norm form and prediction

hi

i want to use ann for forecasting water table

i have two set of data. water table from past and temperature.

it means two matrix with 12 column and 15 row

12 column means 12 month

and each row are separated data from a year(15 year).

it mean i have two data for month and year

i devided my data into 2 matrix

first 11 year for input and 4 year for output

i want to use nntool to forecast and predict water table for future years with under consideration of temperature. now my questions:

1. some of my data don't have value and matlab give them 0 in my table. is this affect the network? and if it is what should i do?
2. do i need to normalize my data before using nntool and how?
3. when i start nnstart, do i need to use time series?
4. how i consider the effect of temperature?

sorry for my long question. I am new in matlab and have a lot of question. thanks a lot.

Answer 1

[Greg Heath]

I would unfold your data into a single 2 dimensional input time series matrix of size

[ 2 180 ] = size(series)

where the inputs are waterlevel and temperature.

The target is just the future waterlevel.

To estimate how far into the future you can predict, obtain the autocorrelation function of waterlevel and crosscorrelation function of waterlevel and temperature.

From the correlation functions you can determine the significant lags of the local correlation peaks.

Then use a subset of the significant lags in a narxnet.

Hope this helps.

Thank you for formally accepting my answer

Greg

[majid a]

thanks for response

I remove some parameter in my question to simplify it.

i have 3 matrix. waterlevel, rain and temperature.

how should i unfold the matrix ? the column and rows are different. one of them is the year and one of them is month.

it get me a string of numbers.

how can i use crosscorrelation for 3 Characteristic?

[Greg Heath]

Assume you have 3 15x12 matrices of temperature, rainlevel, and waterlevel which are unfolded into 3 timeseries of N = 180 days and combined into one 3x180 data matrix.

For a prediction of d days ahead

 [ I N-d ] = size(input)  = [ 3 N-d ]
 [ O N-d ] = size(target) = [ 3 N-d ]

You would then have 3 autocorrelation functions and 9 (corrected from 6) cross-correlation functions to use for estimates of what d should be.

Of course you can reduce the number of outputs if you wish.

Hope this helps

Greg

[majid a]

thanks again for helping.

i divide my matrices and now i have 4 matrix.

1 for rain [1*180]

1 for temperature [1*180]

1 for waterlevel [1*180]

and a combine of those 3 [3*180]

when i use code :

    xcorr(rain)
%for finding auto correlation of rain

its get me a 1*355 matrix!

when i set the lag to 1 it gets me this

    xcorr(rain,1)

ans =

   1.0e+07 *

    2.0621    2.1315    2.0621

i searched in the internet and find a function that calculate acf but it needed lag p.

how can i know lag p?

and when i run it with lag 1 this happen:

    acf(rain,1)

ans =

    0.8492

i am really confused..

sorry to bother you again.

[Greg Heath]

Search in both the NEWSGROUP and ANSWERS using

 greg nncorr

 greg nncorr narxnet

Hope this helps.

Greg

Answer 2

[majid a]

%%Change Data
DD = [];
for ii=1:13
  D = Data(1:12,ii);
  DD = [DD ; D]
end

for ii=13:(numel(DD)-5)
  D_0(ii-12,1) = DD(ii-0);
  D_1(ii-12,1) = DD(ii-1);
  D_2(ii-12,1) = DD(ii-2);
  D_3(ii-12,1) = DD(ii-3);
  D_6(ii-12,1) = DD(ii-6);
  D_12(ii-12,1) = DD(ii-12);
  D_plus_3(ii-12,1)=DD(ii+3);
  D_plus_1(ii-12,1)=DD(ii+1);
end

 X = [D_0 D_1 D_2 D_3 D_6 D_12];
 Y = D_plus_1;
 DataNum = size(X,1);
 InputNum = size(X,2);
 OutputNum = size(Y,2);

 %%Normalization
 MinX = min(X);
 MaxX = max(X);

 MinY = min(Y);
 MaxY = max(Y);

 XN=X;
 YN=Y;

 for ii=1:InputNum
     XN(:,ii)=Normalize_Fcn(X(:,ii),MinX(ii),MaxX(ii));
 end
 for ii=1:OutputNum
     YN(:,ii) = Normalize_Fcn(Y(:,ii),MinY(ii),MaxY(ii));
 end

 %%Test And Train Data
 TrPercent=80;
 TrNum = round(DataNum * TrPercent/100);
 TsNum = DataNum - TrNum;

 R=randperm(DataNum);
 trIndex = R(1 : TrNum);
 tsIndex = R(1+TrNum : end);

 Xtr = XN(trIndex,:);
 Ytr = YN(trIndex,:);

 Xts = XN(tsIndex,:);
 Yts = YN(tsIndex,:);

 %%Network Structure
 pr =  [-1 1]
 PR = repmat(pr,InputNum,1);
 Network = newff(PR,[10 5 OutputNum],{'tansig' 'tansig' 'tansig'});

 %%Training
 [Network,NN] = train(Network,Xtr',Ytr');
 view(Network);
 %%Assesment
 YtrNet = sim(Network,Xtr')';
 YtsNet = sim(Network,Xts')';

 MSEtr = mse(YtrNet - Ytr)
 MSEts = mse(YtsNet - Yts)

I couldn't find out how to consider cross correlation in my network. and how i can forecast in the future??? thanks

[Greg Heath]

Did you search my nncorr postings?