load forecasting using NARNET

Hi folks,

I am using narnet to forecast one step ahead of power system load. I am using NARNET wizard. First of all, let's give it an array of ten elements and predict 11th element.

if true
  targetSeries = tonndata(yyt,true,false);
  %yyt is an array of ten elemnts:
[749.8398  739.8419  719.8462  709.8483  689.8526  709.8483  739.8419 759.8376  709.8483  739.8419]

   feedbackDelays = 1:9;
   % I am confused here, what the f**k is going on with this delay thing? if I give it 9, it forecasts the 10th point which I already have it. So I thought I would need to remove one delay to get one more forecast. Am I right?

   hiddenLayerSize = 10; %whatever
   net = narnet(feedbackDelays,hiddenLayerSize); %whatever
   [inputs,inputStates,layerStates,targets] = preparets(net,{},{},targetSeries); %same sh*t
   net.divideParam.trainRatio = 70/100;
   net.divideParam.valRatio = 15/100;
   net.divideParam.testRatio = 15/100;

   [net,tr] = train(net,inputs,targets,inputStates,layerStates);% Apparently training

   outputs = net(inputs,inputStates,layerStates);
   %This should give me next value of my series, the dumb net just gives me the 10th point I already    entered

   %How about this?
   net = removedelay(net,1);
   %Now if I train net and get the output, it gives me a 1*2 cell. Clearly, I will go with the second point as my forecast, right?

end

Alright, now let's look at what I got out of the script:

Actual 749.8397827 739.841938 719.846189 709.8483443 689.8525953 709.8483443 739.841938

One step ahead Forecast 750.0746601 749.8397827 740.2006302 719.8708696 708.7710368 689.8413695 717.4414728

Look at the bold numbers. Do you see the shifting pattern? it seems that my two step ahead forecast and actual value are almost equal. However, I do not have two step ahead forecast(e.g forecast(t=2)) unless I have the actual load of one step prior (actual(t=1)). What is going on here?

Answer 1

[Greg Heath]

You do not have enough data (N=11) to do anything useful. Use one or more of the MATLAB example data sets

 help nndatasets
 doc nndatasets

[ O N ] = size(target)
 NFD    = No. of feedback delays
 H      = No. of hidden nodes 
 Ntrn   = N-2*round(0.15*N) Default No. of training examples
 Ntrneq = Ntrn*O % No. of training equations
 Nw     = (NFD*O+1)*H+(H+1)*O % No. of unknown weights

Want Ntrneq >= Nw. Prefer Ntrneq >> Nw

With your numbers

 O = 1, N=11, NFD = 9, H = 10, Ntrn = 8,Ntrneq = 8, Nw = 10*10++11 = 111

If you use all 11 for training, no hidden layer (Just a linear 1-D feedback model) and NFD = 1

 O = 1, N = Ntrn = 11, NFD = 1, H = [], Ntrneq = 11, Nw = (1+1)*1 +1 = 3 (1 feedback and 2 biases)

 net = narnet(1,[]);

I suggest you get more data (e.g., a MATLAB Example dataset.

Hope tis helps.

Greg