CODEWITHSUNDEEP
pythonmachine-learningtime-seriesprediction
Zikry Kickbuttowski
Zikry Kickbuttowski

Reputation: 11

How can I use machine learning for time series problem

Hello I have a time series data that basically behaves in a sawtooth manner. After each maintenance period, the signal always goes up before going down until a maintenance happens which will cause the signal to increase again. I am trying to predict the signal and see what happens to the signal if I schedule future maintenance.

I am new to time series and I am not sure which model should I use to predict the data. I have looked at cross correlation but it doesnt seem to take into account any events that will influence the signal like my problem.

I just what what happens after each maintenance event and the signal follws a similar trend all the time after each maintenance period where it goes up and down. Any suggestions?

Upvotes: 1

Views: 97

Answers (1)

theletz
theletz

Reputation: 1805

You are looking for a ML model for time-series data. This is a huge field but I`ll try to write a few important notes:

  1. Try to generate a dataframe where each row is a timestamp and each column is a feature.
  2. Now you can generate rolling features - for example rolling mean/std of your features, using a few different time windows.
  3. split your data to train and test - this is a very tricky part in time series data. You should be very careful with this. You have to split the data by time (not randomly), in order to simulate the real world where you learn from the past and predict the future. You must verify that you haven't a leakage - for example, if you use as a feature "rolling mean of the last week", you must verify that you didn't calculated your signal for validation using data from the train set.
  4. Train a baseline model using classic ML methods - for example boosting trees etc.
  5. In the next steps you can improve your baseline and then continue with more advanced models (LSTM etc)

Upvotes: 2

Related Questions