SUGGESTIONS TO IMPROVE OPEN LOOP NARNET MODEL PERFORMANCE FOR MULTISTEP AHEAD PREDICTIONS (WIND SPEED DATASET)

Question

Saira AlZadjali 2018 年 10 月 18 日

1
リンク

この質問への直接リンク

https://jp.mathworks.com/matlabcentral/answers/424696-suggestions-to-improve-open-loop-narnet-model-performance-for-multistep-ahead-predictions-wind-spee

Dears, I am Student working on Project (to predict wind Speed), and I am Student of MATHWORKS too learning how to build my model it have been almost a year i am reading and fixing. I got the steps and idea/suggestions specially from Sir Greg and by following the steps and idea i can get good results for DataSet in MATLAB, but for my DataSet its not working. let me explain... the below steps are explained by Sir Greg in many Posts, so i started to follow the steps...

% 1. Use NNCORR to determine the statistically significant input and/or % feedback lags to be stored in row vectors ID and/or FD with lengths % LID and LFD, respectively. % SZ: Since I am using NARnet, I will only find FD.

% 2. Use an outer loop search over a range of ~10 values (for number of % hidden nodes h = Hmin:dH:Hmax )

% 3. An inner loop search over Ntrials ~10 RNG states (i=1:Ntrials) that % determine random initial weights and random trn/val/tst data division.

%4. MSE goal ~ 0.01*mean(var(target',1)) (Models ~99% of target variance) % 5. MinGrad = MSEgoal/100

all these steps are followed, with data in MATLAB "simplenar_dataset" as below, .. everything went smoothly. but when i used my dataset, i am not getting FD from NNCORR steps (I am getting a huge number like ~8400, which is too much to apply my computer get stuck and cant continue with the selections). Please Advise what I need to try else to make my MODEL to predict multi step ahead with better accuracy as its working with Matlab dataset.

