Sequence-to-Sequence-Based Beta-VAE Combined With IECA Attention Mechanism for Energy Disaggregation Algorithm
Haimeng Yu, Chengxin Pang, Yang Xuan, Yongbo Chen, Xinhua Zeng
Abstract
Nonintrusive load monitoring (NILM) is a technology that measures and identifies individual appliances’ energy consumption by analyzing the total power consumption collected by building sensors. Using the energy disaggregation method, the total power consumption can be disaggregated down to the device level. With the emergence of deep learning, energy disaggregation methods have improved in recent years. However, NILM still faces significant challenges in its generalization capabilities for different households and in disaggregating complex multi-state devices. To address these issues, on the basis of sequence-to-sequence, this paper proposes a Beta-Variational Autoencoder (Beta-VAE) combined with Improved Efficient Channel Attention (IECA) mechanism for energy disaggregation algorithm. In this model, Beta-VAE leverages its strong feature learning and disentanglement capabilities to improve the accuracy of multi-state device disaggregation by learning the feature representation of the latent space. Meanwhile, a new improved channel attention IECA is also introduced, which is combined with skip connection technology and is applied between the encoder and decoder to capture multi-scale features and improve the reconstruction of power signals. Furthermore, we apply the Gaussian Error Linear Unit (GELU) activation function and instance normalization to the feature extraction network to improve the convergence speed and generalization capability of the model. Finally, we tested and compared our method with several state-of-the-art NILM algorithms on the UK-DALE and REFIT datasets and obtain significant improvement. The results showed that the Mean Absolute Error (MAE) of all appliances decreased by an average of 46% and the F1 Score increased by 20%.