Energy disaggregation: Nonintrusive load monitoring

Energy disaggregation: Nonintrusive load monitoring: The search for practical methods

Janssen, Walter (TU Delft Mechanical, Maritime and Materials Engineering; TU Delft Delft Center for Systems and Control)

De Schutter, Bart (mentor)

Delft University of Technology

2018-01-29

In this MSc. thesis steady-state disaggregation methods for the Non-Intrusive Load Monitoring are researched. The main target is to find methods that have the potential to be practically applicable. These methods are characterized by the fact that they use data that is collected by smart meters, since smart meters are more and more the standard in Western households. Appliances are modeled by Hidden Markov Models and households are modeled by super-state Hidden Markov Models or Factorial Hidden Markov Models. These models are
trained by iterative K-means or expectation maximization, where iterative K-means turns out as the superior training method. Basically three disaggregation algorithms are researched: the Viterbi algorithm, the particle filter, and the newly proposed Global Transition Minimization (GTM).
The disaggregation algorithms are tested on a synthetic dataset of 6 appliances and on part of the Dutch Residential Energy Dataset (DRED). The results on the synthetic dataset vary between an accuracy of 81% for the GTM and 95% for the Viterbi algorithm. On the DRED the highest accuracy that is achieved is 77% for the particle filter. It turned out that the accuracy of the particle filter is not always improved by increasing the number of particles when real-world data is considered.
Improvements for the disaggregation algorithms by using time information data and by using reactive power as extra input data are proposed. However, both suggestions do not lead to more accurate disaggregation results. A new modeling framework, where state transitions are also included in the model, is suggested as a future research topic.

Viterbi
GTM
particle filter
hidden Markov model
NILM
disaggregation

http://resolver.tudelft.nl/uuid:60db0db5-e051-4ce9-b378-38b945944401

Student theses

master thesis

© 2018 Walter Janssen