A Weakly Supervised Learning Framework Integrating Uncertainty-Aware and Class-Specific Attention for Non-Intrusive Load Monitoring

Abstract

Non-intrusive load monitoring (NILM) is essential for fine-grained energy management, but its practical deployment is hindered by the high cost of acquiring fine-grained strong labels and cross-domain distribution shifts. To address this limitation, we propose UACA-Net, a hybrid weakly supervised framework for load disaggregation. UACA-Net integrates a class-specific attention pooling mechanism to improve temporal localization of transient appliance activations under weak supervision. In addition, a homoscedastic uncertainty-driven dynamic weighting strategy is introduced to adaptively balance multi-task objectives. A two-stage warm-start training scheme is further designed to stabilize joint optimization and mitigate early-stage training instability. Extensive experiments on the UK-DALE and REFIT datasets show that UACA-Net achieves a Macro-F1 of 0.9143 and a Macro-TECA of 0.9198 using only 20% strong labels. Moreover, the model maintains strong cross-domain generalization when adapted to unseen target domains using only weak labels. These results demonstrate that UACA-Net effectively reduces reliance on strong annotations while achieving competitive disaggregation performance.