how can i bring mse near to 0 as it is very large on training data and how 5 steps ahead prediction is done with ntstool by using closeloop net

i trained neural network with data set now i need to do 4 steps ahead prediction, for this in targets it is better to give the real dataset or the each prediction to find 4 step ahead forecasting . what should be target value ? values from real data set or previously predicted once should be given as targets.

how to improve performance on traing data ? i tried by changing number of hidden layers , neurons and delay but its not working fine

the last thing can anyone help me regarding finding the proper lagged values as delay to be used in neural net with autocorrelation kind of things

Answer 1

[Greg Heath]

Insufficient detail.

What are the input and targets?

size(input) = ?

size(target)= ?

What are the significant lags of the target autocorrelation function and input/target cross correlation function?

[SAM]

hi greg,

please help me in identifying significant lag terms with both correlation and autocor, i have understood but i am unable to do it. if possible please send me an7y of ur link

[SAM]

Answer 2

[SAM]

HI GREG , thanks alot for responding , after reading many of your posts i m trying to find out the lagged values with autocorr and cross corr, but truly it is still not cleared to me as i hv recently started .

i have temperature data series of total 2825, inputs are 4*2795 and output 1*2795 , i need to forecast 5 steps ahead, rest of 30 values are used as test data . i m using narx close loop n open loo nets.....

truely speaking i took various values of ID and FD ...1. (1:4) 2.(1:2) 3(1:8) i tried with these delay values but now i cum to know that theses should be choosen through correlation by reading your posts

meanwhile i simulated many times by changing number of hidden neurons 20, 50, [50 30], 80 and 100.

i am really thankful to you .

if u dnt mind i have one more thing to ask can i get some link for time series prediction with deep neural networks

[cheersl cheer]

Hi，SAM,I learned that you have done a lot researches in time series prediction based on deep neural networks,I am doing the same work,I am confused about the form of input .I saw that you wrote "i have temperature data series of total 2825, inputs are 4*2795 and output 1*2795 ",what is the number 4 means? I have read many papers about time series forecasting based DNN , but I just cannot figure out how did they do it in practical. Look forward for your replay

Thank you a lot for your patience and replay in advance

[Greg Heath]

Matrix Dimensions:

 [ I N ] = size(input)   % [ 4 2795 ]
 [ O N ] = size(target)  % [ 1 2795 ]

[cheersl cheer]

Hi Greg Heath,Thank you for your answer. I have some questions about time series prediction,I have read that many papers use the codes of Hinton which is used to classification to do prediction,i am confused .I don't konw how can I use the code of DBN to predicte time series,can you give me any suggestions or links about time series predictions,I am doing the siol moisture prediction. Looking forward for your reply,Thank you a lot for your patience in advance

[SAM]

did you find the solution of time series prediction with DNN, are following the matlab tool box of Deep NN

Answer 3

[SAM]

dear Greg,

i am trying to go through all of your previous posts regarding narxnet and nncorr, please kindly guide me or send me any link of post. because i am not able to understand how to find significant lags autocorrelated and cross correlated btwn output/output and input/output i did

N = length(maxT)% maxT is 2895 samples of output zy=zscore(maxT,1) autocorr=nncorr(zy,zy,N-1,'unbiased') plot(autocorr)

 <</matlabcentral/answers/uploaded_files/29181/untitled.jpg>>

[Greg Heath]

You will have to choose one or two of my posts and ask specific questions. Good search terms to include (not all at once)

 greg nncorr sigthresh95 or thresh95