Non-intrusive load monitoring based on self-attention mechanism

Demand-side energy management relying on load monitoring technology is an important guarantee for promoting smart grid construction. Non-intrusive load decomposition (NILD) technology has received a lot of attention worldwide due to its low cost, easy maintenance and high security. In this work, we propose a neural network combining an on/off state classification subnetwork with a power regression subnetwork. We incorporate a tailored self-attention module into the power regression subtask to improve the generalization of the model. The experimental results show the proposed deep neural network outperforms other SGN models.