The much step forecast with usage of the recurrence neural networks

How it is possible to make the multistage forecast with usage of the recurrence networks?

Answer 1

[Greg Heath]

You have to fill in the details. The main theme is

help/doc narxnet

help/doc preparets

help/doc closeloop

search narxnet greg

Find significant delays from the autocorrelation of t and the cross-correlation of x and t. Create delay row vector inputs ID and FD

Find H by trial and error. First try values 0 <= H <= Hub

Hub = -1 + ceil( (Neq-O) / (I+O+1))

net = narxnet(ID,FD,H);

net.divide.Param = 'divideblock';

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

[net tr Ys Es Xf Af ] = train(net,Xs,Ts,Xi,Ai);

netc = closeloop(net);

[ netc trc Ysc Esc Xfc Afc ] = train(netc, Xs,{},Xic Aic);

plot hold on plot(Ts) plot(Ysc,'r')

trc = trc

Hope this helps.

Greg

Answer 2

[Vadim]

NARNET works remarkably, but for the decision of my problem hybrid NARNET and the RECURRENT GRID is necessary. The forecast for 1 step forward is calculated excellently, but the multistep-by-step forecast does not work. I think the reason in wrong displacement because of delay lines. Believed that someone faced a similar problem, it would not be desirable to look at once in source codes and to understand there.

[Greg Heath]

Can't help without seeing your code. Would help if you used it on one of the timeseries nndatasets so that we can compare results.

help nndatasets

Greg