Latent abstraction bridge transformer for generalizable nonintrusive load monitoring

Nonintrusive load monitoring (NILM) is an effective approach for energy management that disaggregates the total power measured at the main power inlet into appliance-level power signals. NILM algorithms have achieved remarkable progress in recent years. However, accurately reconstructing appliance-level power signals from unseen, complex, and diverse aggregated data remains a formidable challenge. To address this challenge, this article proposes a novel hybrid load disaggregation model, the Latent Abstraction Bridge (LAB) Transformer, built on a sequence-to-sequence (S2S) framework that integrates a convolutional neural network (CNN) and a Transformer architecture with an embedding-constrained generative network termed LAB. The LAB effectively balances local discrete details and global information by leveraging a soft vector-quantized variational autoencoder (SoftVQ-VAE) and a beta-variational autoencoder (Beta-VAE) to constrain the encoder’s output representations, thereby considerably improving the model’s ability to generalize and discriminate in latent space. Moreover, we use parameter-free linear interpolation to recover the lengths of Beta-VAE output vectors, preserving essential global information while suppressing unnecessary local details, thereby substantially reducing the parameter count. The effectiveness of the proposed model is validated on two datasets: UK-DALE and REFIT. Experimental results indicate that it achieves the best F1 score, while lowering the mean absolute error (MAE) and signal aggregation error (SAE) by 22.8% and 24.7%, respectively, compared to several recent state-of-the-art models.

Scientific Reports

Scientific Reports

16

2045-2322

N/A

1

17