Non-intrusive load monitoring applied to a home energy management system

• Main Author: Costa, Tomás Oliveira da
• Publication Date: 2022
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10773/39399
• topic: Non-Intrusive Load Monitoring; Energy disaggregation; Home energy management systems; Machine learning; Recurrent neural networks; Convolutional neural networks; Ensemble learning

In recent years, there’s been a huge advancement in the field of the Internet of Things (IoT), this is due to the information that these systems are able to report and the automation that they can provide, but this comes with a problem. The abundance of these equipment’s has risen energy consumption by manyfold since there are now appliances in every corner of a household. One way to reduce this high consumption is by increasing the energetic literacy of homeowners. This need for teaching energetic demands allied with the need to reduce global energy consumption leads way for solutions such as Nonintrusive Load Monitoring (NILM) to appear. This solution collects the data of a given household in a nonintrusive fashion and then performs what is called energy disaggregation, which is the individualization of an appliance’s electrical readings from an aggregate reading. Inserted in a project, the proposition of this dissertation is to develop models that provide a good NILM solution and convert this solution to a functioning Home Energy Management System (HEMS). The hypothesis we wanted to test with this work was to verify that the average results for an ensemble would yield better results than the individual models contained in the ensemble. After validating this argument, we proposed two ensemble models, which we named Appliance Specific Ensembles (ASE) and then evaluate them with the other top-performing NILM models. Here we discover that our models outperform the best individual model of the state-of-the-art in two out of five appliances. We also proved that our ensembles are more resilient and fault-tolerant than the individual models contained in them, as some models would fail to correctly predict the usage and only predict a constant value. Using these models for predictions on different homes also yielded good and consistent results for our proposed ASE models, which is possible to visualize on our developed HEMS platform.