DeepDFML-NILM: um modelo baseado em aprendizado profundo para detecção, extração de características e classificação de sinais de NILM

Abstract

In the coming decades, the growth in energy consumption will require renewable resources and intelligent solutions to manage consumption. In this sense, Non-Intrusive Load Monitoring (NILM) techniques emerge as a way to provide detailed consumption information to users, enabling better power management and avoiding energy losses. These load monitoring techniques typically require four steps: event detection, signal disaggregation, feature extraction, and load identification. However, for high-frequency NILM methods, state-of-the-art approaches, mainly based on deep learning solutions, do not provide a complete NILM architecture that includes all the required steps. To fill this gap, this work presents an integrated method for the detection, feature extraction, and classification of high-frequency NILM signals for the publicly available LIT-Dataset. For detection, the results were above 90% in most cases, while the state-of-the-art methods were below 70% for eight simultaneous loads. For classification, the performance of the proposed model on the evaluated metrics was comparable to other recent works, reaching over 97% for F-score and accuracy. The proposed architecture also includes a multi-label classification strategy to avoid the disaggregation stage, indicating the loads connected at a given time and allowing the identification of multiple loads simultaneously. This work also evaluates the robustness of the proposed method to noise insertion. Finally, this work presents results in an embedded system, a topic also underexplored in the recent literature, demonstrating the feasibility of the proposal for real-time signal analysis and practical applications involving NILM.