Deep attention and generative neural networks for nonintrusive load monitoring

In recent years, Nonintrusive Load Monitoring (NILM) has been considered a crucial problem for energy monitoring and management, especially in the residential sector. In this area, deep learning has shown state-of-the-art performance due to its many tunable parameters that help its generalization power extract highly nonlinear patterns from the residential load. This paper reviews the recent discriminative deep neural networks proposed for NILM and develops attention-based deep neural architectures to improve the accuracy of current studies. In this context, we present a novel attention-based variational autoencoder and sparse coding model which uses the long short-term memory (LSTM) network and a dictionary learning module for generative disaggregation of residential loads. We show how the attention mechanisms can help deep learning better understand the correlations between NILM features and how generative pattern recognition can help capture the underlying probability densities of the features thereby improving the NILM accuracy for cutting-edge deep learning solutions. Our numerical results on the real-world Reference Energy Disaggregation Dataset (REDD) show remarkable improvements compared to the state-of-the-art using disaggregation accuracy and F-score metrics.