The behavior of occupants has a significant impact on building energy consumption. To extract their behavior patterns, data mining methods are widely utilized, with clustering algorithms such as K-means based on “Euclidean” distance being the representatives. However, these algorithms have limitations in characterizing occupants’ behavior over time. Although some studies have started to use time-series clustering algorithms, the resulting data characteristics differ since occupants’ behavior is characterized by both persistent (energy demand) and non-persistent (heat demand, cold demand, and lighting demand) features.
Previous studies focused more on analyzing household energy data using time-series algorithms than non-persistent behavior. Thus, this study compared three time-series clustering algorithms (K-shape, Dynamic Time Warping (DTW), and Derivative Dynamic Time Warping (DDTW)) in terms of occupants’ heat demand behavior and explored possible reasons for differences while proposing three evaluation methods.
The results indicate that: (1) all three algorithms achieved clustering effects for non-continuous behavior (heat demand), with DTW capturing heating demand pattern changes more accurately; (2) due to the discontinuity of heat demands, these three types of time-series algorithms did not outperform the K-means algorithm significantly in clustering effects; and (3) the proposed evaluation methods performed relatively well compared to commonly used metrics for evaluating clustering effects.
This study provides a reference for selecting appropriate clustering algorithms to study non-persistent residential behavior and has potential value in improving energy prediction and management.
