Data mining of smart meters for load category based disaggregation of residential power consumption

Non-Intrusive Load Monitoring (NILM) techniques have been leveraged by new instrument and machine learning algorithms to provide customers the breakdown of their energy usage. The state-of-art indicates a large amount of high-frequency measurement (>1 Hz) can lead to accurate disaggregation. This paper, however, proposes a disaggregation algorithm relies on hourly smart meter readings, aiming to extend the application of the low-frequency data that is accessible by both utilities and customers. The output of the disaggregation includes the breakdown of energy into load-category-based components that have different average power factors. The disaggregated data will support small-scale planning, e.g., in microgrid, by revealing the variance and patterns in different load categories. Our approach is built on a top-down structure that requires no prior knowledge or general models of individual loads. Using clustering and optimization techniques, we infer the load signatures of each category based on the active and reactive power from smart meters. The signatures are updated periodically using the most recent smart meter data. The results show that our disaggregation approach could be applied to random houses in different seasons and to single house and small neighborhood in both offline and quasi-real-time context.