NARX multi step predictions for external test data by using training data?

I am trying to build a network to do some long term predictions. I've uploaded the data and the network architecture. My data is comprised of two datasets (training and test subsets as attached). NARX neural network is used for training Input_Data and Output_Data. Then I tried to make multiple step predictions for test set (Input_Data1 and Output_Data1) by using trained net function but I cannot do predictions longer than length of test day and the predictions are very poor. When I replace the test input data with train data to predict output time steps as {x2 = X(1,predictOutputTimesteps); >> x2 = X2(1,predictOutputTimesteps);} and {LI=length(Input_Data1); >> LI=length(Input_Data);}, the result follows same pattern with train data.

How can I correctly form multi step predictions for test data by using net functions of training data?

I hope I was clear in my query.

[Greg Heath]

Having trouble with *.mat

How about posting the data in a *.m or *.txt

Greg

[Oguz BEKTAS]

thnx, I posted the new "narx & data.m "files

[Greg Heath]

1. CORRECTION: Replace *.tst with *.txt

2. WARNING: You cannot name a file narx.m

3. The data.m file is missing X = Input_Data % 191 x 24

[Oguz BEKTAS]

Hi, sorry for the confusion. I updated the data files and the .m file. Please, import the data as tables.

[Greg Heath]

>Please, import the data as tables.

I have no idea what that means.

Please give me the commands to read these 4 data files into matrices.

[Oguz BEKTAS]

Hi I am a bit confused after i saw your last comment which you wrote after your answer below. I added the data on pro_d.m file and the M_prediction.m is the NARX prediction file. I tried to look at your answer and the correlation and crosscorrelation functions but I could not understand how it is related. Recently, I have received some values after I fitted the Output_data but they are very poor. I would be very grateful if you can explain your answer more. Regards

Answer 1

[Greg Heath]

I have found several problems with your exercise.

You will see the major problem if you just plot your data.

You should also see that your choice of lags and number of hidden nodes may be inadequate.

Look at the correlation and crosscorrelation functions to determine an effective subset of lags.

Thank you for formally accepting my answer

Greg

[Oguz BEKTAS]

Hi, thanks for your answer. I tried to look at the correlation and crosscorrelation functions but I could not understand how it is related. I have received some values after I fitted the Output_data but they are very poor. I would be very grateful if you can explain your answer more. Regards

[Greg Heath]

Search the NEWSGROUP and ANSWERS with

 greg NARXNET nncorr

Pay special attention to the tutorial.

[Oguz BEKTAS]

Hi

I looked at your codes on the the correlation and crosscorrelation functions to determine an effective subset of lags and applied them to my input and target series. It gave successful results but so what??? In the answers and newsgroups, I could not find how to use these vectors in narxnet training.

What is the point that I am missing? Can you tell me what can I do after the code below?

X = zscore(cell2mat(x));

T = zscore(cell2mat(t));

[ I N ] = size(X)

[ O N ] = size(T)

crosscorrXT = nncorr(X,T,N-1);

autocorrT = nncorr(T,T,N-1);

crosscorrXT(1:N-1) = []; % Delete negative delays

autocorrT(1:N-1) = [];

sigthresh95 = 0.21 % Significance threshold

sigcrossind = crosscorrXT( crosscorrXT >= sigthresh95 )

sigautoind = autocorrT( autocorrT >= sigthresh95 )

inputdelays = sigcrossind(sigcrossind <= 35)

feedbackdelays = sigautoind(sigautoind <= 35)

feedbackdelays(1)=[] % Delete zero delay