NARX Closed Loop Network for one step prediction

Hi everyone,

i am using NARX Closed Loop Networks for timeseries prediction. I got a network with 6 input delays and 6 feedback delays, as showed in the figure:

I understand that this network is able to predict y(t) = f(y(t-1),...,y(t-6),u(t-1),...,u(t-6)). If this is true, i should be able to predict one y value having 6 past values from y and u. The problem is, when I use

[xc,xic,aic,~] = preparets(net,input,{},output);

with this dataset of 6 samples, the argument xc (shifted inputs) is empty. I make my prediction:

[yc] = net(xc,xic,aic);

And yc is also empty. As far as I understand NARX Networks, if I'm feeding the network with 6 past input samples and 6 past output samples it should be enough, but it isn't. It needs 7 input values to make one prediction. Why is this?

Thank you for the help.

Answer 1

[Krishna]

Hey Pablo,

I think you need 6 previous values as input and for that the network predicts the 7th value.

Also same goes for the training part.