Non-intrusive load disaggregation in distribution systems using machine learning

O'Malley, Elliott 2025. Non-intrusive load disaggregation in distribution systems using machine learning. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Download (30MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (132kB)

Abstract

This work is motivated by the need to support decarbonisation and manage the growing impact of low-carbon technologies (LCTs), which place increasing strain on electrical distribution networks. Non-Intrusive Load Monitoring (NILM) employs machine learning to disaggregate household or feeder-level power signals into appliance-level consumption, offering a solution to monitor LCTs such as electric vehicles (EVs) and solar photovoltaics (PV). Existing NILM methods often demand extensive manual tuning, high computational resources, and domain-specific expertise, which limit scalability and model development. This thesis investigates the use of Automated Machine Learning (AutoML) for NILM, comparing three leading frameworks: AutoGluon, H2O, and FLAML, on the REFIT and Pecan Street datasets. A Seq2Point (S2P) sequencing approach, is employed using a 599-sample window. Feature engineering configurations are systematically explored, ranging from baseline temporal features, through intermediate statistical features, to full raw aggregate sequences. Simulations are conducted at the household level and extended to a quasi substation scale through aggregation of multiple households. The results demonstrate that AutoML can deliver competitive NILM performance, with AutoGluon and H2O consistently outperforming FLAML across evaluations. Careful tuning and selection of AutoML features can offer improved performance compared to state of the art literature, with up to a 62.24% improvement in Mean Absolute Error (at household-level). Substation-NILM was found to be less accurate than Household-NILM, with only a small deviation observed at the Micro-Substation level (single appliance signal) compared to the larger errors at the Macro-Substation scale (grouped appliance signals), reflecting the greater complexity and overlapping signals. Performance improvements were achieved through the introduction of an additional feature capturing the appliance states. The numerical framework was extended to develop a method that utilised disaggregation outputs to predict the number of active EV chargers at each timestep, enabling forecasts of expected changes in quasi-substation power trends. These findings highlight the viability of AutoML as a scalable and flexible approach to NILM, with implications for monitoring household demand and low carbon technology uptake at household and substation levels.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1. Non-Intrusive Load Monitoring (NILM) 2. Automated Machine Learning (AutoML) 3. Artificial Intelligence (AI) 4. Low Carbon Technologies 5. Energy Disaggregation
Funders:	EPSRC
Projects:	EP/S022996/1
Date of First Compliant Deposit:	5 May 2026
Last Modified:	05 May 2026 10:59
URI:	https://orca.cardiff.ac.uk/id/eprint/186669

Actions (repository staff only)

Edit Item