1
$\begingroup$

I have a large dataset of signals (composed of time series). All time series describe the same process, but each series has a different duration (number of points). Based on these time series, I want to train some neural network, so that then I give a new time series as input and it predicts 100 further points. I have two questions:

  1. What transformations are there to reduce all signals to one size?
  2. What are the methods for solving such problems (predict time series)? I know the popular ARMA and ARIMA models, but they work with the same time series. The goal of my task is to find patterns between time series in order to learn how to predict the further behavior of a new time series. Thanks for any help!
$\endgroup$
1
$\begingroup$
  1. I don't think you can compress time series because there is a risk of losing valuable data. Rather than that, you can set a the max size as the default size, and set zeros to the left for smaller data.

If the sampling is too high (ex: milli seconds), do not hesitate to reduce it for all data (ex: seconds) taking the average values, as long as the prediction objectives allows it. Furthermore, the further you want to predict, the worst the prediction generally is: that's why a lower sampling rate could be useful.

  1. RNN and LSTM are also good solutions, in addition to ARIMA. However, they are quite sensitive to noise: if your signals are quite noisy, try to reduce the noise to have good predictions.

Keep in mind that time series prediction with NN is not an exact science: you may have to apply many modifications and improvements on your data to reach very good results.

Here is a notebook that could be useful:

https://github.com/ageron/handson-ml2/blob/master/15_processing_sequences_using_rnns_and_cnns.ipynb

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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