Abstract:

In smart-meter based non-intrusive load disaggregation with low-frequency sampling， load switching events are sparse and the class distribution is imbalanced. The shortage of minority-class and boundary samples tends to cause missed detections and misclassifications during load ON states and transition stages. Existing oversampling methods still have limitations in precisely controlling the number of synthesized samples， selecting nearest neighbors， and defining boundary samples， and random interpolation strategies may further introduce redundant samples or cross-class mixed samples. To address these issues， this paper proposes an improved algorithm that combines K-means clustering with Borderline-SMOTE （KB-SMOTE）： minority boundary samples are first extracted and clustered， and then centroid-guided within-cluster interpolation is performed to generate new samples， thereby reducing redundancy and enhancing boundary separability. At the model level， to overcome the limited capability of conventional sequence networks in capturing key transient and local features， a Bi-LSTM based load disaggregation model embedded with a convolutional block attention module is designed. By jointly leveraging channel and spatial attention， the model adaptively reweights features and strengthens key information relevant to load operating states. Experiments on the UK-DALE dataset show that， compared with baseline models including DAE， Seq2point， and the basic Bi-LSTM， the proposed method achieves better performance on multiple evaluation metrics， validating its effectiveness in imbalanced-load scenarios.

Key words: non-intrusive load monitoring（NILM）, imbalanced data, oversampling, attention mechanism, K-means clustering