Residential Load Disaggregation Using Deep Learning Methods
(2025)Department of Automatic Control
- Abstract
- The electrical energy landscape is changing rapidly with increasing electrification and renewable energy expansion. Smart meters deployed in households create opportunities for understanding energy consumption patterns through load disaggregation, the task of identifying individual appliances directly from aggregate electrical power data. Load disaggregation could be used for better understanding of past energy consumption, both for the consumers and for electricity providers, and providing estimates of current or future power consumption for home energy management systems. This thesis develops and evaluates three neural network approaches for short-term historical load disaggregation on residential data, focusing on energy-intensive... (More)
- The electrical energy landscape is changing rapidly with increasing electrification and renewable energy expansion. Smart meters deployed in households create opportunities for understanding energy consumption patterns through load disaggregation, the task of identifying individual appliances directly from aggregate electrical power data. Load disaggregation could be used for better understanding of past energy consumption, both for the consumers and for electricity providers, and providing estimates of current or future power consumption for home energy management systems. This thesis develops and evaluates three neural network approaches for short-term historical load disaggregation on residential data, focusing on energy-intensive appliances. Three architectural types were investigated - recurrent neural networks, state-space models, and attention mechanisms, each designed to handle sequential data. Evaluations were done on three appliance types, electric vehicle chargers, air conditioners, and clothes dryers. Results show that recurrent and state-space models, specifically bidirectional Gated Recurrent Units and Mamba, outperformed attention-based approaches in identifying appliances. External conditioning with temperature and time-of-day data improved air conditioner prediction accuracy. Fine-tuning models on unseen households could be achieved with comparable performance to full fine-tuning approaches using a few-shot segmentation approach. The few-shot approach simulates asking customers when their appliances were active, reducing data requirements while maintaining accuracy. Although results are promising, more data would be needed for comprehensive training and evaluation of these methods. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9212313
- author
- Fredriksson, Viktor
- supervisor
- organization
- year
- 2025
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6294
- other publication id
- 0280-5316
- language
- English
- id
- 9212313
- date added to LUP
- 2025-09-18 14:16:23
- date last changed
- 2025-09-18 14:16:23
@misc{9212313,
abstract = {{The electrical energy landscape is changing rapidly with increasing electrification and renewable energy expansion. Smart meters deployed in households create opportunities for understanding energy consumption patterns through load disaggregation, the task of identifying individual appliances directly from aggregate electrical power data. Load disaggregation could be used for better understanding of past energy consumption, both for the consumers and for electricity providers, and providing estimates of current or future power consumption for home energy management systems. This thesis develops and evaluates three neural network approaches for short-term historical load disaggregation on residential data, focusing on energy-intensive appliances. Three architectural types were investigated - recurrent neural networks, state-space models, and attention mechanisms, each designed to handle sequential data. Evaluations were done on three appliance types, electric vehicle chargers, air conditioners, and clothes dryers. Results show that recurrent and state-space models, specifically bidirectional Gated Recurrent Units and Mamba, outperformed attention-based approaches in identifying appliances. External conditioning with temperature and time-of-day data improved air conditioner prediction accuracy. Fine-tuning models on unseen households could be achieved with comparable performance to full fine-tuning approaches using a few-shot segmentation approach. The few-shot approach simulates asking customers when their appliances were active, reducing data requirements while maintaining accuracy. Although results are promising, more data would be needed for comprehensive training and evaluation of these methods.}},
author = {{Fredriksson, Viktor}},
language = {{eng}},
note = {{Student Paper}},
title = {{Residential Load Disaggregation Using Deep Learning Methods}},
year = {{2025}},
}