Semi-Intrusive Load Monitoring via Smart Plug Data Fusion for Robust NILM
Abstract
Non-Intrusive Load Monitoring (NILM) aims to infer the consumption of individual electrical devices in a building from mains meter data. Despite substantial progress, state-of-the-art approaches still face several key barriers. Event-based models do not scale well to the number of devices in a typical household and typically rely on data resolutions that are unrealistically high for today’s smart meters. In contrast, eventless architectures often fail to generalize across the diversity of device signatures. To address the latter, we introduce a new sampling strategy for eventless NILM models, increasing the context size via Gaussian sampling whilst keeping the model dimensionality, achieving a reduction of the mean absolute error (MAE) of up to 57%. We further propose a simple multimodal data fusion approach that augments NILM inputs with smart plug context to improve contextual representation and robustness. Our proof-of-concept experiments on realistic public household data show that this semi-intrusive approach can help overcome these barriers. Semi-intrusive event-based models improve the baseline performance of a 27.3% macro F1 score to 57.9% for a complete household using low-frequency data. In the eventless scenario, we adapt a non-intrusive Transformer to operate semi-intrusively and achieve 65% reduction of the MAE and 85% reduction of the error-standard-deviation across devices, thereby improving generalization across device types.
Justus Breyer, M.Sc.
Researcher
Leonardo Pompe, M.Sc.
Researcher
Kai Gützlaff
Nils Bollwerk
Prof. Dr. Muhammad Hamad Alizai
LUMS, Lahore, Pakistan
