Narxnet Multistep forecast within the time series

Hi all,

I've been reading these forums for a while now, and they've been quite useful. Nevertheless, now I'm in front of a problem that I haven't been able to solve, unfortunately, and I cannot find the answer in these forums.

I'm currently working with a Narxnet to predict climatology. Previously, just by using narnet I had a code that allowed me to do a forecast anywhere within the time series (E.g. If my trained time series has 100 elements(t=100), I was able to see what would happen with the closed loop at t=20, 25, 30, ... Basically, to assess the forecast at any given point within my time series and not necessarily at the end (t>100)). Then I changed my ANN to Narxnet but I've encountered the problem that I cannot make a prediction using the closed loop if the delays of my Inputs and Targets are different. Let's say I want to use the previous 5 days to predict Sea Surface Temperature but just using the previous 2 days of Heat Flux information as Input.

Is it possible to make a multi step prediction using different delays for inputs and targets? Or do I always have to consider same size cell arrays for input and targets when using preparets?

Thank you all,

Cesar

Answer 1

[Greg Heath]

1. NARXNET and TIMEDELAYNET require:

All input delays to be increasing (not necessarily consecutive) and NONNEGATIVE. All inputs have the same delays.

2. NARXNET and NARNET require:

All feedback delays to be increasing (not necessarily consecutive) and POSITIVE. All outputs have the same delays.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Cesar Quilodran Casas]

Thank you for your answer. It solved one of my main doubts. If you don't mind I have another tiny question but I didn't want to open a new thread to ask you about this.

In a previous post Multi step ahead predictions using NARNET http://uk.mathworks.com/matlabcentral/newsreader/view_thread/342655

you mention that if we want to get predictions beyond the target, then we need to input an array of empty cells:

Prediction=cell(1,N)

with N, being the amounts of time steps we want to predict ahead. I can see that this example works when we have one dimension, nevertheless, I'm currently working with 100 EOFs, and therefore this won't be a single vector problem. Do you know what is the input for empty cells in this case? It seems that matlab is waiting for me to give it a cell array of 1,N where each element of the cell has a shape of 100x1.

Thank you for your time and help, it's highly appreciated

César

[Greg Heath]

Isn't it obvious that the first dimension of the cell must be the first dimension of the orignal target?

Greg

[Cesar Quilodran Casas]

Dear Greg,

I tried that option too, but then I receive:

 ??? Error using ==> network.sim at 121
 Number of inputs does not match net.numInputs.true

I used the pollution dataset from the Matlab examples, and the following is a short version of the advanced script from the Neural Network Toolbox

 load pollution_dataset

 inputSeries = pollutionInputs;
 targetSeries = pollutionTargets;

 inputDelays = 1:2;
 feedbackDelays = 1:2;
 hiddenLayerSize = 10;
 net = narxnet(inputDelays,feedbackDelays,hiddenLayerSize);
 [inputs,inputStates,layerStates,targets] = preparets(net,inputSeries,{},targetSeries);

 %Closed Loop
 [net,tr] = train(net,inputs,targets,inputStates,layerStates);
 netc = closeloop(net);
 netc.name = [net.name ' - Closed Loop'];

 [xc,xic,aic,tc] = preparets(netc,inputSeries,{},targetSeries);
 yc = netc(xc,xic,aic);

 %Multi step Prediction
 xc2=cell(3,50) %3: Size of the targets; 50: prediction time steps ahead
  yc = netc(xc2,xic,aic);

In the last line of code is where I receive the error, since I'm not completely sure what xc2 should be in order to make the multi step prediction ahead.

Thank you,

César

[Greg Heath]

% Narxnet Multistep forecast within the time series % Asked by Cesar Quilodran Casas on 18 Apr 2016 at 17:40

 close all, clear all, clc, load pollution_dataset
 inputSeries = pollutionInputs; targetSeries = pollutionTargets;
 inputDelays = 1:2; feedbackDelays = 1:2; hiddenLayerSize = 10;
 net = narxnet(inputDelays,feedbackDelays,hiddenLayerSize);
 [inputs,inputStates,layerStates,targets] = ...
            preparets(net,inputSeries,{},targetSeries);
 [net,tr] = train(net,inputs,targets,inputStates,layerStates);

GEH1 = 'Well, is it a good design? If not, you have to do it over ' ...
              'before closing the loop!'

 %Closed Loop
 netc = closeloop(net);
 netc.name = [net.name ' - Closed Loop'];
 [xc,xic,aic,tc] = preparets(netc,inputSeries,{},targetSeries);
 yc = netc(xc,xic,aic);

GEH2 = 'Should have [ yc xfc afc ] = ...
GEH3 = 'Well, is it a good design? If not, you have to do it over ' ...
               'before trying to predict beyond tc(:,end)'

 %Multi step Prediction
 xc2=cell(3,50) %3: Size of the targets; 50: prediction time steps ahead
%   yc = netc(xc2,xic,aic);
 % In the last line of code is where I receive the error, since 
% I'm not completely sure what xc2 should be in order to 
% make the multi step prediction ahead.

GEH3 = 'WHOOPS! I told you target dimension instead of input dimension!'
GEH4 = 'Also need above final values to initialize the new prediction'

xc2 = cell(8,50);
xic2 = xfc;
aic2 = afc;
yc2 = netc(xc2,xic2,aic2);

GEH5 = 'Need numerous plots for a decent understanding of the problem'

Hope this helps.

Greg

[Cesar Quilodran Casas]

Dear Prof. Heath Thank you for your fast response. I've tried your modified code but even if I put

xc2=cell(8,N)%8 is the dimension of inputs and N the forecasted time steps

I keep getting the same error. The only thing that works for me is

 xc2=cell(0,N);

where N is the number of desired forecasted time steps. I got this idea from one of your previous posts. I suspect that this might not be the optimal answer since Matlab is expecting me to give any inputs. But in my case I want a complete "free running" model, not even giving inputs and completely targetless.

Cheers,

César