A Few-Shot Load Identification Method with Class-Incremental Learning

Abstract

Non-intrusive load monitoring (NILM) enables fine-grained energy awareness and demand-side energy management in residential and commercial buildings, but practical deployment requires load identification models to accommodate newly introduced appliances from only a few labeled samples without forgetting previously learned classes. To address this challenge, this paper proposes a few-shot class-incremental load identification method for long-term NILM operation. Task-specific Adapters augment the shared backbone representation through independent lightweight adaptation branches, enabling discriminative representations of novel appliance classes to be learned under few-shot supervision. Prototype-based constraints further promote intra-class compactness and inter-class separation in the enhanced feature space, while exemplar replay preserves historical knowledge and mitigates representation drift during incremental updates. Experiments on the public PLAID and WHITED datasets under a unified five-stage, 10-shot class-incremental protocol show that the proposed method achieves average accuracies of 93.54% and 98.38%, respectively, outperforming the strongest baselines by 8.86 and 4.02 percentage points. After the final incremental stage, accuracies of 84.56% and 96.70% are retained. These results demonstrate that the proposed framework effectively balances novel-class adaptation and old-class knowledge retention, providing a reliable approach to appliance monitoring for energy-efficient building operation and intelligent demand-side management.