Categorization algorithms, take 1

With data in hand (66 events over 20 minutes from one data set with RMS volts for normalization purposes) we started post-processing the data. This will convert the raw values (of power, reactive power, etc.) to more useful metrics (of delta power, delta reactive power, etc.).

Step one was to map the human-tagged events to more precise timestamps based on transitions in real power. We took the set of all events and selected the nearest timestamp with a change in real power of greater than 5 watts, provided it was less than 4 seconds away from the event. This did a nice job of locating the events at a more precise time relative to the actual change in power, since the light switch and LabView click are not quite synchronized.

Next we wanted the changes in real power at each one of these event times. The sample rate in this data set is 20 Hz (50 ms intervals) so it's not helpful to look at inter-sample changes (though the rate of change might be useful, this method is both susceptible to noise and also will fail to capture the full jump from off-power to on-power). Instead, we discovered (with a little trial-and-error) that averaging the power from all samples in the interval 4 seconds before the event and calling that the before-power, then subtracting the after-power averaged from the 4 seconds after the event gave reasonable intervals. It was able to absorb relatively short start-up transients, but the microwave has a transient with a large spike, then a more gradual descent down to the steady operating power. This resulted in a start-up power change of +300W but a turn-off power change of -90W; reliable for comparing among on-events, but not so good for pairing on-events and off-events.

While none of the appliances we have been testing so far have similar real power consumption, we noticed that the difference between the high and low states of the fan was about 10W, similar to the difference between the on and off states of the CF bulb. We hoped that using reactive power as a secondary axis would help to differentiate those two events, as it did for Hart (1992). However, when we applied the same transformations to the normalized reactive power metrics, we found little correlation between delta-Q values within events of a given appliance transition (eg. fan changing from high to low). In fact, for the incandescent lightbulb on-to-off transition we saw +3W twice and -12W once! Investigating this oddity (and remember that such a resistive load should effect no change in reactive power at all) we discovered that the -12W delta Q transition occurred during a microwave on-cycle. This produced about 40VAR fluctuation in reactive power, so the changes we were seeing were entirely due to the microwave noise, not the light turning off.

For harmonics we are starting to filter the extracted tones to exclude frequencies outside of +/-15Hz of the 1st, 3rd, and 5th harmonics. Then we will look at the normalized 3rd and 5th (e.g., the ratio of the Nth harmonic to the 1st, or fundamental frequency) and the changes to those values around the event times. So far it is clear that some of the harmonics drop out entirely during periods that are visually correlated with certain appliance events. Hopefully the harmonics will help us to filter out some of the noise as it has helped others to isolate continuously variable loads.