       % 1. Use NNCORR to determine the statistically significant input and/or 
      % feedback lags to be stored in row vectors ID and/or FD with lengths 
      % LID and LFD, respectively.  For NARnet you need FD only
      % NNCORR
      close all, clear all, clc , plt = 0;
      S = simplenar_dataset;
      T=S(1:80);             % remaining 20 for prediction and test
      t  = cell2mat(T) ;
      [ O N ] = size(t)      % [ 1 80]
       zt        = zscore(t,1);   tzt = [ t ; zt ];
       meantzt   = mean( tzt' )'       %[ 0.7233 ;-0.0000]
       vartzt1   = var( tzt', 1)'      % [0.0635 ; 1 ]
       minmaxtzt = minmax(tzt)
       % minmaxtzt = 
       %  0.1622    0.9999
       % -2.2273    1.0981
       Ntst   = round(0.15*N), Nval = Ntst % 12, 12
       Ntrn   = N-Nval-Ntst       % 56
       trnind = 1:Ntrn;           % 1 : 56
       valind = Ntrn+1:Ntrn+Nval; % 57 : 68
       tstind = Ntrn+Nval+1:N;    % 69 : 80
       ttrn = t(trnind); tval = t(valind);ttst = t(tstind);
     plt  = plt+1, figure(plt), hold on
     plot( trnind, ttrn,'b' )
     plot( valind, tval,'g' )
     plot( tstind, ttst, 'r' )
     legend('TRAIN','VAL','TEST')
     title( ' DATASET TIMESERIES ' )
     rng('default');
     Ltrn  = floor(0.95*(2*Ntrn-1)) % 105
     imax = 100
     for i = 1:imax
        n                = randn(1,Ntrn);
        autocorrn        = nncorr( n,n, Ntrn-1, 'biased');
        sortabsautocorrn = sort(abs(autocorrn));
        thresh95(i)      = sortabsautocorrn(Ltrn);
    end
                              % imax =  100    
     minthresh95  = min(thresh95)     % 0.0983  
     medthresh95  = median(thresh95)  % 0.1841  
     meanthresh95 = mean(thresh95)    % 0.1950 
     stdthresh95  = std(thresh95)     %  0.0515 
     maxthresh95  = max(thresh95)     % 0.3360 
     sigthresh95  = meanthresh95      % 0.1950
     zttrn        = zscore(ttrn',1)';
     autocorrttrn = nncorr( zttrn, zttrn, Ntrn-1, 'biased' );
     siglag95     = find(abs(autocorrttrn(Ntrn+1:2*Ntrn-1)) ...
                  >= sigthresh95);                 
     Nsiglag95    = length(siglag95)            % 30 ~ 0.54*Ntrn
     d            = min(find(diff(siglag95)>1)) % 5
     proof        = siglag95(d-1:d+1)           % 4     5     7
    plt = plt+1; figure(plt);  hold on
    plot(0:Ntrn-1, -sigthresh95*ones(1,Ntrn),'b--' );
    plot(0:Ntrn-1, zeros(1,Ntrn),'k');
    plot(0:Ntrn-1, sigthresh95*ones(1,Ntrn),'b--' );
    plot(0:Ntrn-1, autocorrttrn(Ntrn:2*Ntrn-1));
    plot( siglag95, autocorrttrn(Ntrn + siglag95), 'ro' );
    title('SIGNIFICANT AUTOCORRELATIONS VALUES')
     to    = t(d+1:N);
     No    = N-d            % 75
     Notst = round(0.15*No) % 11
     Noval = Notst          % 11
     Notrn = No-Noval-Notst % 53
%  --------------------------------------------------------------------- %
% 2. Use an outer loop search over a range of ~10 values (for number of 
% hidden nodes h = Hmin:dH:Hmax ) 
% 3. An inner loop search over Ntrials ~10 RNG states (i=1:Ntrials) that 
% determine random initial weights and random trn/val/tst data division. 
% -----------------------------------------------------------------------%
% finding the best suited FD and H 
   H  = 10              % Default
   neto                 = narnet(1:d,H);           % d=5
   neto.divideFcn       = 'divideblock';
   [ Xo, Xoi, Aoi, To ] = preparets(neto,{},{},T);
   to = cell2mat(To);     varto1 = var(to,1)       % 0.0628
   Hub     = (Notrn*O-O)/(d+O+1)                   % 7.4286
   Hmin    = 1, dH = 1, Hmax = 7
   Ntrials = 5
   rng('default')
   j  =  0
   for h = Hmin:dH:Hmax
       j = j+1
      neto.layers{1}.dimensions = h;
      for i = 1:Ntrials
          neto      = configure(neto,Xo,To);
          [neto tro Yo Eo Xof Aof ]=train(neto,Xo,To,Xoi,Aoi);
          %[Yo Xof Aof =net(Xo,Xoi,Aoi);Eo=gsubtract(To,Yo);
          NMSEo(i,j)    = mse(Eo)/varto1;            
          NMSEtrno(i,j) = tro.best_perf/varto1;     
          NMSEvalo(i,j) = tro.best_vperf/varto1;   
          NMSEtsto(i,j) = tro.best_tperf/varto1;    
      end
   end
   plotresponse(To,Yo)
   NumH      = Hmin:dH:Hmax
   Rsqo      = 1-NMSEo         
   Rsqtrno   = 1-NMSEtrno       
   Rsqvalo   = 1-NMSEvalo        
   Rsqtsto   = 1-NMSEtsto       
%  NumH =
% 
%      1     2     3     4     5     6     7
% 
% 
% Rsqo =
% 
%     0.9981    0.9992    0.9990    1.0000    1.0000    1.0000    1.0000
%     0.9981    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9981    1.0000    0.9988    1.0000    1.0000    1.0000    1.0000
%     0.9980    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9981    0.9994    1.0000    1.0000    1.0000    1.0000    1.0000
% 
% 
% Rsqtrno =
% 
%     0.9979    0.9991    0.9988    1.0000    1.0000    1.0000    1.0000
%     0.9979    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9979    1.0000    0.9986    1.0000    1.0000    1.0000    1.0000
%     0.9979    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9979    0.9993    1.0000    1.0000    1.0000    1.0000    1.0000
% 
% 
% Rsqvalo =
% 
%     0.9986    0.9999    0.9991    1.0000    1.0000    1.0000    1.0000
%     0.9986    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9986    1.0000    0.9989    1.0000    1.0000    1.0000    1.0000
%     0.9986    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9986    0.9998    1.0000    1.0000    1.0000    1.0000    1.0000
% 
% 
% Rsqtsto =
% 
%     0.9984    0.9992    0.9996    1.0000    1.0000    1.0000    1.0000
%     0.9984    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9984    1.0000    0.9993    1.0000    1.0000    1.0000    1.0000
%     0.9984    1.0000    1.0000    1.0000    1.0000    1.0000    1.0000
%     0.9984    0.9993    1.0000    1.0000    1.0000    1.0000    1.0000
   %-------------------------------------------------------------------%
   % From above FD is obtained to be 5 and H neurons, actually we
   % should select the min H which gives Rsqtsto >0.995, in this case H=1 is
   % suitable.
   %------------------------------------------------------------------%
% now using the FD & H to train the network and predict multi steps ahead
% 4. MSE goal ~ 0.01*mean(var(target',1))  (Models ~99% of target variance) 
% 5. MinGrad = MSEgoal/100 
% 
%  MSEgoal = ~ 6.2800e-04
%  MinGrad = ~6.2800e-06 
close all, clear all, clc , plt = 0; 
S = simplenar_dataset;
T= S(1:80);
[ I N ] = size(T) % [ 1 100 ]
FD = 1:5; H = 1; % obtained from NNCORR
neto = narnet( FD, H );
neto.divideFcn = 'divideblock';
%view(neto)
[ Xo, Xoi, Aoi, To ] = preparets( neto, {}, {}, T );
to = cell2mat(To);     varto1 = var(to,1) 
neto.trainParam.epochs = 1000; % DEfault
%  MSEgoal = ~ 6.2800e-04
%  MinGrad = ~6.2800e-06
neto.trainParam.goal= 6.2800e-04 ;
neto.trainParam.min_grad=6.2800e-06 ;
rng( 'default' )
[ neto, tro, Yo, Eo, Aof, Xof ] = train( neto, Xo, To, Xoi, Aoi );
%[ Yo Xof Aof ] = net( Xo, Xoi, Aoi )
%Eo = gsubtract( To, Yo )
%view( neto )
NMSEo = mse( Eo ) /varto1 %  0.0024
yo = cell2mat( Yo );
% plot( d+1:N, yo, 'ro', 'LineWidth', 2 )
% axis( [ 0 100 0 1.3 ] )
% legend( 'TARGET', 'OUTPUT' )
% title( 'OPENLOOP NARNET RESULTS' )
% now converted to closed loop to do multi step predictions 
[ netc, Xci, Aci ] = closeloop( neto, Xoi, Aoi );
%view( netc )
[ Xc, Xci, Aci, Tc ] = preparets( netc, {}, {}, T );
isequal( Tc, To ) , tc = to ; % 1
[ Yc, Xcf, Acf ] = netc( Xc, Xci, Aci );
Ec = gsubtract( Tc, Yc );
yc = cell2mat( Yc );
NMSEc = mse(Ec) /var(tc,1) %0.0245
Tp= S(81:100)
m=1; 
n=6;
Tf= Tp (m:n);
[ Xc1, Xci1, Aci1, Tp1 ] = preparets( netc, {}, {}, Tf );
[ Yc1, Xcf, Acf ] = netc( Xc1, Xci1, Aci1 );
Xc2 = cell(1,10);
[ Yc2, Xcf2, Acf2 ] = netc( Xc2,Xcf, Acf  ); 
yc2 = cell2mat(Yc2);
Tpp =cell2mat(Tp);
    plt = plt+1; figure(plt), hold on
    plot( yc2,'b', 'LineWidth', 2 )
    plot( Tpp(7:16), 'r', 'LineWidth', 2 )
    legend( 'OUTPUT','TARGET')
    title( 'PREDICTION RESULTS' )
% >> yc2
% yc2 =
%    0.9985    0.4726    0.6192    0.9853    0.8055    0.3500    0.9204    0.9794    0.3134    0.7738
%>> Tpp(7:16)
%ans = 
%    0.9599    0.5351    0.6340    0.9911    0.7724    0.3774  0.9389    0.9925    0.2680    0.7762
%>> Ep = gsubtract( Tpp(7:16), yc2 )
%Ep =
%   -0.0386    0.0625    0.0148    0.0058   -0.0331    0.0274    0.0186    0.0131   -0.0454    0.0024
%>> mse (Ec)
%ans =
%    0.0010

