1975‐2012
Nb. Of observations(months)
480 444
Statisticaldescriptionof MonthlyPrecipitation
Average(mm) 46.86 32.69
Max(mm) 311.1 192.3
Standarddeviation
54.21 37.54
W1j
W2j
W3j
W4j
Wnj
fYj
X1
X2
X3
X4
Xn
Input ValuesWeights
comminationfunction
Weightedsum Aj
Activationfunction
Activation
Bias
j
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1303
Figure2:LocationofrainfallstationsusedwithinthebasinInaouen
3.2 Elaboration and processing of thedatabase
3.2.1 DataCollection
The purpose of this step is to collect data, both todevelopdifferentpredictionmodelsandtotestthem.
In the case of applications on real data, the goal is togather a sufficient number of data to provide arepresentativebasisofdatathatmayoccurduringuseofdifferentpredictionmodels.
The data used in this study are related to themeasurementcollectedduringtheperiod1971‐2011fortheBabMarzoukastationandtheperiod1975‐2012forthefirstIdrissstation.
The independent variables are standardizedprecipitation index, rainfall and the North AtlanticOscillationindex.(Figure3)
Figure3:Inputs/Outputsofpredictionmodel
3.2.2 Dataanalysis
Afterdatacollection,ananalysisisrequiredtodeterminethe discriminating features to differentiate or detectthese data. These characteristics are the input of theneuralnetwork.
The determination of the characteristics hasconsequencesonboththesizeofthenetwork(andthusthesimulationtime),onsystemperformance(separation
power,predictionrate),anddevelopmenttime(learningtime).
Datafilteringcanhelpremovethosethatareabsurdandredundant.
3.2.3.Datanormalization
Ingeneral,databasesrequireapretreatmentinordertobe adapted to the inputs and outputs of the stochasticmathematical models. A common preprocessing is toconduct a propernormalization that takes into accountthemagnitudeofthevaluesacceptedbythemodels.
Normalizationofeachinputxiisgivenbytheformula:
x 2 ∗x min x
max x min x1
However, the output variable is normalized between 0and1bytheformula:
x x min x
max x min x
Thisprovidesastandardizeddatabasebetween‐1and1forthedifferentvariationsof the independentvariables(model inputs) and between 0 and 1 for the differentvariationsofthedependentvariables(modeloutput).
3.3 Designing the architecture of artificialneuralnetworks
WhendevelopinganANNmodel,themainobjectiveistoachieve theoptimalANNarchitecture that captures therelationship between the input and output variables.First, a three‐layerperceptronMLPwaschosensince itwas found that a three‐layer model is sufficient forforecasting and simulation in the field ofwater science[17].Thenthetaskistoidentifythenumberofneuronsineachlayer.Fortheseinputandoutput,itissimpleanditisnormallydictatedbytheinputvariablesandoutputconsidered,tomodelaphysicalprocess.Forthenumberof neurons in the hidden layer, it should be optimizedusingtheavailabledata.Todoso,weproceededbytrialand error based on the measurement of the meanabsolute error (MAE) for thedata used to test for eachmodel. Figure 4 illustrates an exemplary model forestablished SPI 24 (SPI calculated for a window of 24months)andshowsthat,26neuronsinthehiddenlayeristheoptimalnumber.
Figure4:Determinationoftheoptimumnumberofneuronsinthehiddenlayer,dependingonthemeanabsoluteerrorfortheSPI24BabMarzoukastation
SPI (t+1) Mathematical model
SPI (t)
SPI (t‐1)
SPI (t‐2)
P (t)
P (t‐1)
NAO (t)
NAO (t‐1)
NAO (t‐2)
P (t‐2)
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1304
3.4
LeardetetraindesiWhefunc
With
ek:t
k:n
OptifuncweigAmosuchMarfast,
Lear(MSEcomoccunetw
3.5
Thiscomandexprexamtakecoeferroared
T
T
M
M
Whe
Learnin
rning, supervermining theningbase, thred output, aen comparinctioncalledgl
h:
theerrorbet
umberofiter
imize the nection.Themightw, of theongthedifferhas"theleasquardt back‐accurateand
rningthemoE) decreases
mpleted20,00ursfirst.Thusworkisuseda
Testan
s step is tomparing the d
the real varessed in twminationofthen in this stufficient of deor (MAE) anddefinedasfol
Thecorrelatio
r∑
∑ x
Thecoefficien
R
MeanAbsolut
MA
MeanSquare
ere x , the
ngofArtific
vised by theweights thaedifferencesand the valug the calcullobalsquared
E12
tweenthedes
rations
etwork thereinimization ise ANN usingrentalgorithmstsquares",w‐propagationdreliable[10
delstopswhs by less tha00iterations,s,theoptimaasanoutput
ndPerforma
o evaluatedifference bealue. The rewo ways: byhegraphs.Thudy are: theeterminationd themean sllows:
oncoefficient
x , x
, x
ntofdetermin
∑ x∑ x
teError(MAE
AE1N
x
Error(MSE)
∑
real value
cialneuraln
neural netwat minimize tsbetween thes of the estlated and acderrorEisca
e
siredvaluean
efore, is tos achievedbyan optimizamsforminimweoptedforn of error alg0].
