Time Series MLP Prediction with exogenous variable

Hi everyone,

I want to do a time series forecast with MLP and exogenous variable. I found MLP node but I do not know how to add exogenous values.

In the attached table are 3 columns: time t, data y and exogenous x

Data y is the history data and I want to forecast about 2 days (109 Datapoints) of y (The empty values in the data y column)

t: is the Time

y: are supporting-values

x: is an Event. 100 means there is an event and 0 means there is no event.

We have seasoanality in Y by a Week (96*7 datapoints) and by a Day (96 datapoints). One Day has 96 datapoints.

A german Description of usinig MLP-Networks to predict timeseries with exogeneous variables is here:




did you already read our white paper about predictive maintenance. There we presented something really similar. The key is the lag column node which enables you to use information from previous data rows.


Anomaly Detection I: Time Alignment and Visualization for Anomaly Detection (2015)

Anomaly Detection II: Anomaly Detection in Predictive Maintenance with Time Series Analysis (2015)

Best regards, Iris 

Thank you for your help.

I read the white papers and with lag columns I can implement MLP Prediction.

But I can't find anything about the MLP predictive modelling with exogenous variables in KNIME.

What can I do?

Best regards, Hermann

This looks like quite the niche topic to me - causality and black-box neural networks still seem "culturally incompatible" to me (see Beiman: http://www.stat.uchicago.edu/~lekheng/courses/191f09/breiman.pdf).

That said, there are people doing it it appears: http://cyrilvoyant.pagesperso-orange.fr/IEE-IC.pdf, so you can surely replicate some of their work somehow in KNIME. No idea how I'd start though, as I've personally pledged for abstention from causal methods a while ago. :)


Hi Ergonomist,

I want to do a time series forecast with MLP and exogenous variable. I found MLP node but I do not know how to add exogenous values to these model.

Question: How is the Implementation of these Case in KNIME?

Hi Hermann,

Sorry, I have only a sliver of an idea, so I'd rather not risk exposing my ignorance. :) Try finding out how authors of other papers have done it!

But again, it appears to be a very niche topic, so you may have to follow these other authors' software choices as well. You can certainly do a whole lot more with KNIME than with many other packages, but this also means that you can treat KNIME much less like a black box. The SPSSs and Statas of this world are more limiting, but can be more forgiving as well. If there's an R script somewhere out there you should be able to "inject" it into KNIME easily for data pre and post processing, but otherwise better go with the flow of the relevant reserach out there.