Politecnico di Torino (logo)

An Unsupervised and Online NILM Approach for ON/OFF Appliances Detection

Angelica Urbanelli

An Unsupervised and Online NILM Approach for ON/OFF Appliances Detection.

Rel. Edoardo Patti, Marco Castangia. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2021


In the last decade, the development of Smart Grids technologies has been enforced by the growing energy demand in the residential sector, as well as the increasing concern about climate changes and the reduction of their consequences. Indeed, one of the six priorities of Von der Leyen commission is to convert the actual energy system into a more integrated, decentralised, and sustainable one. In this context, one of the most promising technologies is NILM (Non-Intrusive Load Monitoring), whose objective is, starting from the aggregate power signal collected in a building with a single smart meter, to disaggregate and distinguish the various household appliances. Midori, a start-up born in Turin in 2011, operates in the smart metering domain, and it has developed its own smart meter and NILM system. Midori’s actual implementation is able to recognise most of the multi-state appliances, that is, those devices that have a repetitive power consumption pattern and that constitute a high part of houses’ total energy consumption. However, there is still a huge category of devices that are actually not addressed: the ON/OFF ones, whose electric fingerprint is too poor to be correctly distinguished by the actual system, thus that kind of power events has been so far labeled as unknown. This work aims to present a house-specific approach able to analyse those unknown events that, due to the lack of annotated data, is completely unsupervised. The idea is to leverage on additional features like event’s duration, time of usage, day of the week and whether the day is a weekday or a weekend day. The data are firstly cleaned in order to get only the events belonging to the target devices. Then they are processed through an online clustering algorithm that tries to find usage patterns among them, exploiting a specific metric. Finally, every time a new pattern is found, a machine learning model is trained to be able to recognise, from that moment on, events belonging to the same pattern, thus incrementally improving the whole system’s ability to distinguish ON/OFF devices.

Relators: Edoardo Patti, Marco Castangia
Academic year: 2021/22
Publication type: Electronic
Number of Pages: 107
Additional Information: Tesi secretata. Fulltext non presente
Corso di laurea: Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering)
Classe di laurea: New organization > Master science > LM-32 - COMPUTER SYSTEMS ENGINEERING
Aziende collaboratrici: Midori Srl
URI: http://webthesis.biblio.polito.it/id/eprint/21295
Modify record (reserved for operators) Modify record (reserved for operators)