SAMNet: Toward Latency-Free Non-Intrusive Load Monitoring via Multi-Task Deep Learning

Abstract

Non-intrusive load monitoring (NILM), including state detection and energy disaggregation, aims to identify the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">on/off</i> state and energy consumption from the aggregate load of a building. By monitoring the electrical behavior of consumers, smart grid applications such as demand response and recommendation services can also be realized for saving energy bills, environmental effectiveness, assisted living, and fault diagnosis. To achieve latency-free monitoring, this paper presents a <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</b> cale- and <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</b> ttention-experts based <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</b> ulti-task neural network (SAMNet) with a large enough context of the view to make full use of the correlation between the tasks of the NILM. A shared expert learner is designed to learn a good summary of the features. Self-attention mechanism is creatively adopted to realize the weighted fusion of different experts. To address the problem of manually setting different threshold values for different appliances to decide the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">on/off</i> states, we designed the network automatically labeling the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">on/off</i> states. Extensive experimental results with the public datasets demonstrate the effectiveness and superiority of our proposed model.