Comparison of Clustering Techniques for Residential Energy Behavior Using Smart Meter Data

Jin, Ling (Lawrence Berkeley National Laboratory) | Lee, Doris (Lawrence Berkeley National Laboratory) | Sim, Alex (Lawrence Berkeley National Laboratory) | Borgeson, Sam (Lawrence Berkeley National Laboratory) | Wu, Kesheng (Lawrence Berkeley National Laboratory) | Spurlock, C. Anna (Lawrence Berkeley National Laboratory) | Todd, Annika (Lawrence Berkeley National Laboratory)

AAAI Conferences 

Current practice in whole time series clustering of residential meter data focuses on aggregated or subsampled load data at the customer level, which ignores day-to-day differences within customers. This information is critical to determine each customer’s suitability to various demand side management strategies that support intelligent power grids and smart energy management. Clustering daily load shapes provides fine-grained information on customer attributes and sources of variation for subsequent models and customer segmentation. In this paper, we apply 11 clustering methods to daily residential meter data. We evaluate their parameter settings and suitability based on 6 generic performance metrics and post-checking of resulting clusters. Finally, we recommend suitable techniques and parameters based on the goal of discovering diverse daily load patterns among residential customers. To the authors’ knowledge, this paper is the first robust comparative review of clustering techniques applied to daily residential load shape time series in the power systems’ literature.