Non-intrusive Load Monitoring Using High-Frequency Measurements and Machine Learning

Non-intrusive load monitoring can be considered as a knapsack or blind source separation problem, complicated by numerous possible combinations of devices that reconstruct total power. Simultaneous appliance usage and the presence of noise further complicate the task of accurately disaggregating the total power consumption into individual appliance contributions. In this paper we aim to solve the inverse problem of disaggregating the active power of each individual appliance from the overall power consumption and current harmonics. Additionally, we compare the performance of random forest, convolutional neural networks, support vector machine and linear regression models on the same set of features.