How to forecast using ANN

HI, please help me to forecast feature groundwater level using feed forward neural network with LM algorithm, i developed the feed forward neural network with LM algorithm with ANN architecture 3-5-1, my model is as follows, but i want to use this model for forecasting at least 3 to 6 point ahead, can any one please help me,

clear all;

clc;

load dannlm.txt;

t=dannlm(:,1);

rain=dannlm(:,2);

et=dannlm(:,3);

gwl=dannlm(:,4);

%[ACF, Lags, Bounds] = autocorr(gwl, [], 2);

%[PACF, Lags, Bounds] = parcorr(gwl, [], 2);

%[CCF, Lags, Bounds] = crosscorr(gwl,rain);

gwlmin=min(gwl);

gwlmax=max(gwl);

rain=(rain-min(rain))/(max(rain)-min(rain));

et=(et-min(et))/(max(et)-min(et));

gwl=(gwl-min(gwl))/(max(gwl)-min(gwl));

%*****Data Preparation****

for t=5:195

      Data(t-4,:)=[rain(t-4) et(t-1) gwl(t-1) gwl(t)];

end

%Define Input patter for Training and Validation

INPTR=Data(1:150, 1:3);

TARTR=Data(1:150,4);

net=newff([0 1; 0 1; 0 1], [5 1], {'logsig', 'purelin'}, 'trainlm');

net.trainParam.epochs=500;

net.trainParam.goal=0.0001;

net.performFcn='mse';

net=init(net);

net=train(net, INPTR', TARTR');

a=sim(net, INPTR');

z=[a' TARTR];

INPVAL=Data(151:191, 1:3);

TARVAL=Data(151:191,4);

y=sim(net, INPVAL');

zv=[y' TARVAL];

%Converting back to Original Flow of gwl Validation

zv=zv*(gwlmax-gwlmin)+gwlmin;

save val1.txt zv -ascii;

z=z*(gwlmax-gwlmin)+gwlmin;

save cal1.txt z -ascii;

CORR_CAL=corrcoef(z)

CORR_VAL=corrcoef(zv)

COM_CAL=z(:,1);

OBS_CAL=z(:,2);

COM_VAL=zv(:,1);

OBS_VAL=zv(:,2);

Eff_ANN_CAL=1-(sumsqr(OBS_CAL-COM_CAL)/sumsqr(OBS_CAL-mean(OBS_CAL)))

Eff_ANN_VAL=1-(sumsqr(OBS_VAL-COM_VAL)/sumsqr(OBS_VAL-mean(OBS_VAL)))

RMSE_ANN_CAL=sqrt(sumsqr(OBS_CAL-COM_CAL)/length(OBS_CAL))

RMSE_ANN_VAL=sqrt(sumsqr(OBS_VAL-COM_VAL)/length(OBS_VAL))

Ex_Var_ANN_CAL=sqrt(sumsqr(COM_CAL-mean(OBS_CAL))/sumsqr(OBS_CAL-mean(OBS_CAL)))

%ERROR_CAL=100*(OBS_CAL-COM_CAL)/OBS_CAL;

%ERROR_VAL=100*(OBS_VAL-COM_VAL)/OBS_VAL;

ANN_PEAK_CAL=(1-max(COM_CAL)/max(OBS_CAL))*100

ANN_PEAK_VAL=(1-max(COM_VAL)/max(OBS_VAL))*100

[Anand Kumar]

Please convert above attachment to forecast at least 3 to 6 point ahead

Answer 1

[Greg Heath]

 help narxnet
 doc narxnet

 net = narxnet(3:6,3:6,5);

===================================================================

% ===>GEH1: WHAT DO THE NAMES DANNLM AND ET STAND FOR?

% ===> GEH2: AUTOCORR(gwl,...), CROSSCORR(GWL,RAIN,...) AND % CROSSCORR(GWL,ET,...) CAN BE USED TO FIND THE SIGNIFICANT DELAYS.

 for t=5:195
    Data(t-4,:)=[rain(t-4) et(t-1) gwl(t-1) gwl(t)];
 end

% ===>GEH3: What is the rationale for this combination?

% ===> GEH4: NEWFF has been obsolete for at least 5 years. Regardless, this is a time series problem which is more easily solved using NARXNET or it's obsolete predecessor.

 z=[a' TARTR];

% ===>GEH5: Why not compute the error mse(a'-TARTR) ???

 INPVAL=Data(151:191, 1:3);
 TARVAL=Data(151:191,4);

% ===> GEH6: This is TEST data, NOT VALIDATION DATA! VALIDATION DATA IS USED TO TUNE PARAMETERS. TEST DATA IS USED TO OBTAIN UNBIASED ESTIMATES OF PERFORMANCE

% THAT IS AS FAR AS I WENT. AFTER CALCULATING

 NMSE = mse(error)/mean(var(target',1) and/or 
 R2   = 1-NMSE

%SUCCESS OR FAILURE IS DETERMINED

Hope this helps.

Thank you for formally accepting my answer

Greg