Non-intrusive industrial load monitoring on a factory in Brazil

Abstract

This thesis addresses the comparison of two techniques of Non-Intrusive Load Monitoring applied to electrical data collected from a factory in Brazil. NILM proposes to separate single-appliance power consumption from consumers total demand without the need for installation of intrusive sensors or more than one meter per building. As the main focus of this thesis is to study NILM on industrial settings and, until the date of writing, no public data were found, IMDELD data set was collected for this research on a poultry feed factory using smart meters. IMDELD has a total of eleven classes of electrical signatures, including eight classes of heavyindustry machines, two different sub-circuits, and a main circuit. The data was collected at a 1 Hz rate for up to a hundred eleven days. To achieve the comparison goal, two methods are implemented: Factorial Hidden Markov Models and Deep Learning (WaveNILM). In comparison to the FHMM models, the Deep Learning-based models have smaller Signal Aggregated Error and Normalized Disaggregation Error. They also identified single-appliances as turned ON or OFF on a larger percentage of the time tested based on F1-Score. Among all appliances, on average WaveNILM F1-scored 0.93±0.07 while FHMM scored 0.79 ± 0.12. WaveNILM predicted machines with average SAE 0.1 ± 0.2 and NDE 0.1 ± 0.2, while FHMM predicted machines with average SAE 0.3 ± 0.2 and NDE 0.3 ± 0.2.