Appliance classification across multiple high frequency energy datasets

Abstract

Non-intrusive load monitoring (NILM) provides several techniques for demand information retrieval to support consumers saving energy usage. Research in NILM often focuses on closed environments, such as single datasets or single households. Disaggregation results are typically not suitable to represent the classification performance under real circumstances due to its data homogeneity of a single dataset. We apply a classification system across four commonly available high frequency energy datasets. The experiments include classification tasks with four different classifiers on 36 spectral and temporal features to perform a cross-, mixed-, and intra-dataset validation. The outcome of this work is a reliable benchmark for appliance recognition in the high frequency domain and its efficiency in smart meters for different use cases and appliance features.