In case you don’t follow our blog: 2019 Data Science Update.
After a stellar 2018, 2019 is proving to be another year of major data science accomplishments for Sense. Work, thus far, is taking place within two broad buckets. Let’s take a look at them in turn.
The quest for “ground truth”
Here at Sense, we often talk about getting more “ground truth” data. For us, it’s the holy grail that helps us drastically boost device detection accuracy and reliability. If you’re not familiar with the term, in machine learning circles, ground truth refers to an accuracy check that pits machine learning algorithms against the real world. For us, it refers to the reality that we’re attempting to predict via our device detection models. So, if we’re working to detect your refrigerator, ground truth would refer to the actual waveform of your refrigerator that Sense is attempting to disaggregate from the cacophonous cluster of devices in your home. If we know exactly what to look for, it becomes much easier to build predictive models.
Now, you might be thinking: Why don’t you at Sense just buy a bunch of appliances and devices, run them at your offices, and build models from that small set? It’s a common misconception that Sense actually works this way — i.e., once Sense is installed in your home, it compares what it sees to a database of pristine, unchanging device waveforms. If only it were so simple! The same model refrigerator can in fact look quite different in two homes due to variances in its manufacture, other devices that are running concurrently, the physical environment in which it’s running, and even the noisy perturbations that exist on your local utility lines and in your home. While we can’t build a predictive database from just one home, we can seed one from many homes, and for that ground truth data is a huge help.
Ground truth, for us, refers to the reality in your home, and 2019 has seen a big step forward in our ground truth data gathering. 2018 saw the release of the smart plug integration, which has given us great data on the devices users are connecting them to. This began our quest for ground truth. In 2019, we’ve taken it even further. As we’ve continued to digest the data from your smart plugs (keep it coming!), we’ve also released beta integrations with certain smart thermostats. Unfortunately, due to reasons outside of our control, we’re unable to release these publicly at this time, but they have given us fantastic data about HVAC devices across our Beta team that is helping us to build out more advanced models for HVAC detection. We have some other ideas to pull ground truth data from smart thermostats, so keep watching for updates.
We’re also pulling circuit-level data from a small set of internal pilot homes, which is helping to provide granular, ground truth data for circuits that feature just one major appliance — think water heaters, pool pumps, HVAC, electric vehicles, and more. We ultimately don’t believe circuit-level monitoring is the solution to energy disaggregation (the fancy term for device detection), with its complex installation, lack of device-specific insights, and potential electrical code concerns. However, these pilot homes can provide valuable data for us to improve model building.
Ground truth doesn’t always need to be so granular. Our recent Device Inventory feature, part of the Home Details + Compare release, gives us better insight into what’s actually in your home, and will eventually allow our models to key in on your devices and not waste time looking for devices that aren’t in your home. We’ve gotten great data so far, but the benefits will take time to see. Be sure to give us a hand and fill out your device inventory.
Improved historical insights
2019 hasn’t just been about ground truth data gathering. More generally, we are working to better model large energy consumers in homes, with specific focus on HVAC and electric vehicles. For EVs, this includes analyzing our expanding data set to both refine existing models and tackle additional vehicles. We are having great success here and recently released brand new Tesla models that will significantly improve our support for Model 3 detection.
Some of these advances will start to show up in better historical reporting of energy use. Some devices aren’t easy to display as real-time bubbles in the Sense app, but can still be tracked historically. For example, one challenge with electric vehicles is that they have complicated “wind down” patterns as the battery is getting full. It’s hard to show a correctly sized bubble as the slow wind-down is happening in real time, but we can model and present the historical data to you, so you can see precisely how much energy your vehicle used over the past week, month, or year. We’re still deep in our work on this front, but hope to share our findings soon.
Beyond these two major buckets, there is a constant stream of work to make incremental updates to device models — like Mahesh’s recent work on ACs— and hopefully you’re seeing positive results from those efforts.
The future for Sense certainly looks bright. In addition to these major device detection updates, we’re excited to continue work on our partnerships with both Schneider Electric and Landis + Gyr to further integrate Sense into the common energy infrastructure. Be sure to follow us here on the blog and via our newsletter for all the latest Sense news.