When do you detrend data or remove the mean?

I have a question regarding data sets and how to make a decision about detrending data or removing the mean. These two options can be used in the identfication toolbox and I wonder for which cases you might use these options?

For example I have a big data set on temperature values of a heat process in the industry and I want to identify a model against validation data.

Would it be a good choice to detrend the data? I tried a few times, but all the models slightly worse, most of the time.

Answer 1

[Star Strider]

Both detrending and removing the mean are essentially removing low-frequency baseline variations or a constant offset, respectively (or both). If these are not part of the process you are modeling (so are instead properties of the data acquisition equipment for instance) removing them would be useful.

For both of these, I generally use a highpass or bandpass digital filter (bandpass also to remove high-frequency noise) on the data first, since the filter design permits filtering the the mean as well as specifying the low-frequency cutoff. There may be advantages to using detrend for this, however for me, using a filter is easier. Filters are also relatively straightforward to design in MATLAB.