Sequence-to-point learning based on spatio-temporal attention fusion network for non-intrusive load monitoring

Most of existing non-invasive load monitoring (NILM) methods usually ignore the complementarity between temporal and spatial characteristics of appliance power data. To tackle this problem, this paper proposes a spatio-temporal attention fusion network with a sequence-to-point learning scheme for load disaggregation. Initially, a temporal feature extraction module is designed to extract temporal features over a large temporal receptive field. Then, an asymmetric inception module is designed for a multi-scale spatial feature extraction. The extracted temporal features and spatial features are concatenated, and fed into a polarized self-attention module to perform a spatio-temporal attention fusion, followed by two dense layers for final NILM predictions. Extensive experiments on two public datasets such as REDD and UK-DALE show the validity of the proposed method, outperforming the other used methods on NILM tasks.