Aumento de dados para monitoramento de carga não intrusiva: uma revisão e estudo empírico.

Abstract

As global energy demand rises, Non-Intrusive Load Monitoring offers a path to enhanced household energy efficiency. However, the deep learning models central to NILM require extensive labeled data, which is costly and time-consuming to collect. This dissertation confronts this data scarcity challenge by investigating the impact of data augmentation tech niques in the NILM domain. The study combines an extensive literature review, categorizing current augmentation strategies, with a rigorous empirical evaluation of selected techniques using the WaveNILM model on real-world household data, focusing on Intra-House and Inter-House generalization. Results unequivocally show that data augmentation significantly improves NILM model performance and generalization. Notably, a merged augmentation strategy consistently outperformed individual methods and a baseline without augmenta tion, delivering substantial gains in disaggregation accuracy and model stability, especially in challenging Inter-House scenarios. This research offers crucial insights into augmenta tion mechanisms, including enhanced noise handling and appliance signature identification, and demonstrates that augmentation facilitates more efficient learning and effective feature representations. The dissertation provides valuable empirical evidence and a structured un derstanding of the augmentation landscape in NILM, offering a practical pathway to mitigate data scarcity, advancing the development of more reliable and deployable NILM solutions for energy conservation.