Multivariate time series

KNIME Analytics Platform
1

Hello everyone,

I am reviewing a multivariate time series example divided in two workflows: training and deployment created by @Kathrin and I would like to know if you can help with the following question:

How I can change the forecast horizon from the next hour to the next 24 hours, for example?

You can find an explanation of the workflow here.

Thank you so much in advance.

Carlos

2

Hello @CarlosEnrique84,

Thank you so much for the post. I am not unable to understand the question, do you mean forecasting for 24 hrs or daily forecasts? Currently, the deployment workflow forecasts for 10 hours, you can split the data set such that the test data is for 24 hrs in the available date.

If you mean to forecast for very 24 hours, then you will have to process the data such that the values are aggregated for every day. You can use, Date&Time Aggregator from the time series analysis extension to aggregate values on daily basis, and apply the workflow from the example to make daily predictions.

Best,
Ali

2 Likes
3

Hello @aliasghar_marvi,

Thank you for your response. I believe there was some confusion regarding my question.

I am referring to the KNIME workflow described in the blog post Multivariate Time Series Analysis: LSTMs & Codeless. In the current setup, the deployment workflow uses a many-to-one LSTM architecture to predict the demand for bikes for the next hour (a single time step ahead).

What I want to do is extend the forecast horizon so that the model predicts 24 time steps ahead (i.e., the demand for bikes for the next 24 hours) while keeping the same hourly resolution.

Could you please explain how to modify the workflow to accomplish this?

Thank you in advance for your help!

Best regards,

Carlos

4

Please, consider what the article says about the workflow:

“This case study, can therefore be framed as a supervised learning problem . The past demand, i.e. the count of the new bike shares, and the past values of the other features over the last 10 hours represent the input. Our target is to predict the demand for bikes in the next hour.”