Need machine learning algorithm to fill in time-series data

Question

I am currently dealing with a time-series data set with cyclical gaps every 30 minutes (30 minutes of data, 30 minutes of no data). Is there a relatively simple way of using scikit-learn (or some other Python library) to predict the missing data using the available data to train it? I assume this would involve a 'supervised training' approach? I have provided a graph for visual reference, the orange line is a 20-minute centralized moving average of the gap data set, and I want to fill in the data to look more like the green 60-minute centralized moving average from a separate complete data set. Thanks!

Answer 1

You can approximate a time series with a polynomial of degree n_degree by using ridge regression. You can try different degree numbers (e.g. [2,3,4,5,6]) and choose the best one. Keep in mind that higher degree models, always get lower error values. So you should somehow penalize higher degrees.

from sklearn.linear_model import Ridge
from sklearn.preprocessing import PolynomialFeatures
from sklearn.pipeline import make_pipeline

for count, degree in enumerate([2, 3, 4, 5, 6]):
    model = make_pipeline(PolynomialFeatures(degree), Ridge())
    model.fit(Time, y)

More details here.

Answer 2

What you are trying to do is called imputing (i.e. filling gaps) in the time-series.

A guide to imputing time-series using python is available here:

https://medium.com/@drnesr/filling-gaps-of-a-time-series-using-python-d4bfddd8c460

The guide uses a few algorithms to impute data, but a more extensive list of algorithms can be found here:

https://otexts.com/fpp2/regression.html