henthemeann 0.01 or ifdependingoalsolutioncatarget.
anceAnalys
the networetween the eesult of they statisticalheindicatorscorrelation
n (R2), the msquared error
t(r)
x ,
∑ x ,
nation(R²)
, x
, x
E)
, x ,
, ,
e of SPI is
network
work consiststhe data sethevaluesof ttimated outpctual, an erralculated:
ndtheoutput
minimize thy adjusting ttion algorith
mizingfunctiotheLevenbergorithm that
nsquarederrthe model honthesituatialculatedbyt
sis
rk formedstimated val evaluationindicators athathavebecoefficient (mean absolur (MSE),whi
x
x
, x , is t
ofoftheut.ror
t
histhehm.onsrg‐t is
rorhasonthe
bylueisndeen(r),uteich
the
esth
Thanre
Thne(F
4
InANhatimraanm
stimatedvaluheaverageva
he best prednd MAE andespectively.
hesestepsliseuron are shFigure5):
Figure5:M
. RESULTS
n this study,NN previousavedevelopeme series ofainfallandSPnd the indexmodelsaresho
ueofSPIbytalueandNist
diction iswheMSE on the
stedtodesignhown in the
Methodologicpredictio
SANDDISC
drought (SPly considereedthreemodf precipitatioPI,andMLP3x north AtlaowninTable
themodel,atthenumbero
en r andR2e other hand
nanetworkmfollowing ch
calflowchartnofdrought
CUSSION
PI values) isd (Figure 5)delsMLP:MLon, MLP 2 b3onvariableantic Oscillat3below:
themomentofsamples.
on the one tend to 1 a
modelofartihart methodo
appliedtoth
s predicted u. In our caseLP1basedonbased on varrainfall,SPIition. These t
ti.xis
handand 0
ificialology
he
usinge, wen theriableindexthree
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1305
Table3:Differenttestedmodels
InputVariables PredictedVariable
ModelName
P(t),P(t−1),P(t−2) SPI(t+1) MLP1
SPI (t ), SPI (t− 1), SPI (t− 2), P(t ),P(t−1),P(t−2)
SPI(t+1) MLP2
SPI (t ), SPI (t− 1), SPI (t− 2), P(t ),P(t − 1),P(t− 2), NAO(t),NAO(t − 1),NAO(t−2)
SPI(t+1) MLP3
For these threemodels, the SPI valueswere calculatedforarangeoftimewindowsvaryingfrom3monthsto24months. Assessing the accuracy of predictions for eachmodel was performed using R2, r, MAE and MSEcoefficients.
TheevaluationofthethreemodelsbuiltshowedthatthepredictionresultsrelatedtoMLP3(Figure6)modelhadthelowesterror(MAEandMSEvaluesarelowerandR2andrvaluesarehigher). It is interesting toexpress theaddition of NAO as an input variable for the MLPimproves the efficiency of prediction models. Thus, inthis study, only the resultsofMLP3arepresentedanddiscussed.
Figure6:ModelANN(three‐layerPerceptron)usedtopredicttheriskofdrought
AllnetworksettingsonMLP3modelaresummarizedinTable4.
