How to train a NARX with multiple datasets and constant external inputs in every dataset?

Dear all,

I am trying to develop a NARX for my problem (following many of the useful suggestions that I found here). The problem that I want to simulate consists of a system having two external inputs, without a delay, and one feedback input with a delay of one timestep. The fact that the system should predict the output at the next timestep given the output only at the previous timestep and the inputs at the current timestep is a requirement.

You can imagine it as a battery, in which the input is the amount of energy extracted at the current timestep, and the feedback input is the state of charge at the previous timestep. I want to build a NARX to simulate only the discharge behavior of this battery.

I am training my NARX with a set of experiments (in a for-loop) in which every experiment has the external inputs constant (i.e. a constant amount of energy extracted every timestep), and the feedback input varies progressively from 0 to 1 (not completely linearly). The idea is that, given a set of (constant) external inputs, the output goes from 0 to 1 within a given amount of timesteps. However, during the training, I see that in the trainng interface the external inputs are disconnected from the first hidden layer. Does it mean that they are automatically excluded because they are constant? Is there a way to avoid that without training the network with experiments in which the inputs are changing during a single experiment (training dataset)?

Thank you very much for your help!

[Marco Pizzoli]

Hello,

Have you by any chance solved this problem?

Because I recently encountered it too and would like to solve it.

Thank you very much.

Marco

Answer 1

[Luke135]

I found the answer: https://nl.mathworks.com/help/nnet/ug/multiple-sequences-with-dynamic-neural-networks.html