Wich feature can i use to classify timeseries?

[muriloasouza]

I have a dataset with N timeseries and 24 hourly measurements of electricity energy consumption (a single day). So, it is a dataset with N rows and 24 columns. Consider these samples as class 0.

I need to generate more N rows, multiplying each of those rows of my original dataset by a random reduction factor between 0.1 and 0.9. These new samples are class 1.

The problem is to solve this timeseries classification problem. One sample from class 0 that has a low energy consumption pattern, could be missclassified as class 1 (since class 1 timeseries has the same format/shape as class 0, but with the reduction factor applied). Wich feature could i use to improve my model predicton?

I have used the mean value wich improved a bit, but still far from acceptable. Anything else i could try?

[dmfolgado]

As a general guideline, the features depend on the problem at hand; therefore, possessing domain knowledge about the problem and its associated data is particularly important.

Based on your description, I recommend exploring features from the statistical set, especially as you are generating new samples for class 1 by scaling the data