Physics-Informed Time-Aware Neural Networks for Industrial Nonintrusive Load Monitoring

Abstract

Nonintrusive load monitoring enables the situational awareness of appliance-level energy consumption without installing appliance-specific sensors. It has been researched for over thirty years, with deep learning methods being the state-of-the-art solutions. However, current works mainly focus on the residential scenario, and industrial load disaggregation as a more challenging problem from the appliance type perspective is much less investigated. Nevertheless, industrial loads play an important role in energy savings and climate change mitigation and adaptation. Therefore, this paper focuses on the industrial nonintrusive load monitoring problem and proposes a physics-informed time-aware neural network method for it. Herein, multiple features of industrial loads are considered, and the physics relationship among them is leveraged to improve the learning process explicitly. In addition, a two-dimensional convolutional layer is further proposed to encode the timestamp for feature enhancement. Experiments on real-world industrial data from ten appliances will verify the effectiveness of the proposed method.