Q&A: A behind the scenes look at AC detection

This seems like a blogpost you all will appreciate: Behind the Scenes: Tracking the Elusive Air Conditioner

Data scientist, Mahesh, talks about the challenges of detecting ACs and how we’re working to meet that challenge (and yes, that includes variable-speed ACs).

@MaheshAtSense has also graciously offered to answer questions folks here might have about AC detection. Please, for his sanity, keep the “Why won’t Sense detect my AC?” questions to Support tickets. Let’s focus on questions about how the process works instead.

1 Like

@MaheshAtSense

Nice intro to the issues. Thank you.

I’d like to prod a little deeper …

Regarding mini-splits, inverters and particularly:

" … we’re exploring ways to match outdoor temperature data to the home’s energy usage. This approach could yield good results for monthly usage data for HVAC systems with identical ACs or mini-split systems."

  • LG and others are now making window inverter ACs. I just installed one. Just to make your life difficult.

  • Some window units have resistive element heating modes.

  • Industry will no-doubt introduce inverter-based compressor heating modes into window units and they will, hopefully quickly, supersede old-school ACs.

  • Many (most) existing mini-split systems also heat … so the heating mode, in your “weather” match, shouldn’t be ignored!

  • Because weather (temperature & humidity that is) is hyper-local (especially in an urban environment) are you looking in to ways to integrate user-based monitoring?

  • Are the cross-overs with Solar output as a weather indicator being considered? Clouds! [The integration here seems key] @kevin1 has done some great exploration on this.

  • BTW: Sump-pump goes on = rain?!

8Great article @MaheshAtSense, and very timely. I have a couple questions driven by my recent experiments.

  • It seems like it is “easier” for Sense to detect on and off events if the on-cycles are relatively short (less than 30min) vs. long (4-5hours continuous). Is that really the case ? If so, why ?

  • You mention a compressor fan, but I really see two fans in a central air system. The furnace blower that pushes the cool air through the house and the compressor fan that circulates air through the outdoor compressor unit. Which one were you talking about in your article ?

  • I have also seen device identifications move around based on season. For instance device AC is associated with my downstairs unit during the spring, but the association moved to AC2 come the late summer. Is that typical ?

BTW - I have investigated my upstairs and downstairs Ecobee AC settings - they are almost exactly the same. The big difference comes purely from the relative differences in compressor size, the location of vents (upstairs in ceiling, downstairs in floor) and the design of our house (big open stairway in the center). I believe that a big percentage of cool air from upstairs flows downstairs and actually cools the downstairs.

Upstairs AC

Downstairs AC

1 Like

Thanks for the reply, @ixu !

LG and others are now making window inverter ACs. I just installed one. Just to make your life difficult.

Some window units have resistive element heating modes.

Industry will no-doubt introduce inverter-based compressor heating modes into window units and they will, hopefully quickly, supersede old-school ACs.

Variable-speed motors are indeed becoming increasingly prevalent! It is exciting that companies are motivated (to whatever extent) to produce more efficient ACs, especially as the world’s cooling needs are monotonically increasing! Even in colder regions, these systems sometimes double as heat-pumps (both inverter-driven and constant speed) and can be efficient alternatives to other heating systems.

This reminds me of another tricky issue in accounting for the seasonality of AC models: lots of people with heat-pumps use the same system in winter (for heating) and summer.

We have done some work in the past modeling inverter-driven fridges. Some of that work might transfer to LG inverter WACs in the future. We haven’t had confirmed examples of users owning this device! Would it be ok if we looked at your data? (tagging @RyanAtSense )

In general, large local variations are more informative to machine learning systems than slow long-term changes.

Side-note: Features defining boundaries of shapes are generally important in image recognition. Thinking in (limited) analogies, constant speed ACs are well defined by the rising and falling edges, while variable speed motors are like a fuzzy blob of unknown extent and fuzziness.

We are continuing to work hard at the problem of recognizing slowly changing shapes! (EVs, inverter-ACs, etc.)

