Unknown Future Prediction by Using ANN

Hi, I am trying to predict future consumption values (2018-2038). But something is wrong with my code I guess. This is a time series of consumption. I can fit real samples and predicted samples with a high score (like %99). But when I come to the end (predict part) after nearly 2 or 4 steps ahead values start to being same. I think it is about normalization, becouse consumption have to increase because of its past and the maximum limit is 1 so values cant go further and it stays under 1. when I use another code for normalization like Xn=normalize(Xn) network cant find any correlation (NaN).

Apart from these when I normalize by using this way how can I denormalize the samples. what will be the maximum and minimum values after training. I am writing the code below. Thanks in advance.

xv=load('elektriktek.txt');

xn=xv';

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

figure,plot(xn)

data=[xn(1,1:end-10);xn(1,2:end-9);xn(1,3:end-8);xn(1,4:end-7);xn(1,5:end-6);xn(1,6:end-5);xn(1,7:end-4);xn(1,8:end-3);xn(1,9:end-2);xn(1,10:end-1)];%%%%5 giriþ 1 ileri tahmin

groups=xn(1,11:end);

data=data';

groups=groups';

indices = crossvalind('Kfold',length(data),3);%%%%

TestMSE_toplam=zeros(3,1);

TestMAE_toplam=zeros(3,1);

for i = 1:3

testy = (indices == i);

testy=testy;

trainy = (~testy);

trainy=trainy;

TestInputData=data(testy,:)';

TrainInputData=data(trainy,:)';

TestOutputData=groups(testy,:)';

TrainOutputData=groups(trainy,:)';

Tahmin=zeros(size(TestOutputData));

Ref=zeros(size(TestOutputData));

for m=1:10

rand('seed',1);

randn('seed',1);

net=newff(minmax(TrainInputData(:,:)),[m 1],{'tansig','tansig','purelin'},'trainlm');

net.trainParam.epochs=300;

net.trainParam.show=20;

%net.trainparam.lr=0.01;

net.trainparam.goal=0.00001;

[net,tr]=train(net,TrainInputData,TrainOutputData);

TrainOutput=sim(net,TrainInputData);

TrainMSE(m,1)=mse(TrainOutputData-TrainOutput);

TrainMAE(m,1)=mae(TrainOutputData-TrainOutput);

TrainReport=[TrainOutput'];

TestOutput=sim(net,TestInputData);

TestMSE(m,1)=mse(TestOutputData-TestOutput);

TestMAE(m,1)=mae(TestOutputData-TestOutput);

traintime(m,1)=sum(tr.time);

end;

[a2,b2]=find(TestMSE==min(min(TestMSE)));

[ax,bx]=find(a2==min(a2))

a2=a2(ax,bx)

b2=b2(ax,bx)

axx(i,1)=a2;

bxx(i,1)=b2;

TestMSE_son(i,1)=TestMSE(a2,b2);

TestMAE_son(i,1)=TestMAE(a2,b2);

traintime_son(i,1)=traintime(a2,b2);

%figure,plot(TestOutputData,'O'); hold on; plot(TestOutput,'*'); hold off

rand('seed',1);

randn('seed',1);

net=newff(minmax(TrainInputData(:,:)),[a2 1],{'tansig','tansig','purelin'},'trainlm');

net.trainParam.epochs=500;

net.trainParam.show=20;

%net.trainparam.lr=0.01;

net.trainparam.goal=0.00001;

[net,tr]=train(net,TrainInputData,TrainOutputData);

TestOutput=sim(net,TestInputData);

TestMSE_son(i,1)=mse(TestOutputData-TestOutput)

TestMAE_son(i,1)=mae(TestOutputData-TestOutput)

TrainMSE_son(i,1)=mse(TrainOutputData-TrainOutput)

TrainMAE_son(i,1)=mae(TrainOutputData-TrainOutput)

R2 = corrcoef(TestOutputData,TestOutput);

R22(i,1)= R2(1,2)^2

[m1,n1,r]=postreg(TestOutputData,TestOutput);

Reg(i,1)=r;

figure,plot(TestOutputData,'-O'); hold on; plot(TestOutput,'-*'); hold off

Tahmin=Tahmin+TestOutput;

Ref=Ref+TestOutputData;

end;

Tahmin_Test=Tahmin/3;

Ref_Data=Ref/3;

%%%%%%%%%%%%%%%%%%

TestMSE=(sum(TestMSE_son))/3

TestMAE=(sum(TestMAE_son))/3

TrainMSE=(sum(TrainMSE_son))/3

TrainMAE=(sum(TrainMAE_son))/3

Corr_R=(sum(R22))/3

Reg_R=sum((Reg))/3

figure,plot(Ref_Data);hold on; plot(Tahmin_Test,'r')

fid1 = fopen('hatalar_ysa.txt','w+');

fprintf(fid1,'TestMSE = %12.6f\n',TestMSE);

fprintf(fid1,'TestMAE = %12.6f\n',TestMAE);

fprintf(fid1,'TrainMSE = %12.6f\n',TrainMSE);

fprintf(fid1,'Corr_R = %12.6f\n',Corr_R);

fprintf(fid1,'Reg_R = %12.6f\n',Reg_R);

miktar=20 %miktar means number of step

tahmin=zeros(miktar,1); %Tahmin means prediction

test1=[xn(1,end-9);xn(1,end-8);xn(1,end-7);xn(1,end-6);xn(1,end-5);xn(1,end-4);xn(1,end-3);xn(1,end-2);xn(1,end-1);xn(1,end)];

for i=1:miktar;

tahmin(i,1)=sim(net,test1);

testn(1,1)=test1(2,1);

testn(2,1)=test1(3,1);

testn(3,1)=test1(4,1);

testn(4,1)=test1(5,1);

testn(5,1)=test1(6,1);

testn(6,1)=test1(7,1);

testn(7,1)=test1(8,1);

testn(8,1)=test1(9,1);

testn(9,1)=test1(10,1);

testn(10,1)=tahmin(i,1);

test1=testn

end;

figure,plot(tahmin)

Answer 1

[BERGHOUT Tarek]

I attached this file ,use this function it is great for norlmalization, but befor you normalize your data between 0 and 1 , i mean befor training try to do that:

1, load you data

2 save the bellow parameters because you will need them later in denormalizing :

- if your data is a matrix then Xmin=(min(min(data)));Xmax=(max(max(data)));

- if your data is a vector just use one max and one min;

3- normalize your data between 0 and 1;

4- train your data and get your outputs,

5- denormalize your data by: output=scaledata(output,Xmin,Xmax);

please accept my answer if it is helpful.

[Hatice GENÇ KAVAS]

Thank you,I put the code after loading data and I used both way for matrix and vextor. but I couldnt get correlation by using both of them

I wrote it like below;

Xmin=(min(min(xn)));Xmax=(max(max(xn)))