Targeted preprossessing for weight reduction in NILM datasets

Abstract

This paper explores variables for NILM dataset creation, focusing on the relationship between measurement frequency, dataset weight, micro and macro characteristics, and prepossessing. Measurements show that a frequency of 2 kHz allows micro analysis by FFT decomposition and high resolution macro analysis by 0.02 s resampling. Storing only features of interest such as events or harmonic frequencies show potential for diminishing the weight of the dataset while keeping insightful data.