Bad accuracy of one step neural network prediction

Hi all,

recently I am doing some one step predictions using neural networks, however, the prediction results are really bad (the network itself is trained very well).

The NN I used is a recurrent NN ( narxnet ) with input u(k) and feedback input y(k), and both input and feedback delays are 1. Here my prediction is a little bit different that, for each prediction y_pre(k+1), I just use u(k) and y(k), instead of using some former data series like [u(k-1), u(k-2), ...] and [y(k-1), y(k-2), ...]. In principle, I think it should work because all delays are 1, but the prediction results are really bad. And another weird thing is that, the more steps I predict, the higher accuracy it has. For example, the predicted y_pre(k+10) is much more close to the real y(k+10) compared with the case of y_pre(k+1) and y(k+1).

I am not sure why this happens, but this phenomenon reminds me that, when I use sim function, the first couple of simulation results are really bad if I use a bad initialization, but after some time it will be recovered. I don't know if there is any thing to do with the initialization stuff or if there is something I did wrong. I appreciate it so much if anyone can help me solve the problem.

Thank you very much!

PS: sorry I didn't put my codes here, because actually I think the whole procedure is pretty normal. I skip the configuration and training parts and just put the prediction part here. The codes I used are:

U_k = repmat(u_k,2,1);
Y_k = repmat(y_k,2,1);
input = tonndata(U_k,false,false);
target = tonndata(Y_k,false,false);
[xs,xi,ls,yt] = preparets(netc,input,{},target);
Ys = netc(xs,xi,ls);
ys = cell2mat(Ys);

Here, u_k and y_k are all row vectors, such as

u_k = ones(1,num_in);
y_k = ones(1,num_out);

I expect this ys can be the prediction of y(k+1) based only on u(k) and y(k).

[Greg Heath]

No one can help without code and/or data.

[Yiming]

I edited my question and put some codes here. Thanks for the help!

[Greg Heath]

"Some" codes is insufficient. I'm not a mind reader (although I would like to be ... I think).