Non-intrusive load monitoring based on pyramid attention network and bigru network

In order to address the current issues of insufficient deep load feature extraction, low decomposition accuracy, and high training costs in load decomposition models, a parallel network model for sequence-to-point load decomposition is proposed. This model consists of two branches: a pyramid attention module and a temporal feature extraction network, achieving multi-scale feature fusion in both time and space. This alleviates the semantic loss problem during feature extraction in the pyramid network, enriches feature representation, and enhances load decomposition accuracy. Additionally, the attention module employs a multi-head self-attention mechanism, where each attention head focuses on different segments of load features, thus enabling key load feature selection from multiple angles to further improve load decomposition performance. Furthermore, a post-processing algorithm is introduced to optimize the load decomposition results. Finally, experiments on the UK_DALE dataset demonstrate that the proposed model outperforms four benchmark models used for comparison in terms of load decomposition performance.