Home activity recognition using infrequently-monitored HEMS Data

This paper proposes a method for estimating household activities based only on the cumulative power consumption data obtained from the HEMS home distribution board every 30 min. The proposed method predicts the activity of each 30 min timeslot from the eight activity labels; household-level waking-up, household-level going-to-bed, room-level waking-up, room-level going-to-bed, cooking, laundry, dishwashing, and bathing. For the prediction, we first identify the branch circuit that is strongly correlated with each activity label and detect the turn-on/off of home appliances on the circuit to detect those activities. We also incorporate machine learning for estimating the other activities based on the circuit’s time series of power consumption. Furthermore, to cope with the difference among households, we apply transfer learning to the constructed model. In collaboration with a Japanese home builder, we conducted an experiment on five households using their HEMS data. In parallel, we obtained verifiable activity labels as our ground truth by the installation of specialized sensors in the respective homes. Under a ±30 min tolerance (i.e. allowing a prediction in the immediately preceding or following half-hour slot), our model achieved an average F1 score of 0.689 across all activities. We also confirmed that transfer learning improved the F1 score of each activity recognition and achieved an average improvement of 0.260 in household-level waking-up, household-level going-to-bed, room-level waking-up, room-level going-to-bed, and bathing activities. • We propose a method for recognizing home activities from low-granularity power consumption data. • We design data-driven input features guided by qualitative analysis of power consumption patterns across households. • The method adapts differences across households using transfer learning. • The proposed method is evaluated based on activity labels obtained from real households.