A Comparative Study of Machine Learning Classifiers for Electric Load Disaggregation based on an extended NILM dataset

The appliance evaluation and the power consumption consciousness are becoming essential for improving demand management and power grid enhancement. Load disaggregation becomes a promising engine for this goal, and some researches efforts have been made in the last years. In this sense, achieving the load characterization is essential to the technique’s success; moreover, the proper feature extraction becomes essential. In this way, this paper presents a comparative study of machine learning classifiers for electric load disaggregation using an enhanced version of a household appliance dataset proposed by Souza et al. of Brazilian appliances (NILMbr). The load characterization is performed through the Conservative Power Theory, a recent power theory that extracts appliance signatures by means of power quantities. Then, it is proposed three machine learning models to validate proper load identification, being: classification algorithms - kNearest Neighbor (k-NN), Support Vector Machine (SVM), and Random Forest (RF). These algorithms were used to assess computational time and performance metrics. Subsequently, the RF algorithm presented the best performance, with an accuracy of 99.5%.