Evaluation of Clustering and Time Series Features for Point Type Inference in Smart Building Retrofit

Metadata inference for building automation system (BAS) is an increasingly important topic to promote wider adoption of smart building technologies. Metadata inference is used to automatically discover semantics within the BAS, such as labelling sensors, discover control variable relationships, etc. Clustering analysis has been applied in many previous research studies to achieve faster smart building retrofits through automated or semi-automated BAS point labelling. However, previous research using clustering only used small data sets of two to five buildings. This research examines the effectiveness of this approach on a broader scale with 40 commercial and institutional buildings and more than 65,000 labelled BAS points. Different clustering strategies with varying feature space and clustering algorithms are examined. Furthermore, this study compares which time series features and generation approach may enhance labelling efficiency. Positive results from this study support the effectiveness of applying clustering for point type inference. Results show the complimentary nature of additional time series features when the existing raw metadata from the BAS is less descriptive.