Recognizing Activities of Daily Living for Activity-Centric Energy Efficiency in Residential Buildings: A Human Sensing and Deep Learning Approach

Recognizing and accounting for activities of daily living (ADL) in building operations holds great promise to improve energy efficiency of residential buildings and enhance occupant comfort. However, recognizing ADL is challenging because most residential buildings are not instrumented with physical sensing devices typically required for daily activity recognition. To address this challenge, this paper proposes a new ADL recognition method, aiming to generalize occupant activity recognition to residential buildings without the sensing capability. The proposed method leverages: 1) existing human sensing effort, the American Time Use Survey, to acquire data about ADL as well as the demographic and socioeconomic characteristics of a representative sample of occupants; and 2) deep sequential learning to learn time-sensitive relationships between occupant activities and occupant characteristics as well as the long-range dependency relationships between activities, to automatically recognize the sequence of ADL that an occupant could conduct during a typical day only based on easily acquirable occupant characteristics data as inputs. The proposed method was tested in both technical and application-oriented evaluations. The technical evaluation showed that the proposed method achieved a precision of 93.1% and a recall of 93.0% in ADL recognition. The application-oriented evaluation showed that, compared to the non-activity-based schedules defined by the ASHRAE Standard 55-2020, operating heating and cooling systems according to the schedules that are defined based on the recognized ADL: 1) saved heating energy by 1.2% and cooling energy by 17.0% and 2) improved thermal comfort in terms of predicted mean vote (PMV) value by 83.3% during the summer and 88.5% during the winter.