UNITY: Semi-supervised Meta-learning Load Monitoring for Resource-restricted Smart Grids

Smart grids rely on massive data from networked smart meters to enable fine-grained energy analytics and improve sustainability. As a representative, Non-Intrusive Load Monitoring (NILM) can infer individual appliance usage from aggregated meter readings, offering energy-saving insights without the need for per-appliance sensors. However, practical deployment at scale is hindered by three challenges: (1) limited per-device resources such as on-board computation and mobile data plans; (2) scarce labeled data due to the high cost of manual annotation; and (3) significant variability across households, including temporal changes in user behavior and appliance status.To address these challenges, we propose UNITY, a semi-supervised meta-learning NILM framework that unifies labeled and unlabeled data across diverse households to learn generalizable representations. UNITY minimizes user-side computation and communication overhead by reformulating inference as a lightweight sequence-retrieval task, accelerated by Discrete Haar Wavelet transforms. Only uncertain samples are uploaded to the server for refinement. To overcome label scarcity, UNITY employs entropy-based transductive learning, gradually enhancing model confidence on unlabeled data. Furthermore, to handle household diversity and temporal distribution shift, UNITY adopts a meta-learning approach that treats each household as a distinct task and incrementally adapts to behavioral drift over time, enabling robust long-term deployment. Experiments on public NILM datasets demonstrate that UNITY achieves 95.21% accuracy, outperforming existing methods under real- world resource constraints.