Multivariate Multi-step Time Series LSTM Model

KNIME Analytics Platform
1

Hi,

I want to know how to configure a LSTM time series model to forecast multi-steps ahead instead of a single-step.

@tim_19801 I read your post about a similar question in 2023. Have you found the solution?

This is the workflow I’m interested in convert to a multi-step problem.

Thank you!

Carlos

2

Hi Carlos -

Not sure if you’re still stuck on this. I asked one of our data scientists about it. His response:

I’m fairly sure the change would just be to the output shape of the LSTM layer. I believe in our examples that’s the dense layer that follows it, from (1) to (3). The training data would need to be reshaped accordingly to provide a target column that was of that shape - could do that with a lag column node.

If that’s not sufficient, let me know and I can do a bit more digging.

3

Hi Scott,

Thank you for your feedback! I really enjoyed your YouTube short. When can we expect a new one?

I definitely need your help digging a little deeper. I found some Python websites that confirm what you mentioned about the dense layer. They explain that each step in a multistep problem corresponds to each neuron in the dense layer. I tried to implement this in a workflow, but I only had partial success. I struggled with generating the number (N) of future steps/values of interest. Instead, my implementation produced N columns as output, not N values.

It would be amazing if you could create a template workflow for a multistep problem using an LSTM network. The whole community would greatly appreciate it.

Thank you!

Carlos

1 Like