@Wick, I completely agree.
Some of my biggest loads (pool pump, EV charging) are easy to spot with hourly data. Machine Learning appears to struggle with these, but customers don’t care what algorithm is used to identify them. It seems like Sense could combine different methods for different types of loads and provide a much better customer experience for those of us with very large “other” categories.
@Wick, I completely agree.
Well put Steve. I couldn’t agree more!
Sense folks have already confirmed they are working on the framework for identifying long rise / fall devices (human-visible on and off events), especially EVs. And to be clear, machine learning can also easily identify this stuff, but only if it is looking at the correct time history window and set of features. Building a flexible framework that enables Sense to experiment on and validate a new family of long rise/fall models is critical - happy to see them working on that.
Sense hasn’t disclosed much about what’s “under the hood”. It would help early-adopter confidence as it relates to shortcomings if they did. Many of us are as knowledgeable in the relevant engineering disciplines, and may be capable of providing useful insight. The company explained they are challenged detecting longer timescale signatures. Hypothesis: It’s possible their entire architecture is predicated on a false assumption - that most devices can be detected by downloading to the sensor short time constant trained signatures. If so, the devices in the field may not be up to the task for devices with erratic unpredictable short segment signatures. Those may only be recognizable from longer windows tolerant of a lot of short time constant noise, and the box may not be able to do that. It does have, after all, finite speed and memory. And real time central server processing may not be economically feasible for them.
From what I know about machine learning and LTSM, I’m not so worried about technical feasibility. Just hard to be patient. I am looking forward to the August briefing.
Let’s imagine, hypothetically, they designed the box to look for signatures no longer than n milliseconds. Let’s also imagine along comes Tesla and it’s got thousands of different such, maybe an unlimited number depending on state of charge, charge limit set, vampire load, age of battery. It’s a complex device. Now macroscopically it’s signature is as plain as the nose on your face. Easily detected by general machine learning. But maybe not so easy if the box lacks the processing power or memory to look for it while running their main strategy. Just one hypothesis that fits the facts. Would be nice if Sense explained it rather than just telling us to be patient. Obvious there’s a shortcoming somewhere re Tesla. Just tell us what it is.
Sense box identifies my Model S like clockwork. So your thesis is off the mark. Only problem is that the identification doesn’t generalize to my X, 3 or Ford Fusion Energi. Just show up for the August briefing - you’ll know more than you know now.
@kevin1-- Encouraging! Does it get the charge ending time correct as well? (Better yet, does the kWh total for a charging session match what your car says?)
I was wondering if Sense could analyze two different “granularities” of the data in parallel for big, obvious loads – one at the micro level (~1MHz) and another at the macro level (1 per hour).
Sense gets the power and duration correct. Car is set to charge at 1am. 20KW charging aligns with 80A dual charger in the S. But apparently Sense has only done one-offs for 3 different car/charger combos.
Part of the challenge is that we’re dealing with some proprietary methods here. I recommend joining the Webinar if you haven’t already and getting some specific questions together: Customer Webinar - 8/17/2018
Tried to sign up but not possible. Advice, Ryan?
It looks like it’s full up. We have to limit capacity to better manage questions and for technical reasons. I’m hoping to expand it a little bit, so watch for news in the Webinar thread.