An ordinary smart meter gives your local utility useful information about how much energy you are using—every hour, or even as often as every minute. This helps utility planners efficiently adjust electricity generation to meet demand or encourage reductions in demand when necessary.
But machine learning systems, looking at that data, can tell something else about your home besides its energy use—they can tell if you are home, or if you are not. That’s what University of California at Berkeley researchers Ming Jin, Ruoxi Jia, and Costas Spanos found out. That information, Jin says, is also useful for utilities—they can call or show up to perform necessary maintenance when you are home, and not waste personnel time trying to reach you.
But they aren’t the only ones who can access this information, given the data is transmitted wirelessly, and isn’t necessarily encrypted at every stage of its journey.
“If you know a person is home, as an advertiser, you can make a phone call. If you know a person isn’t home, that information could be used for home intrusion or other bad activities,” Jin says.
In a recent paper, Jin and his colleagues demonstrated that machine learning systems can be trained to detect occupancy without any initial information from a home owner. “You just need a smart meter that listens over time,” he says, “as well as the basic assumption that different types of buildings have different occupancy patterns, for example, commercial buildings are typically occupied during the day and not the night and homes are the opposite.” Using this assumption, the machine learning algorithms were able to tease out more detailed characteristics about power consumption when a home is occupied; they then are able to tell when someone is home or not, even when that person’s patterns are outside the norm.
How to keep occupancy data private and still provide the information utilities need to manage their grids is the next area of research, Jin says. “Right now, meters are sending accurate information about energy consumption. To protect privacy, you could add some noise to that data. We are now looking to determine the optimal size of the added noise that would mask information about occupancy and still give the utility company an accurate enough reading for its needs.”