Leveraging Smartphones and Machine Learning to Optimize Building Energy Use
A team of researchers led by College of Information and Computer Sciences (CICS) doctoral student Amee Trivedi received the 2020 Best Paper Award at the International Green and Sustainable Computing Science Conference for the paper, “Phone-based Ambient Temperature Sensing Using Opportunistic Crowdsensing and Machine Learning.”
The researchers have developed a novel technique that leverages sensors on a smartphone to sense indoor ambient temperature, essentially turning any smartphone into a digital thermometer. The system, developed by Trivedi with doctoral student Phuthipong Bovornkeeratiroj and Professor Prashant Shenoy of the Laboratory for Advanced Software Systems (LASS) at CICS, along with LASS alumnus Joseph Breda and UMass College of Engineering faculty Jay Taneja and David Irwin, leverages machine learning techniques to sense subtle changes in the battery temperature of a phone in order to infer ambient air temperature at that location.
“We wanted to focus on increasing the energy efficiency in buildings, which account for nearly 70 percent of total electricity usage in the US,” says Shenoy. “Poorly-configured HVAC systems, especially in older buildings, present us with a commonly-found challenge and an opportunity to widely optimize energy use.”
The researchers have created a crowd-sourcing system to gather data from multiple users’ smartphones as they move through indoor spaces and uses a “random forest” ensemble learning model to combines temperature predictions from all of the phones in a certain room or floor of a building in order to to create an accurate temperature heatmap of a building’s interior.
In their prototype, the researchers combined a WiFi-based system which determined the location of phones in the building with a web service that read data from phones in an opportunistic fashion using their new learning model.
“To predict ambient temperature, we need to be able to train phones to understand both their placement and position and the strenuousness of their internal activities,” explains Trivedi. “For example, is a participating phone exposed or in a pocket—and is it in an idle state or running a CPU-intensive app that generates more heat?” To quickly deploy this machine learning model on participating phones, the team used a “few shot” training approach using a limited amount of sample data.
They anticipate that their system can combine personal thermal preferences of users and fine-grain crowd-sourced temperature measurements to optimize HVAC energy use while also enhancing personal comfort. The system can be easily deployed in older buildings to be used by HVAC systems for improved temperature data and control.