Contribution of Non-Intrusive Load Monitoring to Home Energy Management Systems
- Authors
- Publication Date
- Dec 14, 2023
- Source
- HAL-Descartes
- Keywords
- Language
- English
- License
- Unknown
- External links
Abstract
The current energy crisis has revealed deep vulnerabilities in energy supplies. It also emphasised the strong fossil fuel dependency of many countries. Some governments decided to tighten their energy policies, particularly on energy efficiency and the use of intermittent and renewable energies. Photovoltaic energy has the advantage of emitting few greenhouse gases in its lifetime operation but has the drawback of being intermittent and variable. Energy storage can mitigate the variability and intermittency of solar energy. However, energy storage's substantial environmental and financial costs may refrain from investing massively. A second way to face renewable drawbacks involves more intelligent demand management. Developing a Home Energy Management System (HEMS) for residential areas has been a growing research trend but has not been widely applied due to technical, social and financial barriers. It is expected to monitor appliances individually in residential areas to understand the user's behaviours accurately to determine the preferred hour of use of each appliance, the duration, or the energy demanded. However, Appliance Load Monitoring (ALM) is challenging due to the high complexity and cost of the sensor network involved. Non-Intrusive Load Monitoring (NILM) is the process of disaggregating the main load into individual appliance loads. NILM aims to mitigate the cost of sensor installation and maintenance by installing a unique sensor on the main load and retrieving the individual load appliances computationally. Recently, deep learning models have been up-and-coming for NILM tasks. The main meter sampling rate is crucial for the trade-off between disaggregation accuracy and cost efficiency. High-frequency data will carry more detailed signatures but with a costly sensor. In this work, we assess the generalisation capabilities of low-frequency NILM and identify the settings having the most significant impact on performances. This work demonstrates the importance of diversity in the training dataset to unlock generalisability. At the same time, the literature underlines data scarcity in the domain, mainly due to the complex task of recording reliable supervised data in real conditions. The present work provides a new data augmentation technique to address this issue by enriching the training set. The technique is thoroughly experimented with to prove its efficiency in improving generalisability. The data augmentation add-on module is implemented inside the NILMTK framework, a famous toolkit to evaluate NILM solutions. This manuscript brought particular care to a practical NILM application. The idea is to assess the contribution of a low-frequency NILM to a HEMS by conducting simulations. The case study applies a load scheduling HEMS to real houses from public databases. The schedules given by the HEMS with NILM are more acceptable for the end-user. The first original contribution of this work original relies on developing a new and straightforward data augmentation technique published in the SEGAN (Sustainable Energy, Grids and Network) journal. Secondly, the thesis gives an original approach to quantitatively evaluating NILM's contribution to HEMS.