Table4:InformationabouttheMLP3modelused
InputLayer
Covariables 9SPI (t ), SPI (t − 1),SPI (t − 2), P(t ),P(t − 1), P(t − 2),NAO(t),NAO(t − 1),
NAO(t−2)
Numberofneurons 9
Hiddenlayer(s)
Number of hiddenlayers
1
Number of neurons inthehiddenlayer
Variable dependingonmodel
ActivationFunction Sigmoïde
OutputLayer
Dependentvariables 1 SPI(t+1)
Numberofneurons 1
Activationfunction Sigmoïde
ErrorFunction Sumofsquares
The estimated precisions of the model MLP3 bycomparisonoftheactualvaluesandthepredictedvaluesfor both stations Idriss 1st and Bab Marzouka fordifferenttimewindows(SPI3toSPI24)andtheoptimalarchitecturesofthemodelsarepresentedinTable5.
Table5:PredictionresultsofSPIvalues(atdifferenttimescales)forMLP3modelatthestationsIdriss1st
andBabMarzouka
Idriss1st
Architecture R2 r MAE MSE
SPI3 [9‐13‐1] 0.47 0.686 0.547 0.486
SPI6 [9‐11‐1] 0.671 0.819 0.433 0.315
SPI9 [9‐17‐1] 0.758 0.871 0.357 0.238
SPI12 [9‐9‐1] 0.828 0.91 0.286 0.153
SPI24 [9‐20‐1] 0.914 0.956 0.214 0.08
BabMarzouka
Architecture R2 r MAE MSE
SPI3 [9‐23‐1] 0.51 0.714 0.536 0.498
SPI6 [9‐20‐1] 0.62 0.788 0.489 0.399
SPI9 [9‐6‐1] 0.786 0.887 0.362 0.219
SPI12 [9‐9‐1] 0.871 0.933 0.275 0.132
SPI24 [9‐26‐1] 0.923 0.961 0.214 0.079
To visualize the evolution of performance indicatorsbasedonthetimewindowofSPI,theresultsinthetable5havebeen transformed into the following twographs(Figure7):
I NPUT L AYER HI DD EN LAYER OU TPU T LAYER
Biais
Biais
SPI (t+1)
SPI (t)
SPI (t-1)
SPI (t-2)
P (t)
P (t-1)
NAO (t)
NAO (t-1)
NAO (t-2)
P (t-2)
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1306
Figure7:Evolutionofperformanceindexesr,R2,MAEandMSEaccordingtothetimewindowofSPIforIdriss1ststations(left)andBabMarzouka(right)
We can see that all the predictions SPI produce verygoodprecisionexceptforSPI3withdifferencesbetweenobservedandpredictedvalues,arenotveryacceptable.Wealsoseethatwhenthetimewindowisincreased,thecorrelationbetweenthemodelpredictionandtheactualvalues increases significantly. This observation can beexplained by the way the time series SPI is calculated.Unliketheseriesofprecipitation,SPIfollowsastandardnormal distribution. This conversion eliminates suddenpeaks leaving a slowly varying smooth curve that iseasiertopredictusingmodelsofneuralnetworks.
Tobettervisualizetheperformanceofthetestedmodel,graphs that compare the observed SPI values and SPIvalues predicted one month ahead, and their scatterdiagramsforthefirstIdrissstationareshowninFigure8.
It can be concluded that the neural network model ofMLP,has successfullypredicted thedroughtonemonthaheadforseveraltimescalesofSPI.
SPI3 SPI6 SPI9 SPI12 SPI24
0
0.2
0.4
0.6
0.8
1
rR2MAEMSE
SPI3 SPI6 SPI9 SPI12 SPI24
0
0.2
0.4
0.6
0.8
1
rR2MAEMSE
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1307
Figure8:ComparisonofpredictedandobservedSPIvaluesforthefirstIdrissstation
5. CONCLUSION
In this study, the drought prediction is made for theregion of Inaouenbasin in northern Morocco using asystembasedonartificialneuralnetworkofMLP.Firstly,the time series of the Standardized Precipitation Index(SPI)builtinfordifferentperiodsoftimerangingfrom3months, 6, 9, 12 and 24 months using the values ofaverage monthly rainfall of the two stations selectedwithinthewatershedweather.ThenanANN‐MLPmodelwas developed and selected for his performance inpredictingSPIcategoriesforperiodsof3,6,9,12and24months.
The model shows superior forecasts when going fromSPI3toSPI24.Theneuralnetworkmodeldevelopedcanbe therefore a very useful tool for planners of waterresources to take the necessary measures in advancewhen there is shortage of water which can eventuallydevelopintodroughtconditions.
