Multi-step ahead forecasting - Artificial Neural Network for time series

Hi, I have speed data for each 5min speed data during 10days. (10day, 288/day) With this data, I am trying to predict values of y by 12 future time-steps (t+1,t+2,t+3,t+4,t+5,t+6...t+12) with the lastest 6 speed data(t-5, t-4,...,t) for one day. I think NAR function is the best for one parameter(speed).(Input and Output are speed). If I use openloop, I can only find 5 min ahead speed(one step ahead) for one day. Right? I found closed-loop to find multi-step ahead forecasting speed but I cannot find the results for multi-steps. How can I know Multi-step ahead forecasting speeds?? They only give the result for only one step. If inputDelay is 1:6 (this means last 6data?), can I get the prediction results at future 30 minutes? If I want to get by one hour???

9days are testing data and last one day is validation data. Inputdata for tarining: 1x(2880-288) data for validation: 1x288

This is my matlab code.

N=length(historicaldata); % number of samples (testing) Nu=length(historicaldata)-288; % number of learning samples (validation) y=historicaldata;

% prepare training data

yt = con2seq(y(1:Nu)');

% prepare validation data

yv = con2seq(y(Nu+1:end)');

%---------- network parameters ------------- % good parameters

inputDelays = 1:6; % input delay vector

hiddenSizes = [10]; % network structure (number of neurons) %-------------------------------------

% nonlinear autoregressive neural network

net = narnet(inputDelays, hiddenSizes);

% [Xs,Xi,Ai,Ts,EWs,shift] = preparets(net,Xnf,Tnf,Tf,EW)

[Xs,Xi,Ai,Ts] = preparets(net,{},{},yt);

% train net with prepared training data

net = train(net,Xs,Ts,Xi,Ai);

% close feedback for recursive prediction

 netc = closeloop(net);

% prepare validation data for network simulation

yini = yt(end-max(inputDelays)+1:end); % initial values from training data

% combine initial values and validation data 'yv'

[Xs,Xi,Ai] = preparets(netc,{},{},[yini yv]);

% predict on validation data

predict = netc(Xs,Xi,Ai);

Answer 1

[Greg Heath]

You are very much off base.

 0. data         =  training + nontraining 
    nontraining  =  validation + testing + unseen
    data         =  design + nondesign 
    design       =  training + validation
    nondesign    =  testing + unseen

1.The basic difference between openloop (OL) and closeloop (CL):

    a. OL uses (known) target data. Design is relatively quick 
       and simple.
    b. CL uses fedback output data. Design can be relatively 
       long and frustrating.
    c. CL designs are used to predict beyond the target data
    d. Easiest way to obtain CL designs is to close the loop 
       on OL designs

2. Use DIVIDEBLOCK:

   net.divideFcn = 'divideblock';
   a. trainInd = 1:Ntrn 
      i. Calculate the SIGNIFICANT autocorrelation 
         lags using e.g., NNCORR 
      ii. Use above info to determine feedback delays, 
          FD(e.g., FD = 1:d)          
     iii. Given FD, determine the minimum number of 
          hidden nodes, H that will yield a satisfactory 
          OL design
    b. valInd =  Ntrn+1 : Ntrn+Nval 
      i. Prevent overtraining an overfit net via 
          Validation Stopping
      ii. Rank multiple designs via SLIGHTLY BIASED 
          performance estimates
    c. testInd  =  Ntrn+Nval+1 : N
       i. Obtain UNBIASED estimates of performance 
          on nontraining (e.g., test + unseen ) data

3. It will help us help you if you try your code on one or more of the MATLAB example datasets

 help nndatasets
 doc nndatasets

4. I have many posts both in the NEWSGROUP and ANSWERS that explain the details with examples. For example, searching

             NEWSGROUP  ANSWERS

 greg nncorr    14        72
 greg narnet    40       145

Start with the most recent posts and work your way backward. Post all questions. Send new post alerts to my email.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

It looks your understanding of validation and testing data is reversed. Reread what I have written: Validation is part of the design process. Therefore it's results are biased. However, it is not directly involved in estimating weight values. Therefore, it is only slightly biased.

Test data estimates are completely unbiased.

Training, validation and test results are obtained with one call of TRAIN and are stored in the training record TR.