Thesis

Non-Intrusive load monitoring and anomaly detection : on the importance of feature selection for supervised and unsupervised learning for sensor applications

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T17621
Person Identifier (Local)
  • 201884058
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Non-Intrusive Load Monitoring (NILM) provides a scalable and cost-effective means of disaggregating household energy consumption from aggregate smart meter data. In addition to disaggregation, NILM holds promise for detecting abnormal appliance behaviour, supporting predictive maintenance, enhancing energy efficiency, and improving household safety. However, several challenges hinder its practical deployment, including high computational cost, lack of transferability across households, sensitivity to low frequency smart meter data, and limited frameworks for anomaly detection. This thesis addresses these challenges through three main contributions. First, it presents a systematic investigation of feature selection for NILM. By analysing both spectrometry and smart meter datasets, it demonstrates that selecting the most informative features reduces computational complexity while improving classification accuracy. Comparative analyses of supervised and unsupervised algorithms—including Artificial Neural Networks (ANN), Decision Trees (DT), K-Means, and DBSCAN—highlight the role of feature selection in balancing performance and efficiency. Second, the thesis evaluates load disaggregation using supervised (DT, KNN) and unsupervised (DBSCAN, UGSP) techniques. Detailed pre-processing (e.g., noise reduction, resampling, segmentation) and post-processing (e.g., power reconciliation, error correction) methods are integrated to improve robustness. A two-stage disaggregation strategy is proposed to enhance detection accuracy, and transfer learning experiments are conducted to assess generalisability across households, offering insights into scalability and dataset adaptability. Third, the thesis develops a novel framework for NILM-based anomaly detection. A hybrid approach, combining Unsupervised Graph Signal Processing (UGSP) with a statistical rule based method, is applied to fridge-freezers and washing machines in the REFIT dataset. By modelling ON-duration thresholds and operational cycles, the framework successfully identifies abnormal appliance behaviour without the need for intrusive submetering. The results show improved precision and recall in detecting anomalies, demonstrating the feasibility of NILM-driven predictive fault detection. Overall, this work advances NILM from a disaggregation-focused task toward an anomaly aware monitoring framework. It contributes: (i) a systematic evaluation of feature selection strategies, (ii) comparative benchmarking of supervised and unsupervised disaggregation methods, (iii) an assessment of cross-dataset transferability, and (iv) the development of a hybrid NILM-based anomaly detection approach. Together, these contributions provide a comprehensive framework that reduces computational overhead, improves robustness, and enables early detection of appliance faults. The findings support the integration of NILM into smart grid infrastructures, offering benefits for energy efficiency, user engagement, and residential safety.
Advisor / supervisor
  • Stanković, Vladimir
  • Stankovic, Lina
Resource Type
DOI
Date Created
  • 2025

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T17621 2026-02-24 Public Download