Training Data for NARX

Hello,

I'm working on a university assignement where I had to create a model predicitve controller from scratch and generate training data to train a NARX neural network.

I have completed the first part and my data consists of an Input matrix U [1x25] and the resulting state (output) X [2x26]

The number of rows is arbitrary and changes according the number of time steps I chose.

The output has one extra row because of the initial state X0

My question is how do I proceed to create a valid dataset to train a NARX network? How do I get from my controller results a compatible data structure ? How should it look like ?

I'm really clueless because I'm new to this and very much willing to learn more.

Any feedback is appreaciated.

Answer 1

[Tejas]

Hello Wissal,

To create a training dataset for a NARX model, follow these steps:

• Convert the input matrix 'U' and the output matrix 'X' into cell arrays. Make sure there is a one-to-one correspondence between the input data and the output data.

    U_cell = num2cell(U, 1); 
    X_cell = num2cell(X(:, 1:end-1), 1);

• Prepare the NARX network according to your requirements. Below is an example code snippet for guidance on this process.

    inputDelays = 1:2;
    feedbackDelays = 1:2; 
    net = narxnet(inputDelays, feedbackDelays, 10);

• Use the 'preparets' function to form the training dataset. For more information on this function, refer to this documentation: https://www.mathworks.com/help/releases/R2022a/deeplearning/ref/preparets.html .

    [Xs, Xi, Ai, Ts] = preparets(net, U_cell, {}, X_cell);
    [net, tr] = train(net, Xs, Ts, Xi, Ai);

[wissal zaher]

Thank you :)