Energy and Buildings, Vol.187, 95-109, 2019
Self-configuring event detection in electricity monitoring for human-building interaction
Monitoring the temporal changes in the operational states of appliances is a key step in inferring the dynamics of operations in smart homes. This information could be leveraged in a variety of energy management applications including energy breakdown of individual loads, inferring the occupancy patterns, and associating the energy use to occupants' activities. The operational states of appliances could be identified through detecting and classifying the events on power time-series. Despite the advancements in the field of event detection, they often require in-situ configuration of model parameters to achieve a higher level of performance according to each new context. In order to address such limitation, in this paper, we have proposed a self-configuring event detection framework for detecting the changes in operational states of appliances. The framework seeks to autonomously learn the contextual characteristics of the loads from the environment and adapt the event detection parameters. The proposed unsupervised framework couples an automated clustering for identifying the recurring motifs, which are representations of the appliances' transient power draw signatures in a given environment and a proximity-based motif matching for detecting the events. The framework was evaluated on EMBED dataset, a publicly available fully labeled electricity disaggregation dataset, collected from three apartments with different categories of the appliances. The evaluations demonstrate that the proposed event detection framework outperforms the conventional event detection in detecting the operational states of different classes of loads across different environments. The proposed framework could also facilitate human-building interactions in training smart home applications by populating motifs to infer the operations of appliances and activities of occupants. (C) 2019 Elsevier B.V. All rights reserved.
Keywords:Electricity disaggregation;Event detection;Unsupervised learning;Non-intrusive load monitoring;Clustering