A Context-Aware Multi-Event Identification Method for Nonintrusive Load Monitoring

Non-intrusive load monitoring (NILM) is a method that provides appliance power consumption information, which will help enhance the smart grid applications. This paper proposes an end-to-end NILM method for multi-event identification, which alleviates the challenges of setting hyper-parameters and detecting multiple events in traditional methods. In this paper, convolutional neural networks are used to extract the local features of target event from the aggregated data in the sliding window. Then, the multi-head self-attention mechanism is introduced to extract the correlation between the sequence of events in the window, and the contextual information is fully used to distinguish similar events. Finally, a multi-scale anchor detection framework is introduced to identify multiple events in the window. In addition, this paper also proposes a novel data augmentation method to resolve the problem of insufficient event samples in the dataset to support model training. Comparative experiments were performed on two public datasets (REDD and UKDALE) with a variety of recently proposed methods in this paper to demonstrate the effectiveness and superiority of our method. The proposed method here achieved an average
$F_{1}$
score of 0.96 for multiple appliances of different power levels, which was 30% higher than that achieved by other compared methods.