Sensor-aided NILM using Deep Learning models

Sensor-aided NILM using Deep Learning models

Mots-clefs : NILM, Environmental Sensors, LSTM, WindowGRU, GMM

Abstract: This thesis explores novel methodologies to enhance the performance and applicability of Non-Intrusive Load Monitoring (NILM) systems, addressing key challenges in energy disaggregation accuracy and adaptability. A context-aware framework was developed by integrating environmental sensor data such as temperature, humidity, and occupancy into NILM processes, significantly improving the identification of appliances with overlapping or environmentally influenced power usage patterns. A central contribution is the creation of a high-resolution dataset, combining energy consumption and environmental metrics, collected from the SmartSense platform. This dataset was integrated into the NILMTK toolkit, providing an accessible and valuable resource for the NILM research community.

On the other hand, the thesis introduces the Spiral Layer, a novel neural network architecture inspired by rotation operations, designed for efficient feature extraction and dimensionality reduction. This architecture was extended to implement a trainable Karhunen-Loève Transform (KLT), offering a neural-network-based approach to orthogonal projections and signal variance maximization. The Spiral Layer demonstrated competitive performance on benchmark tasks, validating its potential for broad applicability.