REFERENCES
[1] McKee TB., Doesken NJ., Kleist J.,“Therelationship of drought frequency and duration to timescales”. In:8th Conf. on Applied Climatology, 17–22January, American Meteoriological Society,Massachusetts,1993,p179–184.
[2] Manssouri T., Sahbi H., Manssouri I.,“Elaboration of stochasticmathematical models for theprediction of parameters indicative of groundwaterquality Case of Souss Massa – Morocco”. IJCERInternational Journal of Computational EngineeringResearch,Vol.04,No.5,2014,p31‐40.
[3] El Badaoui H., Abdallaoui A., Manssouri I.,OusmanaH.,(2013)“Thepredictionofmoisturethrough
the use of neural networksMLP type”. IOSR Journal ofComputerEngineering(IOSR‐JCE),Vol.11,No.6,2013,p66‐74.
[4] EL‐SHAFIE,AmrH.,EL‐SHAFIE,A.,ELMAZOGHI,Hasan G., et al. Artificial neural network technique forrainfall forecasting applied to Alexandria, Egypt.International Journal of Physical Sciences, Vol. 6, No. 6,2011,p.1306‐1316.
[5] Somvanshi V. K., Pandey O. P., Agrawal P. K.,Kalanker1 N.V., Prakash M.R. and Chand R., “Modellingandpredictionofrainfallusingartificialneuralnetworkand ARIMA techniques”, J. Ind. Geophys. Union, Vol.10,No.2,2006,p141‐151.
[6] EdwardsD.C.andMcKeeT.B.“Characteristicsof20th century drought in the United States at multipletimescales”.AtmosphericSciencePaper,No.634,1997.
[7] THOMH. C. S.,“Somemethods of climatologicalanalysis”. World Meteorological Organization TechnicalNoteGeneva,No.81,1966,53p.
[8] YoungKC.“Athree‐waymodelfor interpolatingfor monthly precipitation values”. Monthly WeatherReview,Vol.120,No.11,1992,p2561–2569.
[9] Abramowitz M. and Stegun A. “Handbook ofmathematical formulas, graphs, and mathematicaltables”.DoverPublicationsIncNewYork,1965.
[10] AdamowskiJ.,SunK.“Developmentofacoupledwavelet transformandneuralnetworkmethod for flowforecasting of non‐perennial rivers in semi‐aridwatersheds”.JournalofHydrology,Vol.390,No.1,2010,p85–91.
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1308
[11] Andronova V.,“Utilisation de données météo etdes réseauxde neurones pour la prédictionde vitessesdevent”.ProjetdeFind’EtudeMaster2èmeannée,Univ.Tech.DeSofia,2006,75p.
[12] Vecci L., Piazza F., and Uncini A. “Learning andapproximation capabilities of adaptive spline activationfunction neural networks”. Neural Networks, Vol. 11,No.2,1998,p259–270.
[13] ELTabach,E.,Lancelot,L.,Shahrour,I.andNajjarY.“Use of artificial neural network simulationmetamodelling to assess groundwater contamination ina road project”.Journal Elsevier Mathematical andcomputermodelling,vol.45,no7,2007,p.766‐776.
[14] RudeJ.“Développementd’unmodèlestatistiqueneuronal pour la description fine de la pollution par ledioxyded’azote”. Thèse d’Etat, Université Paris XII‐ValdeMarne,Paris,2008,140p.
[15] Sibari H., Haidi S., Ait Fora A. “Typologie descrues et érosion mécanique dans un bassin versant dezone semi‐aride : bassin versant de l'Inaouène”,Maroc.Science et changements planétaires. Revue Sécheresse.Vol.12,No.3,2001,p187‐93.
[16] El Garouani A. &Tribak A., “Relation entrehydrologie et climat dans le bassin versant de l’OuedInaouen(pré ‐Rif marocain) ”. Climate variability andchange –Hydrological Impacts (Proceedings of the FifthFRIENDWord Conference held at Havana, Cuba); IASHPubl.No.308,2006,p447‐453.
[17] WangW., Ding J. “Wavelet NetworkModel andIts Application to the Prediction of Hydrology”. NatureandScience,Vol.1,No.1,2003,p67‐71.
BOUDAD B. et al.
Date of Publication: August 27, 2014
ISSN: 2348-4098
Volume 2 Issue 6 August 2014
International Journal of Science, Engineering and Technology- www.ijset.in 1309