Attention-guided Temporal Convolutional Network for Non-intrusive Load Monitoring

Abstract

With the prevalence of smart meter infrastructure, data analysis on consumer side becomes more and more important in smart grid systems. One of the fundamental tasks is to disaggregate users' total consumption into appliance-wise values. It has been well noted that encoding of temporal dependency is a key issue for successful modelling of the relations between the total consumption and its decomposed consumption on an appliance historically, and therefore has been implemented in many state-of-the-art models. However, how to encode the varied long-term and short-term dependency coming from different appliances is yet an open and under-addressed question. In this paper, we propose an attention-guided temporal convolutional network (ATCN), which generates different temporal residual blocks and provides an attention mechanism to indicate the importance of those blocks with respect to the appliance. Ul-timately, we aim to address these two questions: i) How to employ both long-term and short-term temporal dependency to better disaggregate future loads while maintaining an affordable memory cost? ii) How to employ attention during the training of an appliance to obtain a better representation of the consumption pattern? We have demonstrated the effectiveness of our approach through comprehensive experiments and show that our proposed ATCN model achieves state-of-the-art performance, particularly on multi-status appliances that are normally hard to cope with regarding disaggregation accuracy and generalization capability.