Please help optimize "ANN forecasting code"

Question

Arya on 28 Apr 2013

0
Link

Direct link to this question

https://www.mathworks.com/matlabcentral/answers/73885-please-help-optimize-ann-forecasting-code

Dear Sir...

i have some data like this :

Jan-07 97.14
Feb-07 98.63
Mar-07 98.86
Apr-07 98.63
May-07 99.06
Jun-07 99.44
Jul-07 100.03
Aug-07 100.69
Sep-07 101.25
Oct-07 102.01
Nov-07 102.15
Dec-07 102.10
Jan-08 103.14
Feb-08 104.26
Mar-08 105.85
Apr-08 105.56
May-08 106.81
Jun-08 109.37
Jul-08 110.69
Aug-08 111.36
Sep-08 112.46
Oct-08 113.06
Nov-08 113.13
Dec-08 112.66
Jan-09 112.56
Feb-09 112.71
Mar-09 113.47
Apr-09 113.28
May-09 113.38
Jun-09 113.54
Jul-09 114.06
Aug-09 114.43
Sep-09 115.77
Oct-09 115.77
Nov-09 115.94
Dec-09 116.25
Jan-10 117.12
Feb-10 117.67
Mar-10 117.43
Apr-10 117.86
May-10 117.88
Jun-10 118.87
Jul-10 120.93
Aug-10 121.57
Sep-10 122.83
Oct-10 122.86
Nov-10 123.64
Dec-10 124.51
Jan-11 125.26
Feb-11 125.11
Mar-11 124.97
Apr-11 124.29
May-11 124.45
Jun-11 124.99
Jul-11 125.83
Aug-11 126.55
Sep-11 127.19
Oct-11 126.95
Nov-11 127.60
Dec-11 128.08
Jan-12 128.62
Feb-12 129.09
Mar-12 129.51
Apr-12 129.69
May-12 130.16
Jun-12 131.05
Jul-12 132.14
Aug-12 133.8
Sep-12 133.67
Oct-12 133.76
Nov-12 133.74
Dec-12 134.29

and i wanna forecast it for 12 months ahead... few days ago, i found a matlab code, from google, like this :

% Outputs:
% - RMSE , root mean squares error.
% - yL, matrix of y's lags.
% - TRAINEDNET , a NET with minimum mean squares error.
% - minRMSE , minimum of root mean squares error.
% - yf , forecast of y.
% Inputs:  diisi manual pake yoy tiap bulan
% - y , a time series in vertical vector form.(eg stock prices)
% - maxlag ,maximume lag that should be entered in model.(eg 5)  12
% - nhiden, number of hidden layer units.(eg 5)  12
% - trset,percent of observations for trainig set.(eg 80)
% - HPF ,number of priods that should be forecasted.(eg 10)
% - lr , Learning rate.(eg 0.1)  0.01 ato 0.1
%---------------------------------------------------------------------
% Building networks with different number of hiden units and input lags.
for lay=1:nhiden
for lag=1:maxlag
PR(1:maxlag,1)=[-1];
PR(1:maxlag,2)=[1];
net(lay,lag)={newff(PR(1:lag,:),[lay 1],{'tansig','purelin'},'trainlm' )};
net{lay,lag}.trainParam.lr = lr;
net{lay,lag}.trainParam.lr_inc = 1.05;
net{lay,lag}.trainParam.lr_dec = .5;
net{lay,lag}.trainparam.epochs=1000;
net{lay,lag}.trainparam.show=100;
net{lay,lag}.trainparam.goal=1e-5;
end
end
%---------------------------------------------------------------------
%defining lags for y
y=y';
[ny,mny,mxy]=premnmx(y);
[nyr,nyc]=size(y);
trset=floor(nyc*trset/100)
for lag=1:maxlag
yL=zeros(lag,nyc);
end
for s1=1:nyc
for s2=1:maxlag
yL(s2,s1)=NaN;
end
end
for lag=1:maxlag
yL(lag,1+lag:nyc)=ny(1:nyc-lag);
end
%---------------------------------------------------------------------
%training nets
for lay=1:nhiden
for lag=1:maxlag
%early stopping
val.P=yL(1:lag,trset+1:nyc);
val.T=ny(1,trset+1:nyc);
[TRAINEDNET{lay,lag},tr]=train(net{lay,lag},yL(1:lag,lag+1:trset),ny(1,lag+1:trset),[],[],val);
%RMSE calculation
yhat=sim(TRAINEDNET{lay,lag},yL(1:lag,trset+1:nyc));
rmse=((sum((yhat-ny(1,trset+1:nyc)).^2))/nyc)^.5;
RMSE(lay,lag)=rmse;
RMSEE=RMSE;
minRMSE=min(min(RMSE));
end
end
for lag=1:maxlag
for lay=1:nhiden
if RMSEE(lay,lag)==min(min(RMSE))
optlay=lay
optlag=lag
end
end
end
yfL=yL(1:optlag,:);
%---------------------------------------------------------------------
%training again the optimal net with complete set of data.
TRAINEDNET{optlay,optlag}=train(TRAINEDNET{optlay,optlag},yL(1:optlag,:),ny(1,:))
%H period forecasting
for o=1:HPF
yf=sim(TRAINEDNET{optlay,optlag},yfL(1:optlag,nyc-1:o+nyc-1));
yf=[ny yf];
for flag=1:optlag
yfL(flag,1+flag:nyc+o)=yf(1:o+nyc-flag);
end
end
yf2=sim(TRAINEDNET{optlay,optlag},yL(1:optlag,optlag+1:nyc));
yf=yf';
yf2=yf2';
%converting the data back into unnormalized units.
yf=postmnmx(yf,mny,mxy);
yf2=postmnmx(yf2,mny,mxy);
yhatopt=sim(TRAINEDNET{optlay,optlag},yL(1:optlag,trset+1:nyc));
yy(1:nyc-trset)=ny(trset+1:nyc);
error=(yhatopt./yy-1);
subplot(2,1,1);
plot([1:nyc-trset],yhatopt,'g-' ,[1:nyc-trset],yy,'r-');
legend('yhatopt','yy' ,0);
ylabel('value');
title(['Frecasted by Optimal Net']);
subplot(2,1,2);
plot([1:nyc-trset],error,'b-')
xlabel('Time');
ylabel('Residuals');
title(['Forecast Error for Test Set' ]);
%END