An Industrial Non-Intrusive Load Disaggregation Method Based on Deep Learning Using Active and Reactive Power Features

Houkun Han

Abstract: Non-intrusive load monitoring provides a cost-effective way to obtain detailed equipment level energy consumption in buildings and industrial facilities. Existing studies mainly focus on residential loads and often rely only on active power features, which limits performance in industrial environments where reactive power is significant. This paper proposes an industrial load disaggregation method based on deep learning using both active and reactive power features. A hybrid architecture combining convolutional neural networks and long short-term memory networks is designed to extract spatial and temporal power signal features, and a channel attention mechanism is introduced to enhance important feature channels. Reactive power is incorporated as auxiliary information to improve the representation of industrial equipment operating patterns. Experiments conducted on the HIPE industrial dataset demonstrate that the proposed approach improves disaggregation accuracy compared with existing methods using only active power. The results confirm that reactive power information significantly enhances industrial load disaggregation performance.

Keywords: Non-intrusive Load Monitoring, industrial load disaggregation, deep learning, load power feature