It has been a year since Sense found my water heater. I’ve continued to monitor its accuracy, and my initial conclusion quoted above is still valid. Below is an updated version of the plot comparing the native detection to a Kasa smart plug.
There are some interesting features in this graph:
- The appliance uses more energy in the winter than the summer. This confirms the trend in my post above (#3 of the current thread). The basement was warmer this winter than last because I set a new space heater at 60°F, but even so the event count is higher in December through February, which correspond to weeks 22 through 34 in this graph.
- At week 20, the color yellow becomes more prominent, meaning more Missed Off events. Apparently this is when the AI engine made a tweak to the detection definition, which it does from time to time with the objective of improved device detection. In this case, it hurt rather than helping.
- At week 44, red becomes prominent again, meaning more Missed On events. This definition tweak by the AI engine improved the Off definition but nearly ruined the On definition.
The current performance with so many Missed On events means the native detection is nearly useless to me. I could delete the device, but it doesn’t hurt anything and the AI engine might someday get it right.
I wish Sense used smart plug data to directly improve their AI engine (they currently use it only in the aggregate). Imagine how quickly they could refine device definitions if they associated a specific smart plug with a specific native detection, as I did here. People often ask if smart plugs can teach Sense: perhaps someday they will!