I am trying to create a neural network using time series as input, in order to train it based on the type of each series. I read that using RNNs you can split the input into batches and use every point of the time series into individual neurons and eventually train the network.

What I am trying to do though is use multiple time series as an input. So for example you might receive input from two sensors. (So two time series), but I want to use both of them in order to get a final result.

Also I am not trying to predict future values of the time series, I am trying to a get a classification based on all of them.

How should I approach this problem?

  • Is there a way to use multiple time series as an input to an RNN?

  • Should I try to aggregate the time series into one?

  • Or should i just use two different neural networks? And if this last approach is correct, if the number of time series increases wouldn't that be too computer intensive?


1 Answer 1


Multivariate time series is an active research topic you will find a lot of recent paper tackling the subject.

To answer your questions, you can use a single RNN. You can input one value for each time step. Nothing keeps you from adding another value at each time step (if your sensor are synchronized). Your model will then learn how to classify with a two dimensional time series.

You check this blog. In your case, only the output is different.

As for the two last points, aggregating the time series into one is risky in the sense that you might lose important information during the process. Finally the main disadvantage of your last point is that you won't be able to use a potential correlation between the two time series for the final classification.

  • $\begingroup$ If you use multiple time series, how will the network react if for some reason for sample1 you have 5 series but for sample2 you have 4, (maybe because you have no data from last sensor). Is it necessary that if you start with 5 series, it should always be 5? Should you include a 5th time series for sample2 with fake averaged data i order to have all 5? $\endgroup$
    – Ploo
    Oct 13, 2017 at 23:42
  • 1
    $\begingroup$ oh well there are different approaches to missing data. I would recommend you to use the value 0 when you have no values. It is often used when we don't have the whole sequence X_t but we still have to input a sequence of length t. It is called padding if you wish to know more about this. $\endgroup$
    – Daerken
    Oct 15, 2017 at 12:58
  • $\begingroup$ Related question: assume that I have two sensors of the same kind that produce two independent, still similar, time series. Example: I have two different temperature sensors in two similar rooms. Is it a meaningful idea to use the training data from two similar time series rather than a single one to train a single network? What if the timestamps overlap? $\endgroup$
    – EM90
    Feb 22 at 12:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.