Meter to Use for Anomaly Detection

Power Laws
1

For anomaly detection, one of the meter ids in the submission file is ‘38_9686’. In the metadata file, this is called a “main meter” for site 38. This is the time series for that meter:

main_meter

This looks very strange and my thought is that the meter measures cumulative demand and is reset at specific intervals. However, this does not appear to happen at the same time every year and I am having trouble analyzing this meter for anomalies.

In the training data, there is another meter for the same building called “main meter: demand” in the meta data file. The time series for this meter looks like:

main_meter_demand

This is a much better signal to analyze and I have had success using this meter instead of that identified in the test file. I wonder what anyone else’s thoughts are about this meter and how to approach analyzing it for anomalies.

1 Like
2

The meter wraps around once it reaches 999,999. What I did was to look for the discontinuities, then incremented everything after that point by 1*10^7.

3

That’s a good catch. I still think the demand meter might be more useful for finding anomalies than a cumulative total. I’ve tried using both and had more success with the demand meter.

4

I applied your fixed and then graphed the difference between consecutive measurements. This data looks like something I can work with!

correction