Gkalinikis, N. V., Nalmpantis, C., & Vrakas, D. (2020, October). Attention in Recurrent Neural Networks for Energy Disaggregation. In International Conference on Discovery Science (pp. 551-565). Springer, Cham.

Author(s): Gkalinikis, N. V., Nalmpantis, C., & Vrakas, D.



Abstract: Energy disaggregation refers to the separation of appliance-level data from an aggregate energy signal originated from a single-meter, without the use of any other device-specific sensors. Due to the fact that deep learning caught great attention in the last decade, numerous techniques using Artificial Neural Networks (ANN) have been developed to accomplish this task. Whereas most of the current research focuses on achieving better performance, the goal of this paper is to design a computationally light deep neural network based on attention mechanism. A thorough analysis shows how the proposed model is implemented and compares the performance of two different attention layers in the problem of energy disaggregation. The novel architecture achieves fast training and inference with minor performance trade-off when compared against other computationally expensive state-of-the-art models.