All Above were the full complete steps to Design a NARnet Model, For each Data different FD, H , MSE goal, MinGrad and for sure Training Algorithm and the data division tech. For my data I tried to change all these parameters and do test and trials for different FD & H (as NNCORR not geving good results too), But Iam not getting the desired NMSEo, to continue ... :( Please Help...

below the results for my data, i randomly selected FD=6 and kept H=10 as default..

    close all, clear all, clc , plt = 0; 
data_1= xlsread('G:\SQU2\Solar_Wind_Project\MAtlabCodes\data_Nov13-Oct15');
T=data_1(1:92394,2);
T=con2seq(T');
[ I N ] = size(T) % [ 1 100 ]
d=6; FD = 1:d; H = 10; % obtained from NNCORR
neto = narnet( FD, H );
neto.divideFcn = 'divideblock';
%view(neto)
[ Xo, Xoi, Aoi, To ] = preparets( neto, {}, {}, T );
to = cell2mat(To);     varto1 = var(to,1)    % 12.7740
neto.trainParam.epochs = 1000; % DEfault
%  MSEgoal = ~ 0.1277
%  MinGrad = ~0.0013
neto.trainParam.goal= 0.1277 ;
neto.trainParam.min_grad=0.0013 ;
rng( 'default' )
[ neto, tro, Yo, Eo, Aof, Xof ] = train( neto, Xo, To, Xoi, Aoi );
%[ Yo Xof Aof ] = net( Xo, Xoi, Aoi )
%Eo = gsubtract( To, Yo )
%view( neto )
NMSEo = mse( Eo ) /varto1 %  0.0291
% now converted to closed loop to do multi step predictions 
[ netc, Xci, Aci ] = closeloop( neto, Xoi, Aoi );
%view( netc )
[ Xc, Xci, Aci, Tc ] = preparets( netc, {}, {}, T );
%[ netc, trc, Yc, Ec, Acf, Xcf ] = train( netc, Xc, Tc, Xci, Aci );
isequal( Tc, To ) , tc = to ; % 1
[ Yc, Xcf, Acf ] = netc( Xc, Xci, Aci );
Ec = gsubtract( Tc, Yc );
yc = cell2mat( Yc );
NMSEc = mse(Ec) /var(tc,1) %1.9321