Because weather (temperature & humidity that is) is hyper-local (especially in an urban environment) are you looking in to ways to integrate user-based monitoring?

A great solution here would be for Sense to integrate with Ecobee and Nest. Various external constraints beyond our control (mainly related to access to the smart-thermostat APIs) have stymied our ability to integrate with these devices. Another benefit of thermostat integrations would give us sources of ground-truth for 240 V devices!

We are currently exploring other approaches of getting weather data. (Side note: NOAA’s temperature data is great source for local weather, https://www.ncei.noaa.gov/data/global-hourly/access/ - but as you might imagine, it is not straightforward to get this as a real-time stream)

An additional challenge with directly using temperature for modeling real-time AC behavior (i.e., going beyond average estimates) is that there are a lot of confounders and a long tail of outliers: ACs cycle long, some cycle more often, some houses have poorly specced systems, etc.

Are the cross-overs with Solar output as a weather indicator being considered? Clouds! [The integration here seems key] @kevin1 has done some great exploration on this.

We have done some exploration with temporal and spatial solar output trends. But we haven’t used it for weather modeling. That’s an interesting idea!

BTW: Sump-pump goes on = rain?!

Likely, but not necessarily? I wonder if there are other sources of water? It would be interesting to look at correlations between cloud cover and sump-pump usage.

(I was trying to flesh out the correlation between sump-pumps and rain, and remembered this story, though it doesn’t explicitly mention rain: https://sensesaves.sense.com/dirty-sump-pump/ )

1 Like

It seems like it is “easier” for Sense to detect on and off events if the on-cycles are relatively short (less than 30min) vs. long (4-5hours continuous). Is that really the case ? If so, why ?

It is possible in certain cases, though not as a general law of device modeling :slight_smile: . Models can sometimes contain a shorter “time-out” parameter than appropriate for your device. It could also happen that, once the algorithm finds an AC “on”, the longer it has to wait for the matching “off”, the more the likelihood that it will encounter noise and falsely associate a different “off” or fail altogether.

We are working on improvements to the detection system and to our device-modeling to better estimate time-outs and juggle hypotheses.

You mention a compressor fan, but I really see two fans in a central air system. The furnace blower that pushes the cool air through the house and the compressor fan that circulates air through the outdoor compressor unit. Which one were you talking about in your article ?

We refer to the compressor fan as just compressor and the air-handler/blower fan as “the fan.” Some air-handlers are 240 V and some are 120 V. Seems like most run in sync with the compressor and some run all the time.

I have also seen device identifications move around based on season. For instance device AC is associated with my downstairs unit during the spring, but the association moved to AC2 come the late summer. Is that typical ?

We have seen this happen with ACs. Some of these issues are due to the general problem of model evolution and associating components: https://blog.sense.com/behind-scenes-lost-detection, and some are due to subtle seasonal differences in how ACs and heat-pumps function. We have made some progress in mitigating duplication (especially obvious duplications) and association issues over the past year and continue to work on it. Hopefully, you won’t see the association switch going into the next season.

1 Like

Great answers ! Thanks. Will let you know if the detections are better established this season, though it might be difficult since I replaced both units mid-season (July 9th). Might have to wait until next summer to be sure.

@MaheshAtSense,

One more question on behalf of many of us who have either had their AC units repaired or replaced this season, some courtesy of Sense early detection. What should we do after we complete the repair / replacement ?

  • Inform someone at Sense so they can tag / label and archive the signature of a failing device ?
  • Delete the previously detected device(s) ?

Those questions seem to be coming up frequently this summer :wink:

  • Inform someone at Sense so they can tag / label and archive the signature of a failing device ?
  • Delete the previously detected device(s) ?

You’ve summarized it well, @kevin1! These seem to be the right things to do.

We will likely detect your device even if you don’t delete the old models, but deletion will likely improve the user experience.

1 Like

Thanks Mahesh !
I have been experimenting to see how long it takes for previously detected, but now gone AC detections, to stop. Seems like deleting might help.
https://community.sense.com/t/sense-vs-ecobee-round-ii/6945/6